// This file is part of OpenCV project. // It is subject to the license terms in the LICENSE file found in the top-level directory // of this distribution and at http://opencv.org/license.html. // This file is modified from the ficus (https://github.com/vpisarev/ficus/blob/master/lib/NN/OpConv_Winograd.fx). // Here is the original license: /* This file is a part of ficus language project. See ficus/LICENSE for the licensing terms */ #include "../../precomp.hpp" #include "convolution.hpp" #include "conv_winograd_f63.simd.hpp" #include "layers/cpu_kernels/conv_winograd_f63.simd_declarations.hpp" // defines CV_CPU_DISPATCH_MODES_ALL=AVX2,...,BASELINE based on CMakeLists.txt content namespace cv { namespace dnn { #if CV_NEON || CV_SIMD128 || CV_TRY_AVX2 enum { VEC_ALIGN = 32, DFT_TYPE = CV_32F }; // Memory alignment. void winofunc_accum_F32(const float* inwptr, const float* wptr, float* outbuf, int Cg, int iblock, const int winoIblock, const int winoKblock, const int winoAtomF32, const int winoNatomF32); /*Input transform*/ void winofunc_BtXB_8x8_F32(const float* inptr, int inpstep, float* outptr, int Cg, const int winoIblock, const int winoAtomF32); /*Output transform*/ void winofunc_AtXA_8x8_F32(const float* inptr, int inpstep, float* bpptr, int bpstep, float* outptr, int outstep, float bias, float minval, float maxval, bool ifMinMaxAct); int runWinograd63(InputArray _input, InputArray _fusedAddMat, OutputArray _output, const Ptr& conv, int ntasks, float minval, float maxval, ActivationLayer* activ, bool ifMinMaxAct) { Mat input = _input.getMat(); Mat output = _output.getMat(); Mat fusedAddMat = _fusedAddMat.getMat(); MatShape inputShape = shape(input); MatShape outputShape = shape(output); CV_Assert(inputShape.size() == 4 && outputShape.size() == 4); int N = inputShape[0], C = inputShape[1], Hi = inputShape[2], Wi = inputShape[3]; // [N, C, H, W] int K = conv->K; int H0 = outputShape[2], W0 = outputShape[3]; int pad_top = conv->pad_top; int pad_left = conv->pad_left; int ngroups = conv->ngroups, Cg = C/ngroups, Kg = K/ngroups; const int CONV_WINO_KBLOCK = 4; #if (CV_NEON && CV_NEON_AARCH64) const int CONV_WINO_IBLOCK = 6; #elif CV_TRY_AVX || CV_TRY_AVX2 const int CONV_WINO_IBLOCK = (conv->useAVX || conv->useAVX2) ? 6 : 3; #else const int CONV_WINO_IBLOCK = 3; #endif #if CV_TRY_AVX || CV_TRY_AVX2 const int CONV_WINO_ATOM_F32 = (conv->useAVX || conv->useAVX2) ? 8 : 4; #else const int CONV_WINO_ATOM_F32 = 4; #endif const int CONV_WINO_NATOMS_F32 = CONV_WINO_AREA / CONV_WINO_ATOM_F32; // for AVX2, it is 8, otherwise, it's 16. int CONV_WINO_ATOM = CONV_WINO_ATOM_F32; int CONV_WINO_NATOMS = CONV_WINO_NATOMS_F32; #ifdef CONV_ARM_FP16 // FP 16 const int CONV_WINO_ATOM_F16 = CONV_WINO_ATOM_F32 * 2; const int CONV_WINO_NATOMS_F16 = CONV_WINO_AREA / CONV_WINO_ATOM_F16; #endif int esz = sizeof(float ); #ifdef CONV_ARM_FP16 const bool useFP16 = conv->useFP16; if (useFP16) { // works at FP 16. CONV_WINO_ATOM = CONV_WINO_ATOM_F16; CONV_WINO_NATOMS = CONV_WINO_NATOMS_F16; esz = sizeof(__fp16); } #endif int Kg_nblocks = (Kg + CONV_WINO_KBLOCK - 1)/CONV_WINO_KBLOCK; const size_t inp_planesize = (size_t)Hi*Wi; const size_t out_planesize = (size_t)H0*W0; int blocks_per_row = (W0+CONV_WINO_STEP-1)/CONV_WINO_STEP; int blocks_per_plane = ((H0+CONV_WINO_STEP-1)/CONV_WINO_STEP)*blocks_per_row; int blocks_per_plane_aligned = ((blocks_per_plane + CONV_WINO_IBLOCK-1)/CONV_WINO_IBLOCK)*CONV_WINO_IBLOCK; size_t totalbufsize = (size_t)N*C*blocks_per_plane_aligned*CONV_WINO_AREA; AutoBuffer _buf; _buf.allocate((totalbufsize + VEC_ALIGN) * esz); char* wbuf_all = alignPtr(_buf.data(), VEC_ALIGN * esz); float* inp = input.ptr(); float* out = output.ptr(); float* fusedAddPtr = fusedAddMat.empty() ? nullptr : fusedAddMat.ptr(); // Phase 1. compute forward Winograd transforms for all input blocks, // all input planes, all samples in the batch. // [TODO]: maybe, if there are too many input channels, it makes sense to // transform only part of input channels at once and then compute the partial // accumulated sums (i.e. update the output buffers several times, // rather than compute them in one pass). parallel_for_(Range(0, ntasks), [&](const Range& r0) { for (int task_id = r0.start; task_id < r0.end; task_id++) { int nc0 = (N*C)*task_id/ntasks; int nc1 = (N*C)*(task_id+1)/ntasks; for(; nc0 < nc1; nc0++) { int n = nc0 / C; int c = nc0 - n*C; int g = c / Cg; c -= g*Cg; for (int block_id = 0; block_id < blocks_per_plane; block_id += CONV_WINO_IBLOCK) { for (int db = 0; db < CONV_WINO_IBLOCK; db++) { size_t inwofs = ((n*ngroups + g)*blocks_per_plane_aligned + block_id)*Cg*CONV_WINO_AREA + (c*CONV_WINO_IBLOCK + db) * CONV_WINO_ATOM; char* inwptr = wbuf_all + inwofs * esz; if (block_id + db < blocks_per_plane) { int y0 = (block_id + db) / blocks_per_row; int x0 = (block_id + db) - y0 * blocks_per_row; y0 = y0*CONV_WINO_STEP - pad_top; x0 = x0*CONV_WINO_STEP - pad_left; bool partial = y0 < 0 || y0 + CONV_WINO_SIZE > Hi || x0 < 0 || x0 + CONV_WINO_SIZE > Wi; int dx1 = 0, dx2 = CONV_WINO_SIZE, dy1 = 0, dy2 = CONV_WINO_SIZE; int inpstep = Wi; float inpbuf[CONV_WINO_AREA]; float* inptr0 = (float*)inp + nc0*inp_planesize + y0*Wi + x0; float* inptr = inptr0; if (partial) { memset(inpbuf, 0, sizeof(inpbuf)); dy1 = -y0 > 0 ? -y0 : 0; dy2 = Hi - y0 < CONV_WINO_SIZE ? Hi - y0 : CONV_WINO_SIZE; if (dy2 < dy1) {dy2 = dy1 = 0;} dx1 = -x0 > 0 ? -x0 : 0; dx2 = Wi - x0 < CONV_WINO_SIZE ? Wi - x0 : CONV_WINO_SIZE; if (dx2 < dx1) {dx2 = dx1 = 0;} inptr0 -= y0*Wi + x0; if (dx1 < dx2 && dy1 < dy2) { for(int dy = dy1; dy < dy2; dy++) memcpy(&inpbuf[dy*CONV_WINO_SIZE + dx1], inptr0 + (y0+dy)*Wi + (x0+dx1), (dx2-dx1)*sizeof(inpbuf[0])); } inptr = inpbuf; inpstep = CONV_WINO_SIZE; } #if CV_TRY_AVX2 if (conv->useAVX2) opt_AVX2::winofunc_BtXB_8x8_F32(inptr, inpstep, (float *)inwptr, Cg, CONV_WINO_IBLOCK, CONV_WINO_ATOM); else #endif #if CV_TRY_AVX if (conv->useAVX) opt_AVX::winofunc_BtXB_8x8_F32(inptr, inpstep, (float *)inwptr, Cg, CONV_WINO_IBLOCK, CONV_WINO_ATOM); else #endif #if CV_NEON && CV_NEON_AARCH64 if (conv->useNEON) { #ifdef CONV_ARM_FP16 if (useFP16) { opt_NEON_FP16::winofunc_BtXB_8x8_F16(inptr, inpstep, inwptr, Cg, CONV_WINO_IBLOCK, CONV_WINO_ATOM); } else #endif opt_NEON::winofunc_BtXB_8x8_F32(inptr, inpstep, (float *)inwptr, Cg, CONV_WINO_IBLOCK, CONV_WINO_ATOM); } else #endif winofunc_BtXB_8x8_F32(inptr, inpstep, (float *)inwptr, Cg, CONV_WINO_IBLOCK, CONV_WINO_ATOM); } else { for (int i = 0; i < CONV_WINO_NATOMS; i++, inwptr += CONV_WINO_IBLOCK * CONV_WINO_ATOM * esz) memset(inwptr, 0, CONV_WINO_ATOM * esz); } } } } }}); // Phase 2. compute elemwise-weighted sums of transformed blocks, // apply inverse Winograd transforms to the sums, // add bias, apply activation function if any and store the results. char* wptr0 = nullptr; #ifdef CONV_ARM_FP16 if (useFP16) { CV_Assert(!conv->weightsWinoBuf_FP16.empty()); wptr0 = (char *)conv->getWeightsWinoFP16(); } else #endif { CV_Assert(!conv->weightsWinoBuf.empty()); wptr0 = (char *)conv->getWeightsWino(); } parallel_for_(Range(0, ntasks), [&](const Range& r0) { for (int task_id = r0.start; task_id < r0.end; task_id++) { size_t out_wbuf_size = CONV_WINO_AREA * CONV_WINO_KBLOCK * CONV_WINO_IBLOCK; size_t outbuf_size = CONV_WINO_AREA; // For saving the accumulation output. AutoBuffer out_wbuf_; out_wbuf_.allocate((out_wbuf_size + VEC_ALIGN) * esz); char* out_wbuf = alignPtr(out_wbuf_.data(), VEC_ALIGN * esz); memset(out_wbuf, 0, out_wbuf_size * esz); // For saving the fuse_Add data. AutoBuffer outbuf_; outbuf_.allocate(outbuf_size + VEC_ALIGN); float* outbuf = alignPtr(outbuf_.data(), VEC_ALIGN); memset(outbuf, 0, outbuf_size * sizeof(outbuf[0])); int ngk0 = (int)(((int64_t)N*Kg_nblocks*ngroups)*task_id/ntasks); int ngk1 = (int)(((int64_t)N*Kg_nblocks*ngroups)*(task_id+1)/ntasks); for(; ngk0 < ngk1; ngk0++) { int n = ngk0 / (Kg_nblocks*ngroups); int gk0 = ngk0 % (Kg_nblocks*ngroups); int g = gk0 / Kg_nblocks; int k0 = (gk0 % Kg_nblocks)*CONV_WINO_KBLOCK; int k1 = k0 + CONV_WINO_KBLOCK <= Kg ? k0 + CONV_WINO_KBLOCK : Kg; for (int block_id0 = 0; block_id0 < blocks_per_plane; block_id0 += CONV_WINO_IBLOCK) { int block_id1 = block_id0 + CONV_WINO_IBLOCK; block_id1 = block_id1 < blocks_per_plane ? block_id1 : blocks_per_plane; size_t inwofs = ((n*ngroups + g)*blocks_per_plane_aligned + block_id0)*Cg*CONV_WINO_AREA; size_t wofs = (g*Kg_nblocks*CONV_WINO_KBLOCK + k0)*Cg*CONV_WINO_AREA; char* inwptr = wbuf_all + inwofs * esz; char* wptr = wptr0 + wofs * esz; #if CV_TRY_AVX2 if (conv->useAVX2) opt_AVX2::winofunc_accum_F32((float *)inwptr, (float *)wptr, (float *)out_wbuf, Cg, block_id1 - block_id0, CONV_WINO_IBLOCK, CONV_WINO_KBLOCK, CONV_WINO_ATOM, CONV_WINO_NATOMS); else #endif #if CV_TRY_AVX if (conv->useAVX) opt_AVX::winofunc_accum_F32((float *)inwptr, (float *)wptr, (float *)out_wbuf, Cg, block_id1 - block_id0, CONV_WINO_IBLOCK, CONV_WINO_KBLOCK, CONV_WINO_ATOM, CONV_WINO_NATOMS); else #endif #if CV_NEON && CV_NEON_AARCH64 if (conv->useNEON) { #ifdef CONV_ARM_FP16 if (useFP16) { opt_NEON_FP16::winofunc_accum_F16(inwptr, wptr, out_wbuf, Cg, block_id1 - block_id0, CONV_WINO_IBLOCK, CONV_WINO_KBLOCK, CONV_WINO_ATOM, CONV_WINO_NATOMS); } else #endif opt_NEON::winofunc_accum_F32((float *)inwptr, (float *)wptr, (float *)out_wbuf, Cg, block_id1 - block_id0, CONV_WINO_IBLOCK, CONV_WINO_KBLOCK, CONV_WINO_ATOM, CONV_WINO_NATOMS); } else #endif winofunc_accum_F32((float *)inwptr, (float *)wptr, (float *)out_wbuf, Cg, block_id1 - block_id0, CONV_WINO_IBLOCK, CONV_WINO_KBLOCK, CONV_WINO_ATOM, CONV_WINO_NATOMS); for (int k = k0; k < k1; k++) { float biasv = conv->biasBuf[g*Kg + k]; for (int block_id = block_id0; block_id < block_id1; block_id++) { int y0 = block_id / blocks_per_row; int x0 = block_id - y0 * blocks_per_row; y0 = y0*CONV_WINO_STEP; x0 = x0*CONV_WINO_STEP; int dy1 = H0 - y0; if (dy1 > CONV_WINO_STEP) dy1 = CONV_WINO_STEP; int dx1 = W0 - x0; if (dx1 > CONV_WINO_STEP) dx1 = CONV_WINO_STEP; assert(dx1 > 0 && dy1 > 0); bool partial = activ || dy1 < CONV_WINO_STEP || dx1 < CONV_WINO_STEP; size_t outofs = (n*K + g*Kg + k)*out_planesize + y0*W0 + x0; int outstep = W0; float* outptr0 = (float*)out + outofs; float* pbptr0 = fusedAddPtr ? fusedAddPtr + outofs : nullptr; float *outptr = outptr0, *bpptr = pbptr0; if (partial) { outptr = outbuf; outstep = CONV_WINO_SIZE; if (pbptr0) { bpptr = outbuf; for (int y = 0; y < dy1; y++) memcpy(outbuf + y*CONV_WINO_SIZE, pbptr0 + y*W0, dx1*sizeof(pbptr0[0])); } } #if CV_TRY_AVX2 if (conv->useAVX2) opt_AVX2::winofunc_AtXA_8x8_F32((float *)out_wbuf + ((k - k0)*CONV_WINO_IBLOCK + (block_id - block_id0))*CONV_WINO_AREA, CONV_WINO_SIZE, bpptr, outstep, outptr, outstep, biasv, minval, maxval, ifMinMaxAct); else #endif #if CV_TRY_AVX if (conv->useAVX) opt_AVX::winofunc_AtXA_8x8_F32((float *)out_wbuf + ((k - k0)*CONV_WINO_IBLOCK + (block_id - block_id0))*CONV_WINO_AREA, CONV_WINO_SIZE, bpptr, outstep, outptr, outstep, biasv, minval, maxval, ifMinMaxAct); else #endif #if CV_NEON && CV_NEON_AARCH64 // NEON optimization is only for ARMv8 device, and for ARMv7 device, we use the Universal intrinsics. if (conv->useNEON) { #ifdef CONV_ARM_FP16 if (useFP16) { opt_NEON_FP16::winofunc_AtXA_8x8_F16(out_wbuf + ((k - k0)*CONV_WINO_IBLOCK + (block_id - block_id0))*CONV_WINO_AREA * esz, CONV_WINO_SIZE, bpptr, outstep, outptr, outstep, biasv, minval, maxval, ifMinMaxAct); } else #endif opt_NEON::winofunc_AtXA_8x8_F32((float *)out_wbuf + ((k - k0)*CONV_WINO_IBLOCK + (block_id - block_id0))*CONV_WINO_AREA, CONV_WINO_SIZE, bpptr, outstep, outptr, outstep, biasv, minval, maxval, ifMinMaxAct); } else #endif winofunc_AtXA_8x8_F32((float *)out_wbuf + ((k - k0)*CONV_WINO_IBLOCK + (block_id - block_id0))*CONV_WINO_AREA, CONV_WINO_SIZE, bpptr, outstep, outptr, outstep, biasv, minval, maxval, ifMinMaxAct); if (partial) { if (activ) activ->forwardSlice(outptr, outptr, CONV_WINO_SIZE*CONV_WINO_STEP, 0, g*Kg + k, g*Kg + k + 1); for (int y = 0; y < dy1; y++) memcpy(outptr0 + y*W0, outptr + y*CONV_WINO_SIZE, dx1*sizeof(outptr0[0])); } } } } } }}); return 1; } /****************************************************************************************\ SIMD for winograd function \****************************************************************************************/ #if CV_SIMD128 void winofunc_accum_F32(const float* inwptr, const float* wptr, float* outbuf, int Cg, int iblock, const int winoIblock, const int winoKblock, const int winoAtomF32, const int winoNatomF32) { #if 1 CV_Assert(winoIblock == 3 && winoKblock == 4 && winoAtomF32 == 4); for (int atom_id = 0; atom_id < winoNatomF32; atom_id++, outbuf += winoAtomF32) { v_float32x4 s00 = v_setzero_f32(), s01 = s00, s02 = s00; v_float32x4 s10 = v_setzero_f32(), s11 = s00, s12 = s00; v_float32x4 s20 = v_setzero_f32(), s21 = s00, s22 = s00; v_float32x4 s30 = v_setzero_f32(), s31 = s00, s32 = s00; for (int c = 0; c < Cg; c++, inwptr += winoIblock*winoAtomF32, wptr += winoKblock*winoAtomF32) { v_float32x4 x0, x1, x2; x0 = v_load(inwptr); x1 = v_load(inwptr + 4); x2 = v_load(inwptr + 8); v_float32x4 w0 = v_load(wptr); s00 = v_fma(w0, x0, s00); s01 = v_fma(w0, x1, s01); s02 = v_fma(w0, x2, s02); w0 = v_load(wptr + 4); s10 = v_fma(w0, x0, s10); s11 = v_fma(w0, x1, s11); s12 = v_fma(w0, x2, s12); w0 = v_load(wptr + 8); s20 = v_fma(w0, x0, s20); s21 = v_fma(w0, x1, s21); s22 = v_fma(w0, x2, s22); w0 = v_load(wptr + 12); s30 = v_fma(w0, x0, s30); s31 = v_fma(w0, x1, s31); s32 = v_fma(w0, x2, s32); } v_store(outbuf, s00); v_store(outbuf + 1*64, s01); v_store(outbuf + 2*64, s02); v_store(outbuf + 3*64, s10); v_store(outbuf + 4*64, s11); v_store(outbuf + 5*64, s12); v_store(outbuf + 6*64, s20); v_store(outbuf + 7*64, s21); v_store(outbuf + 8*64, s22); v_store(outbuf + 9*64, s30); v_store(outbuf + 10*64, s31); v_store(outbuf + 11*64, s32); } #else // Naive C++ code, the