20 template <
typename GridwiseGemm,
26 typename CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
27 typename AElementwiseOperation,
28 typename BElementwiseOperation,
29 typename CElementwiseOperation,
30 typename Block2CTileMap,
31 bool HasMainKBlockLoop>
33 #if CK_USE_LAUNCH_BOUNDS
37 const BDataType* __restrict__ p_b_grid,
38 CDataType* __restrict__ p_c_grid,
39 const AGridDesc a_grid_desc,
40 const BGridDesc b_grid_desc,
41 const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
42 c_grid_desc_mblock_mperblock_nblock_nperblock,
43 const AElementwiseOperation a_element_op,
44 const BElementwiseOperation b_element_op,
45 const CElementwiseOperation c_element_op,
46 const Block2CTileMap block_2_ctile_map)
48 #if(defined(__gfx11__) || defined(__gfx12__))
49 __shared__
char p_shared[GridwiseGemm::SharedMemTrait::lds_size];
51 GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
57 c_grid_desc_mblock_mperblock_nblock_nperblock,
68 ignore = c_grid_desc_mblock_mperblock_nblock_nperblock;
72 ignore = block_2_ctile_map;
80 typename CShuffleDataType,
85 typename CGridDesc_M_N,
86 typename AElementwiseOperation,
87 typename BElementwiseOperation,
88 typename CElementwiseOperation,
97 typename ABlockTransferThreadClusterLengths_K0_M_K1,
98 typename ABlockTransferThreadClusterArrangeOrder,
99 typename ABlockTransferSrcAccessOrder,
100 index_t ABlockTransferSrcVectorDim,
101 index_t ABlockTransferSrcScalarPerVector,
102 index_t ABlockTransferDstScalarPerVector_K1,
103 bool AThreadTransferSrcResetCoordinateAfterRun,
105 bool ABlockLdsExtraM,
106 typename BBlockTransferThreadClusterLengths_K0_N_K1,
107 typename BBlockTransferThreadClusterArrangeOrder,
108 typename BBlockTransferSrcAccessOrder,
109 index_t BBlockTransferSrcVectorDim,
110 index_t BBlockTransferSrcScalarPerVector,
111 index_t BBlockTransferDstScalarPerVector_K1,
112 bool BThreadTransferSrcResetCoordinateAfterRun,
114 bool BBlockLdsExtraN,
115 index_t CShuffleMRepeatPerShuffle,
116 index_t CShuffleNRepeatPerShuffle,
117 typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
118 index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
119 index_t NumGemmKPrefetchStage = 1,
136 static constexpr
auto MWaves = MPerBlock / (MRepeat * MPerWmma);
137 static constexpr
auto NWaves = NPerBlock / (NRepeat * NPerWmma);
138 static constexpr
auto WmmaK =
K1 == 16 ? 32 : 16;
144 NumGemmKPrefetchStage,
152 constexpr
auto a_block_desc = [&]() {
153 if constexpr(AEnableLds)
156 constexpr
auto K0PerBlock = KPerBlock /
K1;
157 constexpr
auto max_lds_align =
K1;
159 if constexpr(ABlockLdsExtraM)
173 constexpr
auto A_KRow =
I2;
174 constexpr
auto KWmmaPerblock = KPerBlock /
WmmaK;
175 constexpr
auto K0PerWmma =
WmmaK / A_KRow /
K1;
200 constexpr
auto b_block_desc = [&]() {
201 if constexpr(BEnableLds)
204 constexpr
auto K0PerBlock = KPerBlock /
K1;
205 constexpr
auto max_lds_align =
K1;
207 if constexpr(BBlockLdsExtraN)
222 constexpr
auto B_KRow =
I2;
223 constexpr
auto KWmmaPerblock = KPerBlock /
WmmaK;
224 constexpr
auto K0PerWmma =
WmmaK / B_KRow /
K1;
249 constexpr
auto a_block_copy_step = [&]() {
250 if constexpr(AEnableLds)
252 constexpr
auto K0PerBlock = KPerBlock /
K1;
258 constexpr
auto KWmmaPerBlock = KPerBlock /
WmmaK;
264 return a_block_copy_step;
269 constexpr
auto b_block_copy_step = [&]() {
270 if constexpr(BEnableLds)
272 constexpr
auto K0PerBlock = KPerBlock /
K1;
278 constexpr
auto KWmmaPerBlock = KPerBlock /
