Skip to content

GitLab

Commit 7610e049 authored by Chao Liu's avatar Chao Liu
Browse files

refactor

parent 1a24ad25
Hide whitespace changes
Inline Side-by-side
Showing with 99 additions and 146 deletions
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4r2.hpp
View file @ 7610e049
...@@ -123,6 +123,8 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2 ...@@ -123,6 +123,8 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
// K1 should be Number<...> // K1 should be Number<...>
static constexpr auto K1 = Number<K1Value>{}; static constexpr auto K1 = Number<K1Value>{};
using ThisThreadBlock = ThisThreadBlock<BlockSize>;
__host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte() __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
{ {
constexpr auto max_lds_align = K1; constexpr auto max_lds_align = K1;
...@@ -476,7 +478,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2 ...@@ -476,7 +478,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
// sanity check // sanity check
auto blockwise_gemm = auto blockwise_gemm =
BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize, BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<ThisThreadBlock,
FloatAB, FloatAB,
FloatAcc, FloatAcc,
decltype(a_k0_m_k1_block_desc), decltype(a_k0_m_k1_block_desc),
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r2.hpp
View file @ 7610e049
...@@ -24,7 +24,7 @@ template <typename GridwiseGemm, ...@@ -24,7 +24,7 @@ template <typename GridwiseGemm,
typename BElementwiseOperation, typename BElementwiseOperation,
typename CElementwiseOperation, typename CElementwiseOperation,
typename Block2CTileMap, typename Block2CTileMap,
bool HasMainK0BlockLoop> bool HasMainKBlockLoop>
__global__ void __global__ void
#if CK_USE_LAUNCH_BOUNDS #if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU) __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
...@@ -48,7 +48,7 @@ __global__ void ...@@ -48,7 +48,7 @@ __global__ void
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__)) #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()]; __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
GridwiseGemm::template Run<HasMainK0BlockLoop>( GridwiseGemm::template Run<HasMainKBlockLoop>(
p_a_grid, p_a_grid,
p_b_grid, p_b_grid,
p_c_grid, p_c_grid,
...@@ -119,7 +119,7 @@ template < ...@@ -119,7 +119,7 @@ template <
index_t CShuffleNXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle,
typename CBlockTransferClusterLengths_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl, typename CBlockTransferClusterLengths_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl,
index_t CBlockTransferScalarPerVector_NWaveNPerXdl, index_t CBlockTransferScalarPerVector_NWaveNPerXdl,
index_t NumPrefetch = 1> index_t NumGemmKPrefetchStage = 1>
struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r2 struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r2
{ {
static constexpr auto I0 = Number<0>{}; static constexpr auto I0 = Number<0>{};
...@@ -134,6 +134,10 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r2 ...@@ -134,6 +134,10 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r2
// K1 should be Number<...> // K1 should be Number<...>
static constexpr auto K1 = Number<K1Value>{}; static constexpr auto K1 = Number<K1Value>{};
using ThisThreadBlock = ThisThreadBlock<BlockSize>;
using GridwiseGemmPipe = GridwiseGemmPipeline_v1<NumGemmKPrefetchStage>;
__host__ __device__ static constexpr auto GetABlockDescriptor_K0PerBlock_MPerBlock_K1() __host__ __device__ static constexpr auto GetABlockDescriptor_K0PerBlock_MPerBlock_K1()
{ {
constexpr auto max_lds_align = K1; constexpr auto max_lds_align = K1;
...@@ -252,21 +256,10 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r2 ...@@ -252,21 +256,10 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r2
