File size: 21,241 Bytes
67a5826 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 | #pragma once
#include <cuda_bf16.h>
#include <cuda_fp16.h>
#include <cuda_runtime.h>
#include <cute/tensor.hpp>
#include "config.hpp"
#include "marlin_utils.hpp"
namespace packbits_utils {
template <class SourceEngine , class SourceLayout ,
class SourceEngine2 , class SourceLayout2 ,
class TargetEngine , class TargetLayout ,
class ScaleEngine , class ScaleLayout ,
class QuantMapEngine , class QuantMapLayout ,
class QuantMapEngine2, class QuantMapLayout2,
class QuantMapEngine3, class QuantMapLayout3,
class NumBits,
config::QuantMapModeEnum QuantMapMode>
struct DequantizationTraits
{
CUTE_DEVICE static
void
apply(
cute::Tensor<SourceEngine , SourceLayout > const& source,
cute::Tensor<SourceEngine2 , SourceLayout2 > const& source2,
cute::Tensor<TargetEngine , TargetLayout > & target,
cute::Tensor<ScaleEngine , ScaleLayout > const& scale,
cute::Tensor<QuantMapEngine , QuantMapLayout > const& qmap,
cute::Tensor<QuantMapEngine2, QuantMapLayout2> const& qmap2,
cute::Tensor<QuantMapEngine3, QuantMapLayout3> const& qmap3)
{
using TQ = cute::uint16_t;
using TQ2 = cute::uint32_t;
using T = typename TargetEngine::value_type;
using TI = cute::conditional_t<cute::is_same_v<T, cute::half_t>, __half , __nv_bfloat16 >;
using T2 = cute::conditional_t<cute::is_same_v<T, cute::half_t>, __half2, __nv_bfloat162>;
CUTE_STATIC_ASSERT(cute::is_same_v<T , cute::half_t > == true ||
cute::is_same_v<T , cute::bfloat16_t > == true);
CUTE_STATIC_ASSERT(cute::is_same_v<TQ, typename SourceEngine ::value_type> == true);
CUTE_STATIC_ASSERT(cute::is_same_v<TQ, typename SourceEngine2 ::value_type> == true);
CUTE_STATIC_ASSERT(cute::is_same_v<T , typename TargetEngine ::value_type> == true);
CUTE_STATIC_ASSERT(cute::is_same_v<T , typename ScaleEngine ::value_type> == true);
CUTE_STATIC_ASSERT(cute::is_same_v<T , typename QuantMapEngine ::value_type> == true);
CUTE_STATIC_ASSERT(cute::is_same_v<T2, typename QuantMapEngine2::value_type> == true);
CUTE_STATIC_ASSERT(cute::is_same_v<T , typename QuantMapEngine3::value_type> == true);
static constexpr int kNumBits = NumBits::value;
static constexpr int kNumPacked2 = NumBits::value == 4 ? 8 : 16;
static constexpr cute::uint16_t kMask = NumBits::value == 4 ? 0x000f : 0x0003;
static constexpr cute::uint32_t kMask2 = NumBits::value == 4 ? 0x000000ff : 0x0000000f;
static constexpr cute::uint32_t kMaskSync = 0xffffffff;
// vectorize the source and target
auto source_vec = cute::recast<TQ2>(source);
auto source2_vec = cute::recast<TQ2>(source2); // unused
auto target_vec = cute::recast<T2 >(target);
auto scale_vec = cute::recast<T2 >(scale);
auto qmap_view = cute::recast<TI >(qmap);
auto qmap2_view = cute::recast<T2 >(qmap2);
auto qmap3_view = cute::recast<TI >(qmap3);
CUTE_STATIC_ASSERT_V(NumBits{} == cute::_4{} || NumBits{} == cute::_2{});
CUTE_STATIC_ASSERT_V(cute::size<0>(source ) == cute::size<0>(source_vec ) * cute::_2{});
