19 template <
typename GridwiseGemm,
22 typename ABK0MK1GridDesc,
23 typename BBK0NK1GridDesc,
24 typename CM0N0M1N1M2M3M4N2GridDesc,
25 typename AElementwiseOperation,
26 typename BElementwiseOperation,
27 typename CElementwiseOperation,
28 typename CBlockClusterAdaptor,
29 bool HasMainKBlockLoop>
31 #if CK_USE_LAUNCH_BOUNDS
35 const FloatAB* __restrict__ p_b_grid,
36 FloatC* __restrict__ p_c_grid,
37 const ABK0MK1GridDesc a_b_k0_m_k1_grid_desc,
38 const BBK0NK1GridDesc b_b_k0_n_k1_grid_desc,
39 const CM0N0M1N1M2M3M4N2GridDesc c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc,
40 const AElementwiseOperation a_element_op,
41 const BElementwiseOperation b_element_op,
42 const CElementwiseOperation c_element_op,
43 const CBlockClusterAdaptor c_block_cluster_adaptor)
45 #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
47 constexpr
index_t shared_block_size =
48 GridwiseGemm::GetSharedMemoryNumberOfByte() /
sizeof(FloatAB);
50 __shared__ FloatAB p_shared_block[shared_block_size];
52 GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
56 a_b_k0_m_k1_grid_desc,
57 b_b_k0_n_k1_grid_desc,
58 c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc,
62 c_block_cluster_adaptor);
67 ignore = a_b_k0_m_k1_grid_desc;
68 ignore = b_b_k0_n_k1_grid_desc;
69 ignore = c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc;
73 ignore = c_block_cluster_adaptor;
82 typename ABK0MK1GridDesc,
83 typename BBK0NK1GridDesc,
85 typename AElementwiseOperation,
86 typename BElementwiseOperation,
87 typename CElementwiseOperation,
96 typename ABlockTransferThreadClusterLengths_K0_M_K1,
97 typename ABlockTransferThreadClusterArrangeOrder,
98 typename ABlockTransferSrcAccessOrder,
99 index_t ABlockTransferSrcVectorDim,
100 index_t ABlockTransferSrcScalarPerVector,
101 index_t ABlockTransferDstScalarPerVector_K1,
102 bool AThreadTransferSrcResetCoordinateAfterRun,
103 bool ABlockLdsExtraM,
104 typename BBlockTransferThreadClusterLengths_K0_N_K1,
105 typename BBlockTransferThreadClusterArrangeOrder,
106 typename BBlockTransferSrcAccessOrder,
107 index_t BBlockTransferSrcVectorDim,
108 index_t BBlockTransferSrcScalarPerVector,
109 index_t BBlockTransferDstScalarPerVector_K1,
110 bool BThreadTransferSrcResetCoordinateAfterRun,
111 bool BBlockLdsExtraN,
112 typename CThreadTransferSrcDstAccessOrder,
113 index_t CThreadTransferSrcDstVectorDim,
114 index_t CThreadTransferDstScalarPerVector>
133 constexpr
auto max_lds_align =
K1;
136 constexpr
auto a_k0_m_k1_block_desc = [&]() {
137 if constexpr(ABlockLdsExtraM)
151 constexpr
auto b_k0_n_k1_block_desc = [&]() {
152 if constexpr(BBlockLdsExtraN)
166 constexpr
auto a_block_space_size =
169 constexpr
auto b_block_space_size =
172 return (a_block_space_size + b_block_space_size) *
sizeof(FloatAB);
176 template <
typename Block2CTileMap>
177 __host__ __device__
static constexpr
bool
179 const BBK0NK1GridDesc& b_b_k0_n_k1_grid_desc,
180 const CMNGridDesc& c_m_n_grid_desc,
181 const Block2CTileMap& block_2_ctile_map)
184 "wrong! K1 need to be known at compile-time");
186 static_assert((MPerBlock % (MPerXDL * MRepeat) == 0) &&
187 (NPerBlock % (NRepeat * NPerXDL)) == 0,
188 "Invalid tuning param!");
190 const auto M = a_b_k0_m_k1_grid_desc.GetLength(
I2);
191 const auto N = b_b_k0_n_k1_grid_desc.GetLength(
I2);
192 const auto K0 = a_b_k0_m_k1_grid_desc.GetLength(
I1);
193 const auto KBatch = a_b_k0_m_k1_grid_desc.GetLength(
I0);
195 if(!(M == c_m_n_grid_desc.GetLength(
I0) && N == c_m_n_grid_desc.GetLength(
I1) &&
196 K0 == b_b_k0_n_k1_grid_desc.GetLength(
I1) &&
197 K1 == a_b_k0_m_k1_grid_desc.GetLength(
I3) &&
198 K1 == b_b_k0_n_k1_grid_desc.GetLength(
I3) &&
199 KBatch == b_b_k0_n_k1_grid_desc.GetLength(
I0)))
202 if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K0 % K0PerBlock == 0))
205 if(!block_2_ctile_map.CheckValidity(c_m_n_grid_desc))
216 const bool has_main_k0_block_loop = K0 > K0PerBlock;
218 return has_main_k0_block_loop;
221 __host__ __device__
static constexpr
auto
224 constexpr
auto max_lds_align =
K1;
227 constexpr
auto a_k0_m_k1_block_desc = [&]() {
228 if constexpr(ABlockLdsExtraM)
242 constexpr
auto b_k0_n_k1_block_desc = [&]() {
243 if constexpr(BBlockLdsExtraN)
256 using BlockwiseGemm =
260 decltype(a_k0_m_k1_block_desc),
261 decltype(b_k0_n_k1_block_desc),
268 return BlockwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_m_n_grid_desc);
276 c_m_n_grid_desc, 8, KBatch);
282 template <
bool HasMainKBlockLoop>
283 __device__
static void Run(
const FloatAB* __restrict__ p_a_grid,
284 const FloatAB* __restrict__ p_b_grid,
285 FloatC* __restrict__ p_c_grid,
286 FloatAB* __restrict__ p_shared_block,
287 const ABK0MK1GridDesc& a_b_k0_m_k1_grid_desc,
288 const BBK0NK1GridDesc& b_b_k0_n_k1_grid_desc,
290 const AElementwiseOperation& a_element_op,
291 const BElementwiseOperation& b_element_op,
292 const CElementwiseOperation& c_element_op,
295 const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
296 p_a_grid, a_b_k0_m_k1_grid_desc.GetElementSpaceSize());
297 const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
298 p_b_grid, b_b_k0_n_k1_grid_desc.GetElementSpaceSize());
299 auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
300 p_c_grid, c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc.GetElementSpaceSize());
302 const auto K0 = a_b_k0_m_k1_grid_desc.GetLength(
I1);
305 const auto block_work_idx =
308 if(!c_block_cluster_adaptor.ValidCTileIndex(
310 make_tuple(c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc.GetLength(
I0),
311 c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc.GetLength(
I1))))
316 const index_t k_batch_id = block_work_idx[
I0];
319 const index_t m_block_data_idx_on_grid =
320 __builtin_amdgcn_readfirstlane(block_work_idx[
I1] * MPerBlock);
322 const index_t n_block_data_idx_on_grid =
323 __builtin_amdgcn_readfirstlane(block_work_idx[
I2] * NPerBlock);
326 constexpr
auto max_lds_align =
K1;
329 constexpr
auto a_k0_m_k1_block_desc = [&]() {
330 if constexpr(ABlockLdsExtraM)
343 constexpr
auto a_b_k0_m_k1_block_desc = [&]() {
344 if constexpr(ABlockLdsExtraM)
361 constexpr
auto b_k0_n_k1_block_desc = [&]() {
362 if constexpr(BBlockLdsExtraN)
375 constexpr
auto b_b_k0_n_k1_block_desc = [&]() {
376 if constexpr(BBlockLdsExtraN)
393 auto a_blockwise_copy =
395 AElementwiseOperation,
399 ABlockTransferThreadClusterLengths_K0_M_K1,
400 ABlockTransferThreadClusterArrangeOrder,
403 decltype(a_b_k0_m_k1_grid_desc),
404 decltype(a_b_k0_m_k1_block_desc),
405 ABlockTransferSrcAccessOrder,
407 ABlockTransferSrcVectorDim,
409 ABlockTransferSrcScalarPerVector,
410 ABlockTransferDstScalarPerVector_K1,
413 AThreadTransferSrcResetCoordinateAfterRun,
415 a_b_k0_m_k1_grid_desc,
418 a_b_k0_m_k1_block_desc,
423 auto b_blockwise_copy =
425 BElementwiseOperation,
429 BBlockTransferThreadClusterLengths_K0_N_K1,
430 BBlockTransferThreadClusterArrangeOrder,
433 decltype(b_b_k0_n_k1_grid_desc),
434 decltype(b_b_k0_n_k1_block_desc),
435 BBlockTransferSrcAccessOrder,
437 BBlockTransferSrcVectorDim,
439 BBlockTransferSrcScalarPerVector,
440 BBlockTransferDstScalarPerVector_K1,
443 BThreadTransferSrcResetCoordinateAfterRun,
445 b_b_k0_n_k1_grid_desc,
448 b_b_k0_n_k1_block_desc,
460 auto blockwise_gemm =
464 decltype(a_k0_m_k1_block_desc),
465 decltype(b_k0_n_k1_block_desc),
472 auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
475 constexpr
auto a_block_space_size =
478 FloatAB* p_a_block = p_shared_block;
479 FloatAB* p_b_block = p_shared_block + a_block_space_size;
481 constexpr
auto a_block_slice_copy_step =
make_multi_index(0, K0PerBlock, 0, 0);
482 constexpr
auto b_block_slice_copy_step =
make_multi_index(0, K0PerBlock, 0, 0);
484 auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
485 p_a_block, a_k0_m_k1_block_desc.GetElementSpaceSize());
486 auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
487 p_b_block, b_k0_n_k1_block_desc.GetElementSpaceSize());
491 a_blockwise_copy.RunRead(a_b_k0_m_k1_grid_desc, a_grid_buf);
492 b_blockwise_copy.
RunRead(b_b_k0_n_k1_grid_desc, b_grid_buf);
494 a_blockwise_copy.RunWrite(a_b_k0_m_k1_block_desc, a_block_buf);
495 b_blockwise_copy.
RunWrite(b_b_k0_n_k1_block_desc, b_block_buf);
499 c_thread_buf.Clear();
502 if constexpr(HasMainKBlockLoop)
504 index_t k0_block_data_begin = 0;
508 a_blockwise_copy.MoveSrcSliceWindow(a_b_k0_m_k1_grid_desc, a_block_slice_copy_step);
511 a_blockwise_copy.RunRead(a_b_k0_m_k1_grid_desc, a_grid_buf);
515 b_blockwise_copy.
RunRead(b_b_k0_n_k1_grid_desc, b_grid_buf);
517 blockwise_gemm.
Run(a_block_buf, b_block_buf, c_thread_buf);
521 a_blockwise_copy.RunWrite(a_b_k0_m_k1_block_desc, a_block_buf);
522 b_blockwise_copy.
RunWrite(b_b_k0_n_k1_block_desc, b_block_buf);
524 k0_block_data_begin += K0PerBlock;
525 }
while(k0_block_data_begin < (K0 - K0PerBlock));
532 blockwise_gemm.
Run(a_block_buf, b_block_buf, c_thread_buf);
537 constexpr
auto c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc =
538 blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
540 constexpr
auto M0 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(
I0);
541 constexpr
auto N0 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(
I1);
542 constexpr
auto M1 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(
I2);
543 constexpr
auto N1 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(
I3);
544 constexpr
auto M2 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(
I4);
545 constexpr
auto M3 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(
I5);
546 constexpr
auto M4 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(
I6);
547 constexpr
auto N2 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(
I7);
549 constexpr
auto c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc =
555 const auto c_thread_mtx_on_block =
556 blockwise_gemm.CalculateCThreadOriginDataIndex(
I0,
I0,
I0,
I0);
558 const index_t m_thread_data_on_grid =
559 m_block_data_idx_on_grid + c_thread_mtx_on_block[
I0];
561 const index_t n_thread_data_on_grid =
562 n_block_data_idx_on_grid + c_thread_mtx_on_block[
I1];
564 const auto m_thread_data_on_grid_to_m0_m1_m2_m3_m4_adaptor =
570 const auto m_thread_data_on_grid_idx =
571 m_thread_data_on_grid_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
579 const auto n_thread_data_on_grid_idx =
580 n_thread_data_on_grid_to_n0_n1_n2_adaptor.CalculateBottomIndex(
586 decltype(c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc),
587 decltype(c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc),
588 CElementwiseOperation,
590 CThreadTransferSrcDstAccessOrder,
591 CThreadTransferSrcDstVectorDim,
592 CThreadTransferDstScalarPerVector,
593 CGlobalMemoryDataOperation,
597 c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc,
599 n_thread_data_on_grid_idx[
I0],
600 m_thread_data_on_grid_idx[
I1],
601 n_thread_data_on_grid_idx[
I1],
602 m_thread_data_on_grid_idx[
I2],
603 m_thread_data_on_grid_idx[
I3],
604 m_thread_data_on_grid_idx[
I4],
605 n_thread_data_on_grid_idx[
I2]),
608 c_thread_copy.Run(c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc,
611 c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc,
#define CK_MIN_BLOCK_PER_CU
Definition: ck.hpp:34
#define CK_MAX_THREAD_PER_BLOCK
Definition: ck.hpp:33
__host__ constexpr __device__ auto integer_least_multiple(X x, Y y)
Definition: math.hpp:78
__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
InMemoryDataOperationEnum
Definition: ck.hpp:267
__global__ void kernel_gemm_xdlops_v2r4(const FloatAB *__restrict__ p_a_grid, const FloatAB *__restrict__ p_b_grid, FloatC *__restrict__ p_c_grid, const ABK0MK1GridDesc a_b_k0_m_k1_grid_desc, const BBK0NK1GridDesc b_b_k0_n_k1_grid_desc, const CM0N0M1N1M2M3M4N2GridDesc c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc, const AElementwiseOperation a_element_op, const BElementwiseOperation b_element_op, const CElementwiseOperation c_element_op, const CBlockClusterAdaptor c_block_cluster_adaptor)
Definition: gridwise_gemm_xdlops_v2r4.hpp:34
__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
__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:429
constexpr detail::ignore_t ignore
Definition: ignore.hpp:20
__device__ index_t get_block_1d_id()
Definition: get_id.hpp:22
__host__ constexpr __device__ auto make_tuple(Xs &&... xs)
Definition: tuple.hpp:211
int32_t index_t
Definition: ck.hpp:289
__device__ void block_sync_lds()
Definition: synchronization.hpp:10
typename remove_cv< T >::type remove_cv_t
Definition: type.hpp:298
Definition: block_to_ctile_map.hpp:539
Definition: blockwise_gemm_smfmac_xdlops.hpp:44
Definition: gridwise_gemm_xdlops_v2r4.hpp:116
static constexpr auto I3
Definition: gridwise_gemm_xdlops_v2r4.hpp:120
ThisThreadBlock< BlockSize > ThisThreadBlock
Definition: gridwise_gemm_xdlops_v2r4.hpp:129
static constexpr auto I6
Definition: gridwise_gemm_xdlops_v2r4.hpp:123
__host__ static constexpr __device__ auto MakeCBlockClusterAdaptor(const CMNGridDesc &c_m_n_grid_desc, index_t, index_t, index_t KBatch)
Definition: gridwise_gemm_xdlops_v2r4.hpp:272
static constexpr auto K1
Definition: gridwise_gemm_xdlops_v2r4.hpp:127
static constexpr auto I5
Definition: gridwise_gemm_xdlops_v2r4.hpp:122
static constexpr auto I7
Definition: gridwise_gemm_xdlops_v2r4.hpp:124
static __device__ void Run(const FloatAB *__restrict__ p_a_grid, const FloatAB *__restrict__ p_b_grid, FloatC *__restrict__ p_c_grid, FloatAB *__restrict__ p_shared_block, const ABK0MK1GridDesc &a_b_k0_m_k1_grid_desc, const BBK0NK1GridDesc &b_b_k0_n_k1_grid_desc, const CM0N0M1N1M2M3M4N2GridDesc &c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc, const AElementwiseOperation &a_element_op, const BElementwiseOperation &b_element_op, const CElementwiseOperation &c_element_op, const CBlockClusterAdaptor &c_block_cluster_adaptor)
Definition: gridwise_gemm_xdlops_v2r4.hpp:283
__host__ static constexpr __device__ auto MakeCM0N0M1N1M2M3M4N2GridDescriptor(const CMNGridDesc &c_m_n_grid_desc)
Definition: gridwise_gemm_xdlops_v2r4.hpp:222
static constexpr auto I4
Definition: gridwise_gemm_xdlops_v2r4.hpp:121
__host__ static constexpr __device__ index_t GetSharedMemoryNumberOfByte()
Definition: gridwise_gemm_xdlops_v2r4.hpp:131
decltype(MakeCBlockClusterAdaptor(CMNGridDesc{}, 1, 1, 1)) CBlockClusterAdaptor
Definition: gridwise_gemm_xdlops_v2r4.hpp:280
static constexpr auto I1
Definition: gridwise_gemm_xdlops_v2r4.hpp:118
static constexpr auto I0
Definition: gridwise_gemm_xdlops_v2r4.hpp:117
__host__ static constexpr __device__ bool CalculateHasMainK0BlockLoop(index_t K0)
Definition: gridwise_gemm_xdlops_v2r4.hpp:214
static constexpr auto I2
Definition: gridwise_gemm_xdlops_v2r4.hpp:119
__host__ static constexpr __device__ bool CheckValidity(const ABK0MK1GridDesc &a_b_k0_m_k1_grid_desc, const BBK0NK1GridDesc &b_b_k0_n_k1_grid_desc, const CMNGridDesc &c_m_n_grid_desc, const Block2CTileMap &block_2_ctile_map)
Definition: gridwise_gemm_xdlops_v2r4.hpp:178
decltype(MakeCM0N0M1N1M2M3M4N2GridDescriptor(CMNGridDesc{})) CM0N0M1N1M2M3M4N2GridDesc
Definition: gridwise_gemm_xdlops_v2r4.hpp:279
Definition: sequence.hpp:43
Blockwise data transfer.
Definition: thread_group_tensor_slice_transfer_v4r1.hpp:46
__device__ void Run(const SrcDesc &src_desc, const SrcBuffer &src_buf, const DstDesc &dst_desc, DstBuffer &dst_buf, Number< ThreadScratchId > thread_scratch_id)
Definition: thread_group_tensor_slice_transfer_v4r1.hpp:137
__device__ void RunRead(const SrcDesc &src_desc, const SrcBuffer &src_buf, Number< ThreadScratchId > thread_scratch_id=Number< ThreadScratchId >{})
Definition: thread_group_tensor_slice_transfer_v4r1.hpp:113
__device__ void MoveSrcSliceWindow(const SrcDesc &src_desc, const Index &step)
Definition: thread_group_tensor_slice_transfer_v4r1.hpp:147
__device__ void RunWrite(const DstDesc &dst_desc, DstBuffer &dst_buf, Number< ThreadScratchId > thread_scratch_id=Number< ThreadScratchId >{})
Definition: thread_group_tensor_slice_transfer_v4r1.hpp:125
Definition: threadwise_tensor_slice_transfer.hpp:39
Definition: integral_constant.hpp:10
Definition: is_known_at_compile_time.hpp:14
Definition: unary_element_wise_operation.hpp:241