if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K0 % K0PerBlock == 0)) if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K0 % K0PerBlock == 0))
return false; return false;
// check NumPrefetch // check gridwise gemm pipeline
if constexpr(NumPrefetch == 1) const auto num_k_loop = K0 / K0PerBlock;
{
// 1-stage prefetch always supported if(!GridwiseGemmPipe::IsSupported(num_k_loop))
}
else if constexpr(NumPrefetch == 2)
{
// 2-stage prefetch currently only support even number of K0 loop
// TODO: add support for odd number of K0 loop
if(!((K0 / K0PerBlock) % 2 == 0))
{
return false;
}
}
else
{ {
return false; return false;
} }
...@@ -296,12 +289,11 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r2 ...@@ -296,12 +289,11 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r2
return grid_size; return grid_size;
} }
// TODO move this function into GEMM-pipeline class __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
__host__ __device__ static constexpr bool CalculateHasMainK0BlockLoop(index_t K0)
{ {
const bool has_main_k0_block_loop = (K0 / (NumPrefetch * K0PerBlock)) > 1; const index_t num_loop = K / (K0PerBlock * K1);
return has_main_k0_block_loop; return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
} }
template <typename CGridDesc_M_N_> template <typename CGridDesc_M_N_>
...@@ -379,7 +371,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r2 ...@@ -379,7 +371,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r2
using DefaultBlock2CTileMap = using DefaultBlock2CTileMap =
remove_cvref_t<decltype(MakeDefaultBlock2CTileMap(CGridDesc_M_N{}, 1, 1))>; remove_cvref_t<decltype(MakeDefaultBlock2CTileMap(CGridDesc_M_N{}, 1, 1))>;
template <bool HasMainK0BlockLoop, typename Block2CTileMap = DefaultBlock2CTileMap> template <bool HasMainKBlockLoop, typename Block2CTileMap = DefaultBlock2CTileMap>
__device__ static void __device__ static void
Run(const FloatAB* __restrict__ p_a_grid, Run(const FloatAB* __restrict__ p_a_grid,
const FloatAB* __restrict__ p_b_grid, const FloatAB* __restrict__ p_b_grid,
...@@ -455,7 +447,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r2 ...@@ -455,7 +447,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r2
1, 1,
AThreadTransferSrcResetCoordinateAfterRun, AThreadTransferSrcResetCoordinateAfterRun,
true, true,
NumPrefetch>( NumGemmKPrefetchStage>(
a_grid_desc_k0_m_k1, a_grid_desc_k0_m_k1,
make_multi_index(0, m_block_data_idx_on_grid, 0), make_multi_index(0, m_block_data_idx_on_grid, 0),
a_element_op, a_element_op,
...@@ -486,7 +478,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r2 ...@@ -486,7 +478,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r2
1, 1,
BThreadTransferSrcResetCoordinateAfterRun, BThreadTransferSrcResetCoordinateAfterRun,
true, true,
NumPrefetch>( NumGemmKPrefetchStage>(
b_grid_desc_k0_n_k1, b_grid_desc_k0_n_k1,
make_multi_index(0, n_block_data_idx_on_grid, 0), make_multi_index(0, n_block_data_idx_on_grid, 0),
b_element_op, b_element_op,
...@@ -503,7 +495,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r2 ...@@ -503,7 +495,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r2
// sanity check // sanity check
auto blockwise_gemm = auto blockwise_gemm =
BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize, BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<ThisThreadBlock,
FloatAB, FloatAB,
FloatAcc, FloatAcc,
decltype(a_block_desc_k0_m_k1), decltype(a_block_desc_k0_m_k1),
...@@ -531,41 +523,23 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r2 ...@@ -531,41 +523,23 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r2
constexpr auto b_block_slice_copy_step = make_multi_index(K0PerBlock, 0, 0); constexpr auto b_block_slice_copy_step = make_multi_index(K0PerBlock, 0, 0);