CUTE_STATIC_ASSERT_V(cute::size<0>(source2) == cute::size<0>(source2_vec) * cute::_2{});
CUTE_STATIC_ASSERT_V(cute::size<0>(target ) == cute::size<0>(target_vec ) * cute::_2{});
CUTE_STATIC_ASSERT_V(cute::size<0>(scale ) == cute::size<0>(scale_vec ) * cute::_2{});
CUTE_STATIC_ASSERT_V(cute::size<1>(source ) == cute::size<1>(source_vec ));
CUTE_STATIC_ASSERT_V(cute::size<1>(source2) == cute::size<1>(source2_vec));
CUTE_STATIC_ASSERT_V(cute::size<1>(target ) == cute::size<1>(target_vec ));
CUTE_STATIC_ASSERT_V(cute::size<1>(scale ) == cute::size<1>(scale_vec ));
CUTE_STATIC_ASSERT_V(cute::size (qmap ) == cute::size (qmap_view ));
CUTE_STATIC_ASSERT_V(cute::size (qmap2 ) == cute::size (qmap2_view ));
CUTE_STATIC_ASSERT_V(cute::size (qmap3 ) == cute::size (qmap3_view ));
CUTE_UNROLL
for (int i = 0; i < cute::size<0>(source_vec); ++i)
{
CUTE_UNROLL
for (int p = 0; p < cute::size<1>(source_vec); ++p)
{
auto src_crd = cute::make_coord(i, p);
CUTE_UNROLL
for (int k2 = 0; k2 < kNumPacked2; k2 += 2)
{
auto k = k2 / 2;
auto tgt_crd = cute::make_coord(i, k * cute::size<1>(source_vec) + p);
auto src_raw = source_vec(src_crd) >> (k2 * kNumBits);
T2 src_val;
if constexpr ((QuantMapMode == config::QuantMapModeEnum::Vectorized ) ||
(QuantMapMode == config::QuantMapModeEnum::Vectorized_32) ||
(QuantMapMode == config::QuantMapModeEnum::Vectorized_16) ||
(QuantMapMode == config::QuantMapModeEnum::Vectorized_8))
{
// vectorized table lookup
src_val = qmap2_view[src_raw & kMask2];
}
else
{
TI src_val_0;
TI src_val_1;
if constexpr (QuantMapMode == config::QuantMapModeEnum::WarpShuffle)
{
// in-register table lookup
src_val_0 = __shfl_sync(kMaskSync, qmap3_view(0), (src_raw >> kNumBits) & kMask);
src_val_1 = __shfl_sync(kMaskSync, qmap3_view(0), (src_raw ) & kMask);
}
else
{
// normal table lookup
src_val_0 = qmap_view[(src_raw >> kNumBits) & kMask];
src_val_1 = qmap_view[(src_raw ) & kMask];
}
if constexpr (cute::is_same_v<T, cute::half_t>)
{
src_val = __halves2half2 (src_val_0, src_val_1);
}
else
{
src_val = __halves2bfloat162(src_val_0, src_val_1);
}
}
// vectorized scaling
target_vec(tgt_crd) = __hmul2(src_val, scale_vec(tgt_crd));
}
}
}
}
};
template <class SourceEngine , class SourceLayout ,
class SourceEngine2 , class SourceLayout2 ,
class TargetEngine , class TargetLayout ,
class ScaleEngine , class ScaleLayout ,
class QuantMapEngine , class QuantMapLayout ,
class QuantMapEngine2, class QuantMapLayout2,
class QuantMapEngine3, class QuantMapLayout3>
struct DequantizationTraits<SourceEngine , SourceLayout ,
SourceEngine2 , SourceLayout2 ,
TargetEngine , TargetLayout ,
ScaleEngine , ScaleLayout ,
QuantMapEngine , QuantMapLayout ,
QuantMapEngine2, QuantMapLayout2,
QuantMapEngine3, QuantMapLayout3,
cute::Int<4>,
config::QuantMapModeEnum::Marlin>
{
CUTE_DEVICE static
void
apply(
cute::Tensor<SourceEngine , SourceLayout > const& source,
cute::Tensor<SourceEngine2 , SourceLayout2 > const& source2,
cute::Tensor<TargetEngine , TargetLayout > & target,