WmmaK;
284 return b_block_copy_step;
288 template <
typename ABlockDesc_>
292 constexpr
auto a_wave_desc = [&]() {
293 if constexpr(AEnableLds)
296 constexpr
auto A_K0 = ABlockDesc_{}.GetLength(
I0);
297 constexpr
auto A_K1 = ABlockDesc_{}.GetLength(
I2);
299 constexpr
auto A_KRow =
I2;
301 constexpr
auto A_KRow =
I1;
316 constexpr
auto KWmma = ABlockDesc_{}.GetLength(
I0);
317 constexpr
auto K0PerWmma = ABlockDesc_{}.GetLength(
I3);
318 constexpr
auto A_KRow = ABlockDesc_{}.GetLength(
I4);
319 constexpr
auto A_K1 = ABlockDesc_{}.GetLength(
I6);
352 template <
typename BBlockDesc_>
355 constexpr
auto b_wave_desc = [&]() {
356 if constexpr(BEnableLds)
359 constexpr
auto B_K0 = BBlockDesc_{}.GetLength(
I0);
360 constexpr
auto B_K1 = BBlockDesc_{}.GetLength(
I2);
362 constexpr
auto B_KRow =
I2;
364 constexpr
auto B_KRow =
I1;
378 constexpr
auto KWmma = BBlockDesc_{}.GetLength(
I0);
379 constexpr
auto K0PerWmma = BBlockDesc_{}.GetLength(
I3);
380 constexpr
auto B_KRow = BBlockDesc_{}.GetLength(
I4);
381 constexpr
auto B_K1 = BBlockDesc_{}.GetLength(
I6);
396 __host__ __device__
static constexpr
auto
400 constexpr
auto c_shuffle_block_desc_mshrepeat_mpershrepeat_nshrepeat_npershrepeat =
407 return c_shuffle_block_desc_mshrepeat_mpershrepeat_nshrepeat_npershrepeat;
411 template <
typename Block2CTileMap>
412 __host__ __device__
static constexpr
bool CheckValidity(
const AGridDesc& a_grid_desc,
413 const BGridDesc& b_grid_desc,
414 const CGridDesc_M_N& c_grid_desc_m_n,
415 const Block2CTileMap& block_2_ctile_map)
418 "wrong! K1 need to be known at compile-time");
420 static_assert((MPerBlock % (MPerWmma * MRepeat) == 0) &&
421 (NPerBlock % (NRepeat * NPerWmma)) == 0,
422 "Invalid tuning param!");
424 const auto GetAProblemsizeMK = [&]() {
425 if constexpr(AEnableLds)
428 a_grid_desc.GetLength(
I0) * a_grid_desc.GetLength(
I2));
432 return make_tuple(a_grid_desc.GetLength(
I1) * a_grid_desc.GetLength(
I2) *
433 a_grid_desc.GetLength(
I5),
434 a_grid_desc.GetLength(
I0) * a_grid_desc.GetLength(
I3) *
435 a_grid_desc.GetLength(
I4) * a_grid_desc.GetLength(
I6));
439 const auto GetBProblemsizeNK = [&]() {
440 if constexpr(BEnableLds)
443 b_grid_desc.GetLength(
I0) * b_grid_desc.GetLength(
I2));
447 return make_tuple(b_grid_desc.GetLength(
I1) * b_grid_desc.GetLength(
I2) *
448 b_grid_desc.GetLength(
I5),
449 b_grid_desc.GetLength(
I0) * b_grid_desc.GetLength(
I3) *
450 b_grid_desc.GetLength(
I4) * b_grid_desc.GetLength(
I6));
454 const auto M = GetAProblemsizeMK()[
I0];
455 const auto N = GetBProblemsizeNK()[
I0];
456 const auto K = GetAProblemsizeMK()[
I1];
458 if(!(M == c_grid_desc_m_n.GetLength(
I0) && N == c_grid_desc_m_n.GetLength(
I1) &&
459 K == GetBProblemsizeNK()[
I1]))
461 printf(
"A: MxK = %d x %d, B: NxK = %d x %d, C: MxN = %d x %d\n",
462 GetAProblemsizeMK()[
I0],
463 GetAProblemsizeMK()[
I1],
464 GetBProblemsizeNK()[
I0],
465 GetBProblemsizeNK()[
I1],
466 c_grid_desc_m_n.GetLength(
I0),
467 c_grid_desc_m_n.GetLength(
I1));
468 printf(
"GridwiseOp err: ProblemSize check");
472 if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K % KPerBlock == 0))
474 printf(
"GridwiseOp err: ProblemSize division");
479 const auto num_k_loop = K / KPerBlock;
481 if(!GridwiseGemmPipe::IsSupported(num_k_loop))
483 printf(
"GridwiseOp err: Pipeline not support this k_loop");
487 if(!block_2_ctile_map.CheckValidity(c_grid_desc_m_n))
495 if(!(a_grid_desc.GetElementSpaceSize() *
