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(__gfx908__) || defined(__gfx90a__) || defined(__gfx94__))
46 constexpr
index_t shared_block_size =
47 GridwiseGemm::GetSharedMemoryNumberOfByte() /
sizeof(FloatAB);
49 __shared__ FloatAB p_shared_block[shared_block_size];
51 GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
55 a_b_k0_m_k1_grid_desc,
56 b_b_k0_n_k1_grid_desc,
57 c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc,
61 c_block_cluster_adaptor);
66 ignore = a_b_k0_m_k1_grid_desc;
67 ignore = b_b_k0_n_k1_grid_desc;
68 ignore = c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc;
72 ignore = c_block_cluster_adaptor;
81 typename ABK0MK1GridDesc,
82 typename BBK0NK1GridDesc,
84 typename AElementwiseOperation,
85 typename BElementwiseOperation,
86 typename CElementwiseOperation,
95 typename ABlockTransferThreadClusterLengths_K0_M_K1,
96 typename ABlockTransferThreadClusterArrangeOrder,
97 typename ABlockTransferSrcAccessOrder,
98 index_t ABlockTransferSrcVectorDim,
99 index_t ABlockTransferSrcScalarPerVector,
100 index_t ABlockTransferDstScalarPerVector_K1,
101 bool AThreadTransferSrcResetCoordinateAfterRun,
102 bool ABlockLdsExtraM,
103 typename BBlockTransferThreadClusterLengths_K0_N_K1,
104 typename BBlockTransferThreadClusterArrangeOrder,
105 typename BBlockTransferSrcAccessOrder,
106 index_t BBlockTransferSrcVectorDim,
107 index_t BBlockTransferSrcScalarPerVector,
108 index_t BBlockTransferDstScalarPerVector_K1,
109 bool BThreadTransferSrcResetCoordinateAfterRun,
110 bool BBlockLdsExtraN,
111 typename CThreadTransferSrcDstAccessOrder,
112 index_t CThreadTransferSrcDstVectorDim,
113 index_t CThreadTransferDstScalarPerVector>
132 constexpr
auto max_lds_align =
K1;
135 constexpr
auto a_k0_m_k1_block_desc = [&]() {
136 if constexpr(ABlockLdsExtraM)
150 constexpr
auto b_k0_n_k1_block_desc = [&]() {
151 if constexpr(BBlockLdsExtraN)
165 constexpr
auto a_block_space_size =
168 constexpr
auto b_block_space_size =
171 return (a_block_space_size + b_block_space_size) *
sizeof(FloatAB);
175 template <
typename Block2CTileMap>
176 __host__ __device__
static constexpr
bool
178 const BBK0NK1GridDesc& b_b_k0_n_k1_grid_desc,
179 const CMNGridDesc& c_m_n_grid_desc,
180 const Block2CTileMap& block_2_ctile_map)
183 "wrong! K1 need to be known at compile-time");
185 static_assert((MPerBlock % (MPerXDL * MRepeat) == 0) &&
186 (NPerBlock % (NRepeat * NPerXDL)) == 0,
187 "Invalid tuning param!");
189 const auto M = a_b_k0_m_k1_grid_desc.GetLength(
I2);
190 const auto N = b_b_k0_n_k1_grid_desc.GetLength(
I2);
191 const auto K0 = a_b_k0_m_k1_grid_desc.GetLength(
I1);
192 const auto KBatch = a_b_k0_m_k1_grid_desc.GetLength(
I0);
194 if(!(M == c_m_n_grid_desc.GetLength(
I0) && N == c_m_n_grid_desc.GetLength(
I1) &&
195 K0 == b_b_k0_n_k1_grid_desc.GetLength(
I1) &&
196 K1 == a_b_k0_m_k1_grid_desc.GetLength(
I3) &&
197 K1 == b_b_k0_n_k1_grid_desc.GetLength(
I3) &&
198 KBatch == b_b_k0_n_k1_grid_desc.GetLength(
I0)))
201 if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K0 % K0PerBlock == 0))
204 if(!block_2_ctile_map.CheckValidity(c_m_n_grid_desc))
215 const bool has_main_k0_block_loop = K0 > K0PerBlock;
217 return has_main_k0_block_loop;
220 __host__ __device__
static constexpr
auto
223 constexpr
auto max_lds_align =
K1;
226 constexpr
auto a_k0_m_k1_block_desc = [&]() {
227 if constexpr(ABlockLdsExtraM)
241 constexpr
auto b_k0_n_k1_block_desc = [&]() {
242 if constexpr(BBlockLdsExtraN)
255 using BlockwiseGemm =
259 decltype(a_k0_m_k1_block_desc),
260 decltype(b_k0_n_k1_block_desc),
267 return BlockwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_m_n_grid_desc);
275 c_m_n_grid_desc, 8, KBatch);
281 template <
bool HasMainKBlockLoop>
282 __device__
static void Run(
const FloatAB* __restrict__ p_a_grid,
283 const FloatAB* __restrict__ p_b_grid,
284 FloatC* __restrict__ p_c_grid,
285 FloatAB* __restrict__ p_shared_block,
286 const ABK0MK1GridDesc& a_b_k0_m_k1_grid_desc,
287 const BBK0NK1GridDesc& b_b_k0_n_k1_grid_desc,
289 const AElementwiseOperation& a_element_op,
290 const BElementwiseOperation& b_element_op,
291 const CElementwiseOperation& c_element_op,
294 const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
295 p_a_grid, a_b_k0_m_k1_grid_desc.GetElementSpaceSize());
296 const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
297 p_b_grid, b_b_k0_n_k1_grid_desc.GetElementSpaceSize());
298 auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
299 p_c_grid, c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc.GetElementSpaceSize());
301 const auto K0 = a_b_k0_m_k1_grid_desc.GetLength(
I1);
304 const auto block_work_idx =
307 if(!c_block_cluster_adaptor.ValidCTileIndex(
309 make_tuple(c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc.GetLength(
I0),
310 c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc.GetLength(
I1))))
315 const index_t k_batch_id = block_work_idx[
I0];
318 const index_t m_block_data_idx_on_grid =
319 __builtin_amdgcn_readfirstlane(block_work_idx[
I1] * MPerBlock);
321 const index_t n_block_data_idx_on_grid =
322 __builtin_amdgcn_readfirstlane(block_work_idx[
I2] * NPerBlock);
325 constexpr
auto max_lds_align =
K1;
328 constexpr
auto a_k0_m_k1_block_desc = [&]() {
329 if constexpr(ABlockLdsExtraM)
342 constexpr
auto a_b_k0_m_k1_block_desc = [&]() {
343 if constexpr(ABlockLdsExtraM)
360 constexpr
auto b_k0_n_k1_block_desc = [&]() {
361 if constexpr(BBlockLdsExtraN)
374 constexpr
auto b_b_k0_n_k1_block_desc = [&]() {
375 if constexpr(BBlockLdsExtraN)
392 auto a_blockwise_copy =
394 AElementwiseOperation,
398 ABlockTransferThreadClusterLengths_K0_M_K1,
399 ABlockTransferThreadClusterArrangeOrder,
402 decltype(a_b_k0_m_k1_grid_desc),
403 decltype(a_b_k0_m_k1_block_desc),
404 ABlockTransferSrcAccessOrder,
406 ABlockTransferSrcVectorDim,
408 ABlockTransferSrcScalarPerVector,
409 ABlockTransferDstScalarPerVector_K1,
412 AThreadTransferSrcResetCoordinateAfterRun,
414 a_b_k0_m_k1_grid_desc,
417 a_b_k0_m_k1_block_desc,
422 auto b_blockwise_copy =
424 BElementwiseOperation,
428 BBlockTransferThreadClusterLengths_K0_N_K1,
429 BBlockTransferThreadClusterArrangeOrder,
432 decltype(b_b_k0_n_k1_grid_desc),
433 decltype(b_b_k0_n_k1_block_desc),
434 BBlockTransferSrcAccessOrder,
436 BBlockTransferSrcVectorDim,
438 BBlockTransferSrcScalarPerVector,
439 BBlockTransferDstScalarPerVector_K1,
442 BThreadTransferSrcResetCoordinateAfterRun,
444 b_b_k0_n_k1_grid_desc,
447 b_b_k0_n_k1_block_desc,
459 auto blockwise_gemm =
463 decltype(a_k0_m_k1_block_desc),
464 decltype(b_k0_n_k1_block_desc),
471 auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
474 constexpr
auto a_block_space_size =
477 FloatAB* p_a_block = p_shared_block;
478 FloatAB* p_b_block = p_shared_block + a_block_space_size;
480 constexpr
auto a_block_slice_copy_step =
make_multi_index(0, K0PerBlock, 0, 0);
481 constexpr
auto b_block_slice_copy_step =
make_multi_index(0, K0PerBlock, 0, 0);
483 auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
484 p_a_block, a_k0_m_k1_block_desc.GetElementSpaceSize());
485 auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
486 p_b_block, b_k0_n_k1_block_desc.GetElementSpaceSize());
490 a_blockwise_copy.RunRead(a_b_k0_m_k1_grid_desc, a_grid_buf);
491 b_blockwise_copy.
RunRead(b_b_k0_n_k1_grid_desc, b_grid_buf);
493 a_blockwise_copy.RunWrite(a_b_k0_m_k1_block_desc, a_block_buf);
494 b_blockwise_copy.
RunWrite(b_b_k0_n_k1_block_desc, b_block_buf);
498 c_thread_buf.Clear();
501 if constexpr(HasMainKBlockLoop)
503 index_t k0_block_data_begin = 0;
507 a_blockwise_copy.MoveSrcSliceWindow(a_b_k0_m_k1_grid_desc, a_block_slice_copy_step);
510 a_blockwise_copy.RunRead(a_b_k0_m_k1_grid_desc, a_grid_buf);
514 b_blockwise_copy.
RunRead(b_b_k0_n_k1_grid_desc, b_grid_buf);
516 blockwise_gemm.
Run(a_block_buf, b_block_buf, c_thread_buf);
520 a_blockwise_copy.RunWrite(a_b_k0_m_k1_block_desc, a_block_buf);
521 b_blockwise_copy.
RunWrite(b_b_k0_n_k1_block_desc, b_block_buf);
523 k0_block_data_begin += K0PerBlock;
524 }
while(k0_block_data_begin < (K0 - K0PerBlock));
531 blockwise_gemm.
Run(a_block_buf, b_block_buf, c_thread_buf);
536 constexpr
auto c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc =
537 blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
539 constexpr
auto M0 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(
I0);
540 constexpr
auto N0 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(
I1);
541 constexpr
auto M1 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(
I2);
542 constexpr
auto N1 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(
I3);
543 constexpr
auto M2 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(
I4);
544 constexpr
auto M3 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(
I5);
545 constexpr
auto M4 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(
I6);
546 constexpr
auto N2 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(
I7);
548 constexpr
auto c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc =
554 const auto c_thread_mtx_on_block =
555 blockwise_gemm.CalculateCThreadOriginDataIndex(
I0,
I0,
I0,
I0);
557 const index_t m_thread_data_on_grid =
558 m_block_data_idx_on_grid + c_thread_mtx_on_block[
I0];
560 const index_t n_thread_data_on_grid =
561 n_block_data_idx_on_grid + c_thread_mtx_on_block[
I1];
563 const auto m_thread_data_on_grid_to_m0_m1_m2_m3_m4_adaptor =
569 const auto m_thread_data_on_grid_idx =
570 m_thread_data_on_grid_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
578 const auto n_thread_data_on_grid_idx =
579 n_thread_data_on_grid_to_n0_n1_n2_adaptor.CalculateBottomIndex(
585 decltype(c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc),
586 decltype(c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc),
587 CElementwiseOperation,
589 CThreadTransferSrcDstAccessOrder,
590 CThreadTransferSrcDstVectorDim,
591 CThreadTransferDstScalarPerVector,
592 CGlobalMemoryDataOperation,
596 c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc,
598 n_thread_data_on_grid_idx[
I0],
599 m_thread_data_on_grid_idx[
I1],
600 n_thread_data_on_grid_idx[
I1],
601 m_thread_data_on_grid_idx[
I2],
602 m_thread_data_on_grid_idx[
I3],
603 m_thread_data_on_grid_idx[
I4],
604 n_thread_data_on_grid_idx[
I2]),
607 c_thread_copy.Run(c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc,
610 c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc,
#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__ 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:276
__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:425
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_tuple(Xs &&... xs)
Definition: tuple.hpp:211
int32_t index_t
Definition: ck.hpp:298
__device__ void block_sync_lds()
Definition: synchronization.hpp:10
typename remove_cv< T >::type remove_cv_t
Definition: type.hpp:295
const GenericPointer< typename T::ValueType > T2 value
Definition: pointer.h:1350
Definition: block_to_ctile_map.hpp:540
Definition: blockwise_gemm_smfmac_xdlops.hpp:44
Definition: gridwise_gemm_xdlops_v2r4.hpp:115
static constexpr auto I3
Definition: gridwise_gemm_xdlops_v2r4.hpp:119
ThisThreadBlock< BlockSize > ThisThreadBlock
Definition: gridwise_gemm_xdlops_v2r4.hpp:128
static constexpr auto I6
Definition: gridwise_gemm_xdlops_v2r4.hpp:122
__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:271
static constexpr auto K1
Definition: gridwise_gemm_xdlops_v2r4.hpp:126
static constexpr auto I5
Definition: gridwise_gemm_xdlops_v2r4.hpp:121
static constexpr auto I7
Definition: gridwise_gemm_xdlops_v2r4.hpp:123
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:282
__host__ static constexpr __device__ auto MakeCM0N0M1N1M2M3M4N2GridDescriptor(const CMNGridDesc &c_m_n_grid_desc)
Definition: gridwise_gemm_xdlops_v2r4.hpp:221
static constexpr auto I4
Definition: gridwise_gemm_xdlops_v2r4.hpp:120
__host__ static constexpr __device__ index_t GetSharedMemoryNumberOfByte()
Definition: gridwise_gemm_xdlops_v2r4.hpp:130
decltype(MakeCBlockClusterAdaptor(CMNGridDesc{}, 1, 1, 1)) CBlockClusterAdaptor
Definition: gridwise_gemm_xdlops_v2r4.hpp:279
static constexpr auto I1
Definition: gridwise_gemm_xdlops_v2r4.hpp:117
static constexpr auto I0
Definition: gridwise_gemm_xdlops_v2r4.hpp:116
__host__ static constexpr __device__ bool CalculateHasMainK0BlockLoop(index_t K0)
Definition: gridwise_gemm_xdlops_v2r4.hpp:213
static constexpr auto I2
Definition: gridwise_gemm_xdlops_v2r4.hpp:118
__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:177
decltype(MakeCM0N0M1N1M2M3M4N2GridDescriptor(CMNGridDesc{})) CM0N0M1N1M2M3M4N2GridDesc
Definition: gridwise_gemm_xdlops_v2r4.hpp:278
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:143
__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:119
__device__ void MoveSrcSliceWindow(const SrcDesc &src_desc, const Index &step)
Definition: thread_group_tensor_slice_transfer_v4r1.hpp:153
__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:131
Definition: threadwise_tensor_slice_transfer.hpp:39
Definition: integral_constant.hpp:20
Definition: is_known_at_compile_time.hpp:14
Definition: unary_element_wise_operation.hpp:334