code should never be run here. for (int atom_id = 0; atom_id < winoNatomF32; atom_id++, outbuf += winoAtomF32) { float sumbuf[winoIblock*winoKblock*winoAtomF32]; memset(sumbuf, 0, sizeof(sumbuf)); for (int c = 0; c < Cg; c++, inwptr += winoIblock*winoAtomF32, wptr += winoKblock*winoAtomF32) { for (int i = 0; i < winoKblock; i++) { for (int j = 0; j < winoIblock; j++) { int i_ = i*winoAtomF32; int j_ = j*winoAtomF32; int ij_ = i_*winoIblock + j_; float s0 = inwptr[j_ + 0]*wptr[i_ + 0]; float s1 = inwptr[j_ + 1]*wptr[i_ + 1]; float s2 = inwptr[j_ + 2]*wptr[i_ + 2]; float s3 = inwptr[j_ + 3]*wptr[i_ + 3]; sumbuf[ij_ + 0] += s0; sumbuf[ij_ + 1] += s1; sumbuf[ij_ + 2] += s2; sumbuf[ij_ + 3] += s3; } } } for (int ij = 0; ij < winoKblock*winoIblock; ij++) { int ij_ = ij*winoAtomF32; int ij_out = ij*CONV_WINO_AREA; outbuf[ij_out + 0] = sumbuf[ij_ + 0]; outbuf[ij_out + 1] = sumbuf[ij_ + 1]; outbuf[ij_out + 2] = sumbuf[ij_ + 2]; outbuf[ij_out + 3] = sumbuf[ij_ + 3]; } } #endif } /*Input transform*/ void winofunc_BtXB_8x8_F32(const float* inptr, int inpstep, float* outptr, int Cg, const int winoIblock, const int winoAtomF32) { CV_Assert(winoIblock == 3 && winoAtomF32 == 4); v_float32x4 x00 = v_load(inptr), x01 = v_load(inptr + 4); v_float32x4 x10 = v_load(inptr + inpstep), x11 = v_load(inptr + inpstep + 4); v_float32x4 x20 = v_load(inptr + inpstep*2), x21 = v_load(inptr + inpstep*2 + 4); v_float32x4 x30 = v_load(inptr + inpstep*3), x31 = v_load(inptr + inpstep*3 + 4); v_float32x4 x40 = v_load(inptr + inpstep*4), x41 = v_load(inptr + inpstep*4 + 4); v_float32x4 x50 = v_load(inptr + inpstep*5), x51 = v_load(inptr + inpstep*5 + 4); v_float32x4 x60 = v_load(inptr + inpstep*6), x61 = v_load(inptr + inpstep*6 + 4); v_float32x4 x70 = v_load(inptr + inpstep*7), x71 = v_load(inptr + inpstep*7 + 4); v_float32x4 z00, z01, z10, z11, z20, z21, z30, z31, z40, z41, z50, z51, z60, z61, z70, z71; { /* Y[0] = [1.f, 0.f, -5.25f, 0.f, 5.25f, 0.f, -1.f, 0.f]*X */ /* Y[7] = [0.f, -1.f, 0.f, 5.25f, 0.f, -5.25f, 0.f, 1.f]*X */ v_float32x4 q5_25 = v_setall_f32(5.25f), t00, t01, t10, t11; t00 = v_sub(x40, x20); t01 = v_sub(x41, x21); t10 = v_sub(x30, x50); t11 = v_sub(x31, x51); v_float32x4 y00 = v_fma(t00, q5_25, v_sub(x00, x60)); v_float32x4 y01 = v_fma(t01, q5_25, v_sub(x01, x61)); v_float32x4 y70 = v_fma(t10, q5_25, v_sub(x70, x10)); v_float32x4 y71 = v_fma(t11, q5_25, v_sub(x71, x11)); /* Y[1] = [0.f, 1.f, 1.f, -4.25f, -4.25f, 1.f, 1.f, 0.f]*X */ /* Y[2] = [0.f, -1.f, 1.f, 4.25f, -4.25f, -1.f, 1.f, 0.f]*X */ v_float32x4 qm4_25 = v_setall_f32(-4.25f); t00 = v_fma(x30, qm4_25, v_add(x10, x50)); t01 = v_fma(x31, qm4_25, v_add(x11, x51)); t10 = v_fma(x40, qm4_25, v_add(x20, x60)); t11 = v_fma(x41, qm4_25, v_add(x21, x61)); v_float32x4 y10 = v_add(t00, t10), y11 = v_add(t01, t11); v_float32x4 y20 = v_sub(t10, t00), y21 = v_sub(t11, t01); /* Y[3] = [0.f, 0.5f, 0.25f, -2.5f, -1.25f, 2.f, 1.f, 0.f]*X */ /* Y[4] = [0.f, -0.5f, 0.25f, 2.5f, -1.25f, -2.f, 1.f, 0.f]*X */ v_float32x4 q0_5 = v_setall_f32(0.5f), q0_25 = v_setall_f32(0.25f); v_float32x4 qm2_5 = v_setall_f32(-2.5f), qm1_25 = v_setall_f32(-1.25f); t00 = v_fma(x10, q0_5, v_add(x50, x50)); t01 = v_fma(x11, q0_5, v_add(x51, x51)); t10 = v_fma(x20, q0_25, x60); t11 = v_fma(x21, q0_25, x61); t00 = v_fma(x30, qm2_5, t00); t01 = v_fma(x31, qm2_5, t01); t10 = v_fma(x40, qm1_25, t10); t11 = v_fma(x41, qm1_25, t11); v_float32x4 y30 = v_add(t00, t10), y31 = v_add(t01, t11); v_float32x4 y40 = v_sub(t10, t00), y41 = v_sub(t11, t01); /* Y[5] = [0.f, 2.f, 4.f, -2.5f, -5.f, 0.5f, 1.f, 0.f]*X */ /* Y[6] = [0.f, -2.f, 4.f, 2.5f, -5.f, -0.5f, 1.f, 0.f]*X */ v_float32x4 q4 = v_setall_f32(4.f), qm5 = v_setall_f32(-5.f); t00 = v_fma(x50, q0_5, v_add(x10, x10)); t01 = v_fma(x51, q0_5, v_add(x11, x11)); t10 = v_fma(x20, q4 , x60); t11 = v_fma(x21, q4 , x61); t00 = v_fma(x30, qm2_5, t00); t01 = v_fma(x31, qm2_5, t01); t10 = v_fma(x40, qm5 , t10); t11 = v_fma(x41, qm5 , t11); v_float32x4 y50 = v_add(t00, t10), y51 = v_add(t01, t11); v_float32x4 y60 = v_sub(t10, t00), y61 = v_sub(t11, t01); /* transpose 8x8 matrix with v_transpose4x4 */ v_float32x4 y000, y100, y200, y300, y010, y110, y210, y310, y400, y500, y600, y700, y410, y510, y610, y710; v_transpose4x4(y00, y10, y20, y30, y000, y100, y200, y300); v_transpose4x4(y01, y11, y21, y31, y010, y110, y210, y310); v_transpose4x4(y40, y50, y60, y70, y400, y500, y600, y700); v_transpose4x4(y41, y51, y61, y71, y410, y510, y610, y710); /* Z[0] = [1.f, 0.f, -5.25f, 0.f, 5.25f, 0.f, -1.f, 0.f]*Y */ /* Z[7] = [0.f, -1.f, 0.f, 5.25f, 0.f, -5.25f, 0.f, 1.f]*Y */ t00 = v_sub(y010, y200); t01 = v_sub(y410, y600); t10 = v_sub(y300, y110); t11 = v_sub(y700, y510); z00 = v_fma(t00, q5_25, v_sub(y000, y210)); z01 = v_fma(t01, q5_25, v_sub(y400, y610)); z70 = v_fma(t10, q5_25, v_sub(y310, y100)); z71 = v_fma(t11, q5_25, v_sub(y710, y500)); /* Z[1] = [0.f, 1.f, 1.f, -4.25f, -4.25f, 1.f, 1.f, 0.f]*Y */ /* Z[2] = [0.f, -1.f, 1.f, 4.25f, -4.25f, -1.f, 1.f, 0.f]*Y */ t00 = v_fma(y300, qm4_25, v_add(y100, y110)); t01 = v_fma(y700, qm4_25, v_add(y500, y510)); t10 = v_fma(y010, qm4_25, v_add(y200, y210)); t11 = v_fma(y410, qm4_25, v_add(y600, y610)); z10 = v_add(t00, t10); z11 = v_add(t01, t11); z20 = v_sub(t10, t00); z21 = v_sub(t11, t01); /* Z[3] = [0.f, 0.5f, 0.25f, -2.5f, -1.25f, 2.f, 1.f, 0.f]*Y */ /* Z[4] = [0.f, -0.5f, 0.25f, 2.5f, -1.25f, -2.f, 1.f, 0.f]*Y */ t00 = v_fma(y100, q0_5, v_add(y110, y110)); t01 = v_fma(y500, q0_5, v_add(y510, y510)); t10 = v_fma(y200, q0_25, y210); t11 = v_fma(y600, q0_25, y610); t00 = v_fma(y300, qm2_5, t00); t01 = v_fma(y700, qm2_5, t01); t10 = v_fma(y010, qm1_25, t10); t11 = v_fma(y410, qm1_25, t11); z30 = v_add(t00, t10); z31 = v_add(t01, t11); z40 = v_sub(t10, t00); z41 = v_sub(t11, t01); /* Z[5] = [0.f, 2.f, 4.f, -2.5f, -5.f, 0.5f, 1.f, 0.f]*Y */ /* Z[6] = [0.f, -2.f, 4.f, 2.5f, -5.f, -0.5f, 1.f, 0.f]*Y */ t00 = v_fma(y110, q0_5, v_add(y100, y100)); t01 = v_fma(y510, q0_5, v_add(y500, y500)); t10 = v_fma(y200, q4, y210); t11 = v_fma(y600, q4, y610); t00 = v_fma(y300, qm2_5, t00); t01 = v_fma(y700, qm2_5, t01); t10 = v_fma(y010, qm5, t10); t11 = v_fma(y410, qm5, t11); z50 = v_add(t00, t10); z51 = v_add(t01, t11); z60 = v_sub(t10, t00); z61 = v_sub(t11, t01); } const int outstep = winoIblock*winoAtomF32*Cg; v_store(outptr, z00); v_store(outptr + outstep, z01); v_store(outptr + outstep*2, z10); v_store(outptr + outstep*3, z11); v_store(outptr + outstep*4, z20); v_store(outptr + outstep*5, z21); v_store(outptr + outstep*6, z30); v_store(outptr + outstep*7, z31); v_store(outptr + outstep*8, z40); v_store(outptr + outstep*9, z41); v_store(outptr + outstep*10, z50); v_store(outptr + outstep*11, z51); v_store(outptr + outstep*12, z60); v_store(outptr + outstep*13, z61); v_store(outptr + outstep*14, z70); v_store(outptr + outstep*15, z71); } /*Output transform*/ /* Inverse Winograd 8x8 transform: out = (A'*inp*A)', where inp is input 8x8 FP32 matrix, A' is [1.f, 1.f, 1.f, 1.f, 1.f, 1.f, 1.f, 0.f, 0.f, 1.f, -1.f, 2.f, -2.f, 0.5f, -0.5f, 0.f, 0.f, 1.f, 1.f, 4.f, 4.f, 0.25f, 0.25f, 0.f, 0.f, 1.f, -1.f, 8.f, -8.f, 0.125f, -0.125f, 0.f, 0.f, 1.f, 1.f, 16.f, 16.f, 1.f/16, 1.f/16, 0.f, 0.f, 1.f, -1.f, 32.f, -32.f, 1.f/32, -1.f/32, 1.f] inp is pre-loaded into xij registers, out will be stored in zij, where (0<=i<=7 for x, 0<=i<=5 for z), 0<=j<=1. After the inverse transform is done, we add bias, optionally add results from the earlier tensors (by-pass), optionally apply activation function and then store the final results. That is, after both forward and then inverse transformation, we get non-transposed result. Of course, for the correct work of Winograd-based convolution, the Winograd-transformed weights should also be transposed. init_conv() (see OpConv.fx) takes care of that. */ void winofunc_AtXA_8x8_F32(const float* inptr, int inpstep, float* bpptr, int bpstep, float* outptr, int outstep, float bias, float minval, float maxval, bool ifMinMaxAct) { v_float32x4 x00 = v_load(inptr), x01 = v_load(inptr + 4); v_float32x4 x10 = v_load(inptr + inpstep), x11 = v_load(inptr + inpstep + 4); v_float32x4 x20 = v_load(inptr + inpstep*2), x21 = v_load(inptr + inpstep*2 + 4); v_float32x4 x30 = v_load(inptr + inpstep*3), x31 = v_load(inptr + inpstep*3 + 4); v_float32x4 x40 = v_load(inptr + inpstep*4), x41 = v_load(inptr + inpstep*4 + 4); v_float32x4 x50 = v_load(inptr + inpstep*5), x51 = v_load(inptr + inpstep*5 + 4); v_float32x4 x60 = v_load(inptr + inpstep*6), x61 = v_load(inptr + inpstep*6 + 4); v_float32x4 x70 = v_load(inptr + inpstep*7), x71 = v_load(inptr + inpstep*7 + 4); v_float32x4 z00, z01, z10, z11, z20, z21, z30, z31, z40, z41, z50, z51; { v_float32x4 s12_0, s12_1, s34_0, s34_1, s56_0, s56_1; s12_0 = v_add(x10, x20); s12_1 = v_add(x11, x21); s34_0 = v_add(x30, x40); s34_1 = v_add(x31, x41); s56_0 = v_add(x50, x60); s56_1 = v_add(x51, x61); v_float32x4 y00 = v_add(v_add(v_add(x00, s12_0), s34_0), s56_0); v_float32x4 y01 = v_add(v_add(v_add(x01, s12_1), s34_1), s56_1); v_float32x4 a0 = v_setall_f32(0.25f), a1 = v_setall_f32(4.0f); v_float32x4 y20 = v_fma(s56_0, a0, v_fma(s34_0, a1, s12_0)); v_float32x4 y21 = v_fma(s56_1, a0 ,v_fma(s34_1, a1, s12_1) ); a0 = v_setall_f32(1.f/16), a1 = v_setall_f32(16.0f); v_float32x4 y40 = v_fma(s56_0, a0, v_fma(s34_0, a1, s12_0)); v_float32x4 y41 = v_fma(s56_1, a0, v_fma(s34_1, a1, s12_1)); s12_0 = v_sub(x10, x20); s12_1 = v_sub(x11, x21); s34_0 = v_sub(x30, x40); s34_1 = v_sub(x31, x41); s56_0 = v_sub(x50, x60); s56_1 = v_sub(x51, x61); a0 = v_setall_f32(1.f/32), a1 = v_setall_f32(32.f); v_float32x4 y50 = v_fma(s56_0, a0, v_fma(s34_0, a1, v_add(x70, s12_0))); v_float32x4 y51 = v_fma(s56_1, a0, v_fma(s34_1, a1, v_add(x71, s12_1))); a0 = v_setall_f32(0.5f), a1 = v_setall_f32(2.f); v_float32x4 y10 = v_fma(s56_0, a0, v_fma(s34_0, a1, s12_0)); v_float32x4 y11 = v_fma(s56_1, a0, v_fma(s34_1, a1, s12_1)); a0 = v_setall_f32(0.125f), a1 = v_setall_f32(8.f); v_float32x4 y30 = v_fma(s56_0, a0, v_fma(s34_0, a1, s12_0)); v_float32x4 y31 = v_fma(s56_1, a0, v_fma(s34_1, a1, s12_1)); v_float32x4 y60 = v_setall_f32(0.f), y61 = y60, y70 = y60, y71 = y60; /* transpose 8x8 matrix with v_transpose4x4 */ v_float32x4 y000, y100, y200, y300, y010, y110, y210, y310, y400, y500, y600, y700, y410, y510, y610, y710; v_transpose4x4(y00, y10, y20, y30, y000, y100, y200, y300); v_transpose4x4(y01, y11, y21, y31, y010, y110, y210, y310); v_transpose4x4(y40, y50, y60, y70, y400, y500, y600, y700); v_transpose4x4(y41, y51, y61, y71, y410, y510, y610, y710); s12_0 = v_add(y100, y200); s12_1 = v_add(y500, y600); s34_0 = v_add(y300, y010); s34_1 = v_add(y700, y410); s56_0 = v_add(y110, y210); s56_1 = v_add(y510, y610); z00 = v_add(v_add(v_add(y000, s12_0), s34_0), s56_0); z01 = v_add(v_add(v_add(y400, s12_1), s34_1), s56_1); a0 = v_setall_f32(0.25f), a1 = v_setall_f32(4.0f); z20 = v_fma(s56_0, a0, v_fma(s34_0, a1, s12_0)); z21 = v_fma(s56_1, a0, v_fma(s34_1, a1, s12_1)); a0 = v_setall_f32(1.f/16), a1 = v_setall_f32(16.0f); z40 = v_fma(s56_0, a0, v_fma(s34_0, a1, s12_0)); z41 = v_fma(s56_1, a0, v_fma(s34_1, a1, s12_1)); s12_0 = v_sub(y100, y200); s12_1 = v_sub(y500, y600); s34_0 = v_sub(y300, y010); s34_1 = v_sub(y700, y410); s56_0 = v_sub(y110, y210); s56_1 = v_sub(y510, y610); a0 = v_setall_f32(1.f/32), a1 = v_setall_f32(32.0f); z50 = v_fma(s56_0, a0, v_fma(s34_0, a1, v_add(y310, s12_0))); z51 = v_fma(s56_1, a0, v_fma(s34_1, a1, v_add(y710, s12_1))); a0 = v_setall_f32(0.5f), a1 = v_setall_f32(2.0f); z10 = v_fma(s56_0, a0, v_fma(s34_0, a1, s12_0)); z11 = v_fma(s56_1, a0, v_fma(s34_1, a1, s12_1)); a0 = v_setall_f32(0.125f), a1 = v_setall_f32(8.0f); z30 = v_fma(s56_0, a0, v_fma(s34_0, a1, s12_0)); z31 = v_fma(s56_1, a0, v_fma(s34_1, a1, s12_1)); v_float32x4 vbias = v_setall_f32(bias); z00 = v_add(z00, vbias); z01 = v_add(z01, vbias); z10 = v_add(z10, vbias); z11 = v_add(z11, vbias); z20 = v_add(z20, vbias); z21 = v_add(z21, vbias); z30 = v_add(z30, vbias); z31 = v_add(z31, vbias); z40 = v_add(z40, vbias); z41 = v_add(z41, vbias); z50 = v_add(z50, vbias); z51 = v_add(z51, vbias); } if (bpptr) { z00 = v_add(z00, v_load(bpptr)); z01 = v_add(z01, v_load_low(bpptr + 4)); z10 = v_add(z10, v_load(bpptr + bpstep)); z11 = v_add(z11, v_load_low(bpptr + bpstep + 4)); z20 = v_add(z20, v_load(bpptr + bpstep * 2)); z21 = v_add(z21, v_load_low(bpptr + bpstep * 2 + 4)); z30 = v_add(z30, v_load(bpptr + bpstep * 3)); z31 = v_add(z31, v_load_low(bpptr + bpstep * 3 + 4)); z40 = v_add(z40, v_load(bpptr + bpstep * 4)); z41 = v_add(z41, v_load_low(bpptr + bpstep * 4 + 4)); z50 = v_add(z50, v_load(bpptr + bpstep * 5)); z51 = v_add(z51, v_load_low(bpptr + bpstep * 5 + 4)); } if (ifMinMaxAct) { v_float32x4 vmax = v_setall_f32(maxval); v_float32x4 vmin = v_setall_f32(minval); z00 = v_min(v_max(z00, vmin), vmax); z01 = v_min(v_max(z01, vmin), vmax); z10 = v_min(v_max(z10, vmin), vmax); z11 = v_min(v_max(z11, vmin), vmax); z20 = v_min(v_max(z20, vmin), vmax); z21 = v_min(v_max(z21, vmin), vmax); z30 = v_min(v_max(z30, vmin), vmax); z31 = v_min(v_max(z31, vmin), vmax); z40 = v_min(v_max(z40, vmin), vmax); z41 = v_min(v_max(z41, vmin), vmax); z50 = v_min(v_max(z50, vmin), vmax); z51 = v_min(v_max(z51, vmin), vmax); } v_store(outptr, z00); v_store_low(outptr + 4, z01); v_store(outptr + outstep, z10); v_store_low(outptr + outstep + 4, z11); v_store(outptr + outstep*2, z20); v_store_low(outptr + outstep*2 + 4, z21); v_store(outptr + outstep*3, z30); v_store_low(outptr + outstep*3 + 4, z31); v_store(outptr + outstep*4, z40); v_store_low(outptr + outstep*4 + 4, z41); v_store(outptr + outstep*5, z50); v_store_low(outptr + outstep*5 + 4, z51); } #endif #else int runWinograd63(InputArray _input, InputArray _fusedAddMat, OutputArray _output, const Ptr& conv, int ntasks, float minval, float maxval, ActivationLayer* activ, bool ifMinMaxAct) { return 0; } #endif }} // namespace cv::dnn