// gridwise GEMM pipeline // gridwise GEMM pipeline
const auto gridwise_gemm_pipeline =
GridwiseGemmPipeline_v1<remove_cvref_t<decltype(a_grid_desc_k0_m_k1)>,
remove_cvref_t<decltype(a_block_desc_k0_m_k1)>,
remove_cvref_t<decltype(a_blockwise_copy)>,
remove_cvref_t<decltype(a_grid_buf)>,
remove_cvref_t<decltype(a_block_buf)>,
remove_cvref_t<decltype(a_block_slice_copy_step)>,
remove_cvref_t<decltype(b_grid_desc_k0_n_k1)>,
remove_cvref_t<decltype(b_block_desc_k0_n_k1)>,
remove_cvref_t<decltype(b_blockwise_copy)>,
remove_cvref_t<decltype(b_grid_buf)>,
remove_cvref_t<decltype(b_block_buf)>,
remove_cvref_t<decltype(b_block_slice_copy_step)>,
remove_cvref_t<decltype(blockwise_gemm)>,
remove_cvref_t<decltype(c_thread_buf)>,
NumPrefetch,
HasMainK0BlockLoop>{};
const index_t K0BlockMainLoop = __builtin_amdgcn_readfirstlane(K0 / K0PerBlock); const index_t K0BlockMainLoop = __builtin_amdgcn_readfirstlane(K0 / K0PerBlock);
gridwise_gemm_pipeline.Run(a_grid_desc_k0_m_k1, GridwiseGemmPipe::template Run<HasMainKBlockLoop>(a_grid_desc_k0_m_k1,
a_block_desc_k0_m_k1, a_block_desc_k0_m_k1,
a_blockwise_copy, a_blockwise_copy,
a_grid_buf, a_grid_buf,
a_block_buf, a_block_buf,
a_block_slice_copy_step, a_block_slice_copy_step,
b_grid_desc_k0_n_k1, b_grid_desc_k0_n_k1,
b_block_desc_k0_n_k1, b_block_desc_k0_n_k1,
b_blockwise_copy, b_blockwise_copy,
b_grid_buf, b_grid_buf,
b_block_buf, b_block_buf,
b_block_slice_copy_step, b_block_slice_copy_step,
blockwise_gemm, blockwise_gemm,
c_thread_buf, c_thread_buf,
K0BlockMainLoop); K0BlockMainLoop);
// shuffle C and write out // shuffle C and write out
{ {
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r3.hpp
View file @ 7610e049
...@@ -25,7 +25,7 @@ template <typename GridwiseGemm, ...@@ -25,7 +25,7 @@ template <typename GridwiseGemm,
typename BElementwiseOperation, typename BElementwiseOperation,
typename CElementwiseOperation, typename CElementwiseOperation,
typename Block2CTileMap, typename Block2CTileMap,
bool HasMainK0BlockLoop> bool HasMainKBlockLoop>
__global__ void __global__ void
#if CK_USE_LAUNCH_BOUNDS #if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU) __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
...@@ -52,7 +52,7 @@ __global__ void ...@@ -52,7 +52,7 @@ __global__ void
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__)) #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()]; __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
GridwiseGemm::template Run<HasMainK0BlockLoop>( GridwiseGemm::template Run<HasMainKBlockLoop>(
p_a_grid, p_a_grid,
p_b_grid, p_b_grid,
p_c_grid, p_c_grid,
...@@ -128,7 +128,7 @@ template < ...@@ -128,7 +128,7 @@ template <
index_t CShuffleNXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle,
typename CBlockTransferClusterLengths_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl, typename CBlockTransferClusterLengths_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl,
index_t CBlockTransferScalarPerVector_NWaveNPerXdl, index_t CBlockTransferScalarPerVector_NWaveNPerXdl,
index_t NumPrefetch = 1> index_t NumGemmKPrefetchStage = 1>
struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r3 struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r3
{ {
static constexpr auto I0 = Number<0>{}; static constexpr auto I0 = Number<0>{};
...@@ -143,6 +143,10 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r3 ...@@ -143,6 +143,10 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r3
// K1 should be Number<...> // K1 should be Number<...>
static constexpr auto K1 = Number<K1Value>{}; static constexpr auto K1 = Number<K1Value>{};
using ThisThreadBlock = ThisThreadBlock<BlockSize>;
using GridwiseGemmPipe = GridwiseGemmPipeline_v1<NumGemmKPrefetchStage>;