sizeof(ADataType) <= TwoGB &&
496 b_grid_desc.GetElementSpaceSize() *
sizeof(BDataType) <= TwoGB))
505 const index_t num_loop = K / KPerBlock;
507 return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
510 __host__ __device__
static constexpr
auto
513 const auto M = c_grid_desc_m_n.GetLength(
I0);
514 const auto N = c_grid_desc_m_n.GetLength(
I1);
516 const auto MBlock = M / MPerBlock;
517 const auto NBlock = N / NPerBlock;
526 return c_grid_desc_mblock_mperblock_nblock_nperblock;
558 .GetElementSpaceSize();
574 template <
bool HasMainKBlockLoop,
typename Block2CTileMap = DefaultBlock2CTileMap>
575 __device__
static void Run(
const ADataType* __restrict__ p_a_grid,
576 const BDataType* __restrict__ p_b_grid,
577 CDataType* __restrict__ p_c_grid,
578 void* __restrict__ p_shared,
579 const AGridDesc& a_grid_desc,
580 const BGridDesc& b_grid_desc,
582 c_grid_desc_mblock_mperblock_nblock_nperblock,
583 const AElementwiseOperation& a_element_op,
584 const BElementwiseOperation& b_element_op,
585 const CElementwiseOperation& c_element_op,
586 const Block2CTileMap& block_2_ctile_map)
591 const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
592 p_a_grid, a_grid_desc.GetElementSpaceSize());
593 const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
594 p_b_grid, b_grid_desc.GetElementSpaceSize());
595 auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
596 p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
601 if(!block_2_ctile_map.ValidCTileIndex(
603 make_tuple(c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(
I0),
604 c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(
I2))))
608 const index_t m_block_data_idx_on_grid = __builtin_amdgcn_readfirstlane(block_work_idx[
I0] * MPerBlock);
609 const index_t n_block_data_idx_on_grid = __builtin_amdgcn_readfirstlane(block_work_idx[
I1] * NPerBlock);
613 const auto K = [&](){
614 if constexpr(AEnableLds){
615 return a_grid_desc.GetLength(
I0) * a_grid_desc.GetLength(
I2);
618 return a_grid_desc.GetLength(
I0) * a_grid_desc.GetLength(
I3)
619 * a_grid_desc.GetLength(
I4) * a_grid_desc.GetLength(
I6);
626 auto a_block_trait = [&](){
628 if constexpr(AEnableLds)
630 constexpr
auto K0PerBlock = KPerBlock/
K1;
631 auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
632 static_cast<ADataType*
>(p_shared),
635 auto a_blockwise_copy =
637 AElementwiseOperation,
641 ABlockTransferThreadClusterLengths_K0_M_K1,
642 ABlockTransferThreadClusterArrangeOrder,
645 decltype(a_grid_desc),
646 decltype(a_block_desc),
647 ABlockTransferSrcAccessOrder,
649 ABlockTransferSrcVectorDim,
651 ABlockTransferSrcScalarPerVector,
652 ABlockTransferDstScalarPerVector_K1,
655 AThreadTransferSrcResetCoordinateAfterRun,
657 NumGemmKPrefetchStage>(
665 return make_tuple(a_block_buf, a_blockwise_copy);
671 constexpr
auto KWmmaPerBlock = KPerBlock /
WmmaK;
672 constexpr
auto K0PerWmma =
WmmaK/2/K1Value;
673 auto a_block_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ADataType>(
674 a_block_desc.GetElementSpaceSize());
677 auto a_blockwise_copy =
680 decltype(a_grid_desc),
681 decltype(a_block_desc),
691 ABlockTransferSrcScalarPerVector,
692 AThreadTransferSrcResetCoordinateAfterRun,
696 m_block_data_idx_on_grid/(
MWaves * MPerWmma),
703 return make_tuple(a_block_buf, a_blockwise_copy);
707 auto b_block_trait = [&](){
708 if constexpr(BEnableLds)
710 constexpr
auto K0PerBlock = KPerBlock/
K1;
711 auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
715 auto b_blockwise_copy =
717 BElementwiseOperation,
721 BBlockTransferThreadClusterLengths_K0_N_K1,
722 BBlockTransferThreadClusterArrangeOrder,
725 decltype(b_grid_desc),
726 decltype(b_block_desc),
727 BBlockTransferSrcAccessOrder,
729 BBlockTransferSrcVectorDim,
731 BBlockTransferSrcScalarPerVector,
732 BBlockTransferDstScalarPerVector_K1,
735 BThreadTransferSrcResetCoordinateAfterRun,
737 NumGemmKPrefetchStage>(
745 return make_tuple(b_block_buf, b_blockwise_copy);
751 constexpr
auto KWmmaPerBlock = KPerBlock /
WmmaK;
752 constexpr
auto K0PerWmma =
WmmaK/2/K1Value;
753 auto b_block_buf = make_static_buffer<AddressSpaceEnum::Vgpr, BDataType>(
754 b_block_desc.GetElementSpaceSize());
757 auto b_blockwise_copy =
760 decltype(b_grid_desc),
761 decltype(b_block_desc),
771 BBlockTransferSrcScalarPerVector,
772 BThreadTransferSrcResetCoordinateAfterRun,
776 n_block_data_idx_on_grid/(
NWaves * NPerWmma),
783 return make_tuple(b_block_buf, b_blockwise_copy);
787 auto a_block_buf = a_block_trait()[
I0];
788 auto a_blockwise_copy = a_block_trait()[
I1];
790 auto b_block_buf = b_block_trait()[
I0];
791 auto b_blockwise_copy = b_block_trait()[
I1];
796 auto blockwise_gemm =
823 const index_t KBlockMainLoop = __builtin_amdgcn_readfirstlane(K / KPerBlock);
824 GridwiseGemmPipe::template Run<HasMainKBlockLoop>(a_grid_desc,
829 a_block_slice_copy_step,
835 b_block_slice_copy_step,
843 constexpr
auto c_thread_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs =
844 blockwise_gemm.GetCThreadDescriptor_MRepeat_MWave_MSubGroup_NRepeat_NWave_NThreadPerSubGroup_MAccVgprs();
847 constexpr
auto c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs_tmp =
848 blockwise_gemm.GetCBlockDescriptor_MRepeat_MWave_MSubGroup_NRepeat_NWave_NThreadPerSubGroup_MAccVgprs();
850 constexpr
auto MWave = c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs_tmp.GetLength(
I1);
851 constexpr
auto MSubGroup = c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs_tmp.GetLength(
I2);
852 constexpr
auto NWave = c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs_tmp.GetLength(
I4);
853 constexpr
auto NThreadPerSubGroup = c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs_tmp.GetLength(
I5);
854 constexpr
auto MAccVgprs = c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs_tmp.GetLength(
I6);
857 constexpr
auto c_shuffle_block_desc_mshrepeat_mpershrepeat_nshrepeat_npershrepeat =
860 auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
865 c_shuffle_block_desc_mshrepeat_mpershrepeat_nshrepeat_npershrepeat,
877 NThreadPerSubGroup))),
883 const auto c_thread_mtx_on_block = blockwise_gemm.CalculateCThreadOriginDataIndex(
I0,
I0);
885 const index_t m_thread_data_on_block = c_thread_mtx_on_block[
I0];
886 const index_t n_thread_data_on_block = c_thread_mtx_on_block[
I1];
888 const auto m_thread_data_on_block_to_mrepeat_mwave_msubgroup_maccvgprs_adaptor =
894 const auto n_thread_data_on_block_to_nrepeat_nwave_nthreadpersubgroup_adaptor =
900 const auto m_thread_data_on_block_idx = m_thread_data_on_block_to_mrepeat_mwave_msubgroup_maccvgprs_adaptor.CalculateBottomIndex(
903 const auto n_thread_data_on_block_idx = n_thread_data_on_block_to_nrepeat_nwave_nthreadpersubgroup_adaptor.CalculateBottomIndex(
907 auto c_thread_copy_vgpr_to_lds =
910 decltype(c_thread_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs),
911 decltype(c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs),
916 CShuffleNRepeatPerShuffle,
926 c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs,
928 m_thread_data_on_block_idx[
I1],
929 m_thread_data_on_block_idx[
I2],
931 n_thread_data_on_block_idx[
I1],
932 n_thread_data_on_block_idx[
I2],
933 m_thread_data_on_block_idx[
I3]),
939 CElementwiseOperation,
940 CGlobalMemoryDataOperation,
942 CShuffleMRepeatPerShuffle * MWave * MPerWmma,
944 CShuffleNRepeatPerShuffle * NWave * NPerWmma>,
945 CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
949 decltype(c_shuffle_block_desc_mshrepeat_mpershrepeat_nshrepeat_npershrepeat),
950 decltype(c_grid_desc_mblock_mperblock_nblock_nperblock),
953 CShuffleBlockTransferScalarPerVector_NPerBlock,
956 {c_shuffle_block_desc_mshrepeat_mpershrepeat_nshrepeat_npershrepeat,
958 c_grid_desc_mblock_mperblock_nblock_nperblock,
964 constexpr
auto sfc_c_vgpr =
970 CShuffleNRepeatPerShuffle,
976 constexpr
auto sfc_c_global =
980 CShuffleMRepeatPerShuffle * MWave * MPerWmma,
982 CShuffleNRepeatPerShuffle * NWave * NPerWmma>>{};
984 constexpr
index_t num_access = sfc_c_vgpr.GetNumOfAccess();
986 static_assert(num_access == sfc_c_global.GetNumOfAccess(),
"wrong!");
993 c_thread_copy_vgpr_to_lds.Run(c_thread_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs,
994 sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
996 c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs,
997 c_shuffle_block_buf);
1003 c_shuffle_block_copy_lds_to_global.Run(
1004 c_shuffle_block_desc_mshrepeat_mpershrepeat_nshrepeat_npershrepeat,
1005 c_shuffle_block_buf,
1006 c_grid_desc_mblock_mperblock_nblock_nperblock,
1009 if constexpr(access_id < num_access - 1)
1011 constexpr
auto c_global_step = sfc_c_global.GetForwardStep(access_id);
1014 c_shuffle_block_copy_lds_to_global.MoveDstSliceWindow(
1015 c_grid_desc_mblock_mperblock_nblock_nperblock, c_global_step);
#define CK_MIN_BLOCK_PER_CU
Definition: ck.hpp:30
#define CK_MAX_THREAD_PER_BLOCK
Definition: ck.hpp:29
__host__ constexpr __device__ auto integer_least_multiple(X x, Y y)
Definition: math.hpp:78
__host__ constexpr __device__ T max(T x)
Definition: math.hpp:84
__host__ constexpr __device__ auto make_multi_index(Xs &&... xs)
Definition: array_multi_index.hpp:15
__host__ constexpr __device__ auto make_naive_tensor_descriptor(const Tuple< Lengths... > &lengths, const Tuple< Strides... > &strides)
Definition: tensor_descriptor_helper.hpp:49
constexpr auto GridwiseGemmPipeline_Selector()
Definition: gridwise_gemm_pipeline_selector.hpp:31
InMemoryDataOperationEnum
Definition: ck.hpp:276
__host__ constexpr __device__ auto make_naive_tensor_descriptor_packed(const Tuple< Lengths... > &lengths)
Definition: tensor_descriptor_helper.hpp:101
__host__ constexpr __device__ auto make_merge_transform(const LowLengths &low_lengths)
Definition: multi_index_transform_helper.hpp:55
__global__ void kernel_gemm_wmma(const ADataType *__restrict__ p_a_grid, const BDataType *__restrict__ p_b_grid, CDataType *__restrict__ p_c_grid, const AGridDesc a_grid_desc, const BGridDesc b_grid_desc, const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock c_grid_desc_mblock_mperblock_nblock_nperblock, const AElementwiseOperation a_element_op, const BElementwiseOperation b_element_op, const CElementwiseOperation c_element_op, const Block2CTileMap block_2_ctile_map)
Definition: gridwise_gemm_wmma.hpp:36
int64_t long_index_t
Definition: ck.hpp:299
__host__ constexpr __device__ auto make_naive_tensor_descriptor_aligned(const Tuple< Lengths... > &lengths, Align align)
Definition: tensor_descriptor_helper.hpp:132
__host__ constexpr __device__ auto make_single_stage_tensor_adaptor(const Transforms &transforms, LowerDimensionOldTopIdss, UpperDimensionNewTopIdss)
Definition: tensor_adaptor.hpp:425
__host__ constexpr __device__ auto make_freeze_transform(const LowerIndex &low_idx)
Definition: multi_index_transform_helper.hpp:98
constexpr detail::ignore_t ignore
Definition: ignore.hpp:20
__device__ index_t get_block_1d_id()
Definition: get_id.hpp:58
__host__ constexpr __device__ auto make_pass_through_transform(const LowLength &low_length)
Definition: multi_index_transform_helper.hpp:12
__host__ constexpr __device__ auto make_tuple(Xs &&... xs)
Definition: tuple.hpp:211
remove_cv_t< remove_reference_t< T > > remove_cvref_t
Definition: type.hpp:297
__host__ constexpr __device__ auto make_unmerge_transform(const UpLengths &up_lengths, integral_constant< bool, Use24BitIntegerCalculation >=integral_constant< bool, false >{})
Definition: multi_index_transform_helper.hpp:90
LoopScheduler
Definition: loop_scheduler.hpp:15
int32_t index_t
Definition: ck.hpp:298
__device__ index_t get_thread_local_1d_id()
Definition: get_id.hpp:52
__host__ constexpr __device__ auto transform_tensor_descriptor(const OldTensorDescriptor &old_tensor_desc, const NewTransforms &new_transforms, NewLowerDimensionOldVisibleIdss, NewUpperDimensionNewVisibleIdss)
Definition: tensor_descriptor.hpp:319
__device__ void block_sync_lds()
Definition: synchronization.hpp:10
PipelineVersion
Definition: gridwise_gemm_pipeline_selector.hpp:18
typename remove_cv< T >::type remove_cv_t
Definition: type.hpp:295
constexpr LoopScheduler make_default_loop_scheduler()
Definition: loop_scheduler.hpp:20
const GenericPointer< typename T::ValueType > T2 value
Definition: pointer.h:1350
Definition: block_to_ctile_map.hpp:260
Definition: blockwise_gemm_wmma.hpp:550
__host__ constexpr __device__ auto & GetCThreadBuffer()
Definition: blockwise_gemm_wmma.hpp:585
Definition: gridwise_gemm_wmma.hpp:538
static constexpr auto c_shuffle_block_space_size
Definition: gridwise_gemm_wmma.hpp:556
static constexpr auto b_block_space_size_aligned
Definition: gridwise_gemm_wmma.hpp:547
static constexpr auto max_lds_align
Definition: gridwise_gemm_wmma.hpp:541
static constexpr auto c_shuffle_block_space_offset
Definition: gridwise_gemm_wmma.hpp:560
static constexpr auto lds_size
Definition: gridwise_gemm_wmma.hpp:562
static constexpr auto a_block_space_size_aligned
Definition: gridwise_gemm_wmma.hpp:543
static constexpr auto a_block_space_offset
Definition: gridwise_gemm_wmma.hpp:552
static constexpr auto b_block_space_offset
Definition: gridwise_gemm_wmma.hpp:553
Definition: gridwise_gemm_wmma.hpp:123
__host__ static constexpr __device__ auto MakeDefaultBlock2CTileMap(const CGridDesc_M_N &c_grid_desc_m_n, index_t, index_t)
Definition: gridwise_gemm_wmma.hpp:530
remove_cvref_t< decltype(MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(CGridDesc_M_N{}))> CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
Definition: gridwise_gemm_wmma.hpp:570
remove_cvref_t< decltype(GridwiseGemmPipeline_Selector< PipelineVer, NumGemmKPrefetchStage, LoopSched, AEnableLds, BEnableLds >())> GridwiseGemmPipe
Definition: gridwise_gemm_wmma.hpp:147
static constexpr auto I6
Definition: gridwise_gemm_wmma.hpp:130
static __device__ void Run(const ADataType *__restrict__ p_a_grid, const BDataType *__restrict__ p_b_grid, CDataType *__restrict__ p_c_grid, void *__restrict__ p_shared, const AGridDesc &a_grid_desc, const BGridDesc &b_grid_desc, const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock &c_grid_desc_mblock_mperblock_nblock_nperblock, const AElementwiseOperation &a_element_op, const BElementwiseOperation &b_element_op, const CElementwiseOperation &c_element_op, const Block2CTileMap &block_2_ctile_map)
Definition: gridwise_gemm_wmma.hpp:575
static constexpr auto I5
Definition: gridwise_gemm_wmma.hpp:129
__host__ static constexpr __device__ auto MakeBBlockDescriptor()
Definition: gridwise_gemm_wmma.hpp:198
__host__ static constexpr __device__ auto MakeBWaveDescriptor(const BBlockDesc_ &)
Definition: gridwise_gemm_wmma.hpp:353
__host__ static constexpr __device__ bool CheckValidity(const AGridDesc &a_grid_desc, const BGridDesc &b_grid_desc, const CGridDesc_M_N &c_grid_desc_m_n, const Block2CTileMap &block_2_ctile_map)
Definition: gridwise_gemm_wmma.hpp:412
__host__ static constexpr __device__ auto MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const CGridDesc_M_N &c_grid_desc_m_n)
Definition: gridwise_gemm_wmma.hpp:511
__host__ static constexpr __device__ auto GetCShuffleBlockDescriptor_MShRepeat_MPerShRepeat_NShRepeat_NPerShRepeat()
Definition: gridwise_gemm_wmma.hpp:398
__host__ static constexpr __device__ bool CalculateHasMainKBlockLoop(index_t K)
Definition: gridwise_gemm_wmma.hpp:503
static constexpr auto I7
Definition: gridwise_gemm_wmma.hpp:131
ThisThreadBlock< BlockSize > ThisThreadBlock
Definition: gridwise_gemm_wmma.hpp:140
__host__ static constexpr __device__ auto MakeBBlockSliceCopyStep()
Definition: gridwise_gemm_wmma.hpp:267
static constexpr auto K1
Definition: gridwise_gemm_wmma.hpp:134
static constexpr auto I4
Definition: gridwise_gemm_wmma.hpp:128
static constexpr auto I1
Definition: gridwise_gemm_wmma.hpp:125
static constexpr auto MWaves
Definition: gridwise_gemm_wmma.hpp:136
__host__ static constexpr __device__ auto MakeABlockSliceCopyStep()
Definition: gridwise_gemm_wmma.hpp:247
static constexpr auto I2
Definition: gridwise_gemm_wmma.hpp:126
remove_cvref_t< decltype(MakeDefaultBlock2CTileMap(CGridDesc_M_N{}, 1, 1))> DefaultBlock2CTileMap
Definition: gridwise_gemm_wmma.hpp:572
__host__ static constexpr __device__ auto MakeAWaveDescriptor(const ABlockDesc_ &)
Definition: gridwise_gemm_wmma.hpp:289
static constexpr auto I3
Definition: gridwise_gemm_wmma.hpp:127
__host__ static constexpr __device__ auto MakeABlockDescriptor()
Definition: gridwise_gemm_wmma.hpp:150
static constexpr auto NWaves
Definition: gridwise_gemm_wmma.hpp:137
static constexpr auto I0
Definition: gridwise_gemm_wmma.hpp:124
static constexpr auto WmmaK
Definition: gridwise_gemm_wmma.hpp:138
Definition: sequence.hpp:43
Definition: tensor_space_filling_curve.hpp:20
Blockwise data transfer.
Definition: thread_group_tensor_slice_transfer_v4r1.hpp:46
Definition: thread_group_tensor_slice_transfer_v6r1.hpp:34
Definition: threadwise_tensor_slice_transfer.hpp:39
Helper structure that facilitates transfer of source (grid) data to destination threads.
Definition: threadwise_tensor_slice_transfer.hpp:234
Definition: integral_constant.hpp:20
Definition: is_known_at_compile_time.hpp:14
Definition: functional2.hpp:33
Definition: unary_element_wise_operation.hpp:334