cute::Tensor<ScaleEngine , ScaleLayout > const& scale,
cute::Tensor<QuantMapEngine , QuantMapLayout > const& qmap,
cute::Tensor<QuantMapEngine2, QuantMapLayout2> const& qmap2,
cute::Tensor<QuantMapEngine3, QuantMapLayout3> const& qmap3)
{
using TQ = cute::uint16_t;
using TQ2 = cute::uint32_t;
using T = typename TargetEngine::value_type;
using TI = cute::conditional_t<cute::is_same_v<T, cute::half_t>, __half , __nv_bfloat16 >;
using T2 = cute::conditional_t<cute::is_same_v<T, cute::half_t>, __half2, __nv_bfloat162>;
CUTE_STATIC_ASSERT(cute::is_same_v<T , cute::half_t > == true ||
cute::is_same_v<T , cute::bfloat16_t > == true);
CUTE_STATIC_ASSERT(cute::is_same_v<TQ, typename SourceEngine ::value_type> == true);
CUTE_STATIC_ASSERT(cute::is_same_v<TQ, typename SourceEngine2 ::value_type> == true);
CUTE_STATIC_ASSERT(cute::is_same_v<T , typename TargetEngine ::value_type> == true);
CUTE_STATIC_ASSERT(cute::is_same_v<T , typename ScaleEngine ::value_type> == true);
CUTE_STATIC_ASSERT(cute::is_same_v<T , typename QuantMapEngine ::value_type> == true);
CUTE_STATIC_ASSERT(cute::is_same_v<T2, typename QuantMapEngine2::value_type> == true);
CUTE_STATIC_ASSERT(cute::is_same_v<T , typename QuantMapEngine3::value_type> == true);
static constexpr int kNumBits = 4;
static constexpr int kNumPacked2 = 8;
// vectorize the source and target
auto source_vec = cute::recast<TQ2>(source);
auto source2_vec = cute::recast<TQ2>(source2); // unused
auto target_vec = cute::recast<T2 >(target);
auto scale_vec = cute::recast<T2 >(scale);
auto qmap_view = cute::recast<TI >(qmap);
auto qmap2_view = cute::recast<T2 >(qmap2);
auto qmap3_view = cute::recast<TI >(qmap3);
CUTE_STATIC_ASSERT_V(cute::size<0>(source ) == cute::size<0>(source_vec ) * cute::_2{});
CUTE_STATIC_ASSERT_V(cute::size<0>(source2) == cute::size<0>(source2_vec) * cute::_2{});
CUTE_STATIC_ASSERT_V(cute::size<0>(target ) == cute::size<0>(target_vec ) * cute::_2{});
CUTE_STATIC_ASSERT_V(cute::size<0>(scale ) == cute::size<0>(scale_vec ) * cute::_2{});
CUTE_STATIC_ASSERT_V(cute::size<1>(source ) == cute::size<1>(source_vec ));
CUTE_STATIC_ASSERT_V(cute::size<1>(source2) == cute::size<1>(source2_vec));
CUTE_STATIC_ASSERT_V(cute::size<1>(target ) == cute::size<1>(target_vec ));
CUTE_STATIC_ASSERT_V(cute::size<1>(scale ) == cute::size<1>(scale_vec ));
CUTE_STATIC_ASSERT_V(cute::size (qmap ) == cute::size (qmap_view ));
CUTE_STATIC_ASSERT_V(cute::size (qmap2 ) == cute::size (qmap2_view ));
CUTE_STATIC_ASSERT_V(cute::size (qmap3 ) == cute::size (qmap3_view ));
CUTE_UNROLL
for (int i = 0; i < cute::size<0>(source_vec); ++i)
{
CUTE_UNROLL
for (int p = 0; p < cute::size<1>(source_vec); ++p)
{
auto src_crd = cute::make_coord(i, p);
CUTE_UNROLL
for (int k4 = 0; k4 < kNumPacked2; k4 += 4)
{
auto k = k4 / 4;
auto src_raw = source_vec(src_crd) >> (k * 8);
auto src_val = marlin_utils::dequant(src_raw);
auto tgt0_crd = cute::make_coord(i, (k * 2 + 0) * cute::size<1>(source_vec) + p);
auto tgt1_crd = cute::make_coord(i, (k * 2 + 1) * cute::size<1>(source_vec) + p);
target_vec(tgt0_crd) = __hmul2(src_val[0], scale_vec(tgt0_crd));
target_vec(tgt1_crd) = __hmul2(src_val[1], scale_vec(tgt1_crd));
}
}
}
}
};
template <class SourceEngine , class SourceLayout ,
class SourceEngine2 , class SourceLayout2 ,
class TargetEngine , class TargetLayout ,
class ScaleEngine , class ScaleLayout ,
class QuantMapEngine , class QuantMapLayout ,
class QuantMapEngine2, class QuantMapLayout2,
class QuantMapEngine3, class QuantMapLayout3,
config::QuantMapModeEnum QuantMapMode>
struct DequantizationTraits<SourceEngine , SourceLayout ,
SourceEngine2 , SourceLayout2 ,
TargetEngine , TargetLayout ,
ScaleEngine , ScaleLayout ,
QuantMapEngine , QuantMapLayout ,
QuantMapEngine2, QuantMapLayout2,
QuantMapEngine3, QuantMapLayout3,
cute::Int<3>,
QuantMapMode>
{
CUTE_DEVICE static
void
apply(
cute::Tensor<SourceEngine , SourceLayout > const& source,
cute::Tensor<SourceEngine2 , SourceLayout2 > const& source2,
cute::Tensor<TargetEngine , TargetLayout > & target,
cute::Tensor<ScaleEngine , ScaleLayout > const& scale,
cute::Tensor<QuantMapEngine , QuantMapLayout > const& qmap,
cute::Tensor<QuantMapEngine2, QuantMapLayout2> const& qmap2,
cute::Tensor<QuantMapEngine3, QuantMapLayout3> const& qmap3)
{
using TQ = cute::uint16_t;
using TQ2 = cute::uint32_t;
using T = typename TargetEngine::value_type;
using TI = cute::conditional_t<cute::is_same_v<T, cute::half_t>, __half , __nv_bfloat16 >;
using T2 = cute::conditional_t<cute::is_same_v<T, cute::half_t>, __half2, __nv_bfloat162>;
CUTE_STATIC_ASSERT(cute::is_same_v<T , cute::half_t > == true ||
cute::is_same_v<T , cute::bfloat16_t > == true);
CUTE_STATIC_ASSERT(cute::is_same_v<TQ, typename SourceEngine ::value_type> == true);
CUTE_STATIC_ASSERT(cute::is_same_v<TQ, typename SourceEngine2 ::value_type> == true);
CUTE_STATIC_ASSERT(cute::is_same_v<T , typename TargetEngine ::value_type> == true);
CUTE_STATIC_ASSERT(cute::is_same_v<T , typename ScaleEngine ::value_type> == true);
CUTE_STATIC_ASSERT(cute::is_same_v<T , typename QuantMapEngine ::value_type> == true);
CUTE_STATIC_ASSERT(cute::is_same_v<T2, typename QuantMapEngine2::value_type> == true);
CUTE_STATIC_ASSERT(cute::is_same_v<T , typename QuantMapEngine3::value_type> == true);
CUTE_STATIC_ASSERT (QuantMapMode == config::QuantMapModeEnum::Vectorized);
CUTE_STATIC_ASSERT_V(cute::size<1>(source2) == cute::size<1>(source) * cute::_2{});
CUTE_STATIC_ASSERT_V(cute::size<0>(source2) == cute::size<0>(source));
static constexpr int kNumBits = 3;
static constexpr int kNumPacked2 = 10;
static constexpr cute::uint32_t kMask2 = 0x0000003f;
static constexpr cute::uint32_t kMaskF2 = 0x00000003;
// vectorize the source and target
auto source0_vec = cute::recast<TQ2>(source);
auto source1_vec = cute::recast<TQ2>(source2);
auto source2_vec = cute::recast<TQ2>(source2); // the same as source2
auto target_vec = cute::recast<T2 >(target);
auto scale_vec = cute::recast<T2 >(scale);
auto qmap_view = cute::recast<TI >(qmap);
auto qmap2_view = cute::recast<T2 >(qmap2);
auto qmap3_view = cute::recast<TI >(qmap3);
CUTE_STATIC_ASSERT_V(cute::size<0>(source ) == cute::size<0>(source0_vec) * cute::_2{});
CUTE_STATIC_ASSERT_V(cute::size<0>(source2) == cute::size<0>(source1_vec) * cute::_2{});
CUTE_STATIC_ASSERT_V(cute::size<0>(source2) == cute::size<0>(source2_vec) * cute::_2{});
CUTE_STATIC_ASSERT_V(cute::size<0>(target ) == cute::size<0>(target_vec ) * cute::_2{});
CUTE_STATIC_ASSERT_V(cute::size<0>(scale ) == cute::size<0>(scale_vec ) * cute::_2{});
CUTE_STATIC_ASSERT_V(cute::size<1>(source ) == cute::size<1>(source0_vec));
CUTE_STATIC_ASSERT_V(cute::size<1>(source2) == cute::size<1>(source1_vec));
CUTE_STATIC_ASSERT_V(cute::size<1>(source2) == cute::size<1>(source2_vec));
CUTE_STATIC_ASSERT_V(cute::size<1>(target ) == cute::size<1>(target_vec ));
CUTE_STATIC_ASSERT_V(cute::size<1>(scale ) == cute::size<1>(scale_vec ));
CUTE_STATIC_ASSERT_V(cute::size (qmap ) == cute::size (qmap_view ));
CUTE_STATIC_ASSERT_V(cute::size (qmap2 ) == cute::size (qmap2_view ));
CUTE_STATIC_ASSERT_V(cute::size (qmap3 ) == cute::size (qmap3_view ));
CUTE_UNROLL
for (int i = 0; i < cute::size<0>(source0_vec); ++i)
{
CUTE_UNROLL
for (int p = 0; p < cute::size<1>(source0_vec); ++p)
{
auto src0_crd = cute::make_coord(i, p);
auto src1_crd = cute::make_coord(i, p * 2);
auto src2_crd = cute::make_coord(i, p * 2 + 1);
CUTE_UNROLL
for (int k2 = 0; k2 < kNumPacked2; k2 += 2)
{
auto k = k2 / 2;
// using `source0_vec` for the stride, since others sahre the same value
auto tgt0_crd = cute::make_coord(i, (k * 3 + 0) * cute::size<1>(source0_vec) + p);
auto tgt1_crd = cute::make_coord(i, (k * 3 + 1) * cute::size<1>(source0_vec) + p);
auto tgt2_crd = cute::make_coord(i, (k * 3 + 2) * cute::size<1>(source0_vec) + p);
auto src0_raw = source0_vec(src0_crd) >> (k2 * kNumBits);
auto src0_val = qmap2_view [src0_raw & kMask2];
target_vec(tgt0_crd) = __hmul2 (src0_val , scale_vec(tgt0_crd));
auto src1_raw = source1_vec(src1_crd) >> (k2 * kNumBits);
auto src1_val = qmap2_view [src1_raw & kMask2];
target_vec(tgt1_crd) = __hmul2 (src1_val , scale_vec(tgt1_crd));
auto src2_raw = source2_vec(src2_crd) >> (k2 * kNumBits);
auto src2_val = qmap2_view [src2_raw & kMask2];
target_vec(tgt2_crd) = __hmul2 (src2_val , scale_vec(tgt2_crd));
}
// handle the last element
auto tgt3_crd = cute::make_coord(i, ((kNumPacked2 / 2) * 3) * cute::size<1>(source0_vec) + p);
auto src3_raw = ((((source0_vec(src0_crd) >> (kNumPacked2 * kNumBits)) & kMaskF2) ) |
(((source1_vec(src1_crd) >> (kNumPacked2 * kNumBits)) & kMaskF2) << 2) |
(((source2_vec(src2_crd) >> (kNumPacked2 * kNumBits)) & kMaskF2) << 4));
auto src3_val = qmap2_view[src3_raw & kMask2];
target_vec(tgt3_crd) = __hmul2(src3_val, scale_vec(tgt3_crd));
}
}
}
};
template <config::QuantMapModeEnum QuantMapMode,
class SourceEngine , class SourceLayout ,
class SourceEngine2 , class SourceLayout2 ,
class TargetEngine , class TargetLayout ,
class ScaleEngine , class ScaleLayout ,
class QuantMapEngine , class QuantMapLayout ,
class QuantMapEngine2, class QuantMapLayout2,
class QuantMapEngine3, class QuantMapLayout3,
class NumBits>
CUTE_DEVICE
void
dequantize(
cute::Tensor<SourceEngine , SourceLayout > const& source,
cute::Tensor<SourceEngine2 , SourceLayout2 > const& source2,
cute::Tensor<TargetEngine , TargetLayout > && target,
cute::Tensor<ScaleEngine , ScaleLayout > const& scale,
cute::Tensor<QuantMapEngine , QuantMapLayout > const& qmap,
cute::Tensor<QuantMapEngine2, QuantMapLayout2> const& qmap2,
cute::Tensor<QuantMapEngine3, QuantMapLayout3> const& qmap3,
NumBits)
{
CUTE_STATIC_ASSERT_V(cute::rank (source ) == cute::_2{}); // ((dim0, dim1), Mma_P)
CUTE_STATIC_ASSERT_V(cute::rank (source2) == cute::_2{}); // ((dim0, dim1), Mma_P * 2)
CUTE_STATIC_ASSERT_V(cute::rank (target ) == cute::_2{}); // ((dim0, dim1), Mma)
CUTE_STATIC_ASSERT_V(cute::rank (scale ) == cute::_2{}); // ((dim0, dim1), Mma)
CUTE_STATIC_ASSERT_V(cute::rank (qmap ) == cute::_1{}); // (2 ** (NumBits),)
CUTE_STATIC_ASSERT_V(cute::rank (qmap2 ) == cute::_1{}); // (2 ** (NumBits * 2),)
CUTE_STATIC_ASSERT_V(cute::rank (qmap3 ) == cute::_1{}); // (1,)
CUTE_STATIC_ASSERT_V(cute::size<0>(target) == cute::size<0>(source));
CUTE_STATIC_ASSERT_V(cute::size<0>(target) == cute::size<0>(source2));
CUTE_STATIC_ASSERT_V(cute::size<0>(target) == cute::size<0>(scale));
CUTE_STATIC_ASSERT_V(cute::size<1>(target) == cute::size<1>(scale));
CUTE_STATIC_ASSERT_V(cute::size (qmap3) == cute::_1{});
CUTE_STATIC_ASSERT (cute::is_same_v<typename SourceEngine ::value_type, cute::uint16_t> == true);
CUTE_STATIC_ASSERT (cute::is_same_v<typename SourceEngine2 ::value_type, cute::uint16_t> == true);
CUTE_STATIC_ASSERT (cute::is_same_v<typename TargetEngine ::value_type, cute::half_t > == true || cute::is_same_v<typename TargetEngine ::value_type, cute::bfloat16_t> == true);
CUTE_STATIC_ASSERT (cute::is_same_v<typename ScaleEngine ::value_type, cute::half_t > == true || cute::is_same_v<typename ScaleEngine ::value_type, cute::bfloat16_t> == true);
CUTE_STATIC_ASSERT (cute::is_same_v<typename QuantMapEngine ::value_type, cute::half_t > == true || cute::is_same_v<typename QuantMapEngine ::value_type, cute::bfloat16_t> == true);
CUTE_STATIC_ASSERT (cute::is_same_v<typename QuantMapEngine2::value_type, __half2 > == true || cute::is_same_v<typename QuantMapEngine2::value_type, __nv_bfloat162 > == true);
CUTE_STATIC_ASSERT (cute::is_same_v<typename QuantMapEngine3::value_type, cute::half_t > == true || cute::is_same_v<typename QuantMapEngine3::value_type, cute::bfloat16_t> == true);
DequantizationTraits<
SourceEngine , SourceLayout ,
SourceEngine2 , SourceLayout2 ,
TargetEngine , TargetLayout ,
ScaleEngine , ScaleLayout ,
QuantMapEngine , QuantMapLayout ,
QuantMapEngine2, QuantMapLayout2,
QuantMapEngine3, QuantMapLayout3,
NumBits,
QuantMapMode>::apply(
source,
source2,
target,
scale,
qmap,
qmap2,
qmap3);
}
} // namespace packbits_utils
|