__host__ __device__ static constexpr auto GetABlockDescriptor_K0PerBlock_MPerBlock_K1() __host__ __device__ static constexpr auto GetABlockDescriptor_K0PerBlock_MPerBlock_K1()
{ {
constexpr auto max_lds_align = K1; constexpr auto max_lds_align = K1;
...@@ -261,21 +265,10 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r3 ...@@ -261,21 +265,10 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r3
if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K0 % K0PerBlock == 0)) if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K0 % K0PerBlock == 0))
return false; return false;
// check NumPrefetch // check gridwise gemm pipeline
if constexpr(NumPrefetch == 1) const auto num_k_loop = K0 / K0PerBlock;
{
// 1-stage prefetch always supported if(!GridwiseGemmPipe::IsSupported(num_k_loop))
}
else if constexpr(NumPrefetch == 2)
{
// 2-stage prefetch currently only support even number of K0 loop
// TODO: add support for odd number of K0 loop
if(!((K0 / K0PerBlock) % 2 == 0))
{
return false;
}
}
else
{ {
return false; return false;
} }
...@@ -305,12 +298,11 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r3 ...@@ -305,12 +298,11 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r3
return grid_size; return grid_size;
} }
// TODO move this function into GEMM-pipeline class __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
__host__ __device__ static constexpr bool CalculateHasMainK0BlockLoop(index_t K0)
{ {
const bool has_main_k0_block_loop = (K0 / (NumPrefetch * K0PerBlock)) > 1; const index_t num_loop = K / (K0PerBlock * K1);
return has_main_k0_block_loop; return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
} }
template <typename CGridDesc_M_N_> template <typename CGridDesc_M_N_>
...@@ -393,7 +385,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r3 ...@@ -393,7 +385,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r3
using DefaultBlock2CTileMap = using DefaultBlock2CTileMap =
remove_cvref_t<decltype(MakeDefaultBlock2CTileMap(CGridDesc_M_N{}, 1, 1))>; remove_cvref_t<decltype(MakeDefaultBlock2CTileMap(CGridDesc_M_N{}, 1, 1))>;
template <bool HasMainK0BlockLoop, typename Block2CTileMap = DefaultBlock2CTileMap> template <bool HasMainKBlockLoop, typename Block2CTileMap = DefaultBlock2CTileMap>
__device__ static void __device__ static void
Run(const FloatAB* __restrict__ p_a_grid, Run(const FloatAB* __restrict__ p_a_grid,
const FloatAB* __restrict__ p_b_grid, const FloatAB* __restrict__ p_b_grid,
...@@ -522,7 +514,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r3 ...@@ -522,7 +514,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r3
// sanity check // sanity check
auto blockwise_gemm = auto blockwise_gemm =
BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize, BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<ThisThreadBlock,
FloatAB, FloatAB,
FloatAcc, FloatAcc,
decltype(a_block_desc_k0_m_k1), decltype(a_block_desc_k0_m_k1),
...@@ -550,41 +542,23 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r3 ...@@ -550,41 +542,23 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r3
constexpr auto b_block_slice_copy_step = make_multi_index(K0PerBlock, 0, 0); constexpr auto b_block_slice_copy_step = make_multi_index(K0PerBlock, 0, 0);
// gridwise GEMM pipeline // gridwise GEMM pipeline
const auto gridwise_gemm_pipeline =
GridwiseGemmPipeline_v1<remove_cvref_t<decltype(a_grid_desc_k0_m_k1)>,
remove_cvref_t<decltype(a_block_desc_k0_m_k1)>,
remove_cvref_t<decltype(a_blockwise_copy)>,
remove_cvref_t<decltype(a_grid_buf)>,
remove_cvref_t<decltype(a_block_buf)>,
remove_cvref_t<decltype(a_block_slice_copy_step)>,
remove_cvref_t<decltype(b_grid_desc_k0_n_k1)>,
remove_cvref_t<decltype(b_block_desc_k0_n_k1)>,
remove_cvref_t<decltype(b_blockwise_copy)>,
remove_cvref_t<decltype(b_grid_buf)>,
remove_cvref_t<decltype(b_block_buf)>,
remove_cvref_t<decltype(b_block_slice_copy_step)>,
remove_cvref_t<decltype(blockwise_gemm)>,
remove_cvref_t<decltype(c_thread_buf)>,
NumPrefetch,
HasMainK0BlockLoop>{};
const index_t K0BlockMainLoop = __builtin_amdgcn_readfirstlane(K0 / K0PerBlock); const index_t K0BlockMainLoop = __builtin_amdgcn_readfirstlane(K0 / K0PerBlock);
gridwise_gemm_pipeline.Run(a_grid_desc_k0_m_k1, GridwiseGemmPipe::template Run<HasMainKBlockLoop>(a_grid_desc_k0_m_k1,
a_block_desc_k0_m_k1, a_block_desc_k0_m_k1,
a_blockwise_copy, a_blockwise_copy,
a_grid_buf, a_grid_buf,
a_block_buf, a_block_buf,
a_block_slice_copy_step, a_block_slice_copy_step,
b_grid_desc_k0_n_k1, b_grid_desc_k0_n_k1,
b_block_desc_k0_n_k1, b_block_desc_k0_n_k1,
b_blockwise_copy, b_blockwise_copy,
b_grid_buf, b_grid_buf,
b_block_buf, b_block_buf,
b_block_slice_copy_step, b_block_slice_copy_step,
blockwise_gemm, blockwise_gemm,
c_thread_buf, c_thread_buf,
K0BlockMainLoop); K0BlockMainLoop);
// shuffle C and write out // shuffle C and write out
{ {
...@@ -623,17 +597,18 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r3 ...@@ -623,17 +597,18 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r3
constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor( constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl, c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
make_tuple( make_tuple(make_freeze_transform(I0), // freeze mblock
make_freeze_transform(I0), // freeze mblock make_pass_through_transform(
make_pass_through_transform( Number<CShuffleMXdlPerWavePerShuffle>{}), // M0 (MXdlPerWave) per
Number<CShuffleMXdlPerWavePerShuffle>{}), // M0 (MXdlPerWave) per shuffle // shuffle
make_unmerge_transform( make_unmerge_transform(
make_tuple(M1, M2, M3, M4)), // M1 = MWave, M2 * M3 * M4 = MPerXdl make_tuple(M1, M2, M3, M4)), // M1 = MWave, M2 * M3 * M4 = MPerXdl
make_freeze_transform(I0), // freeze nblock make_freeze_transform(I0), // freeze nblock
make_pass_through_transform( make_pass_through_transform(
Number<CShuffleNXdlPerWavePerShuffle>{}), // N0 (NXdlPerWave) per shuffle Number<CShuffleNXdlPerWavePerShuffle>{}), // N0 (NXdlPerWave) per
make_unmerge_transform( // shuffle
make_tuple(N1, N2))), // M1 = MWave, M2 * M3 * M4 = MPerXdl make_unmerge_transform(
make_tuple(N1, N2))), // M1 = MWave, M2 * M3 * M4 = MPerXdl
make_tuple(Sequence<0>{}, make_tuple(Sequence<0>{},
Sequence<1>{}, Sequence<1>{},
Sequence<2>{}, Sequence<2>{},
......
library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instance.cpp
View file @ 7610e049
#include <stdlib.h> #include <stdlib.h>
#include "config.hpp" #include "config.hpp"
#include "device_gemm_xdl_c_shuffle.hpp" #include "device_gemm_xdl_cshuffle.hpp"
#include "element_wise_operation.hpp" #include "element_wise_operation.hpp"
#include "device_operation_instance.hpp" #include "device_operation_instance.hpp"
...@@ -20,26 +20,28 @@ using S = ck::Sequence<Is...>; ...@@ -20,26 +20,28 @@ using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough; using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// Compilation parameters for a[m, k] * b[n, k] = c[m, n] // Compilation parameters for a[m, k] * b[n, k] = c[m, n]
using device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instances = std::tuple< using device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instances = std::tuple<
// clang-format off // clang-format off
//#####################| AData| BData| CData| AccData| CShuffle| ALayout| BLayout| CLayout| A| B| C| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Num| //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle| A| B| C| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//#####################| Type| Type| Type| Type| DataType| | | | Elementwise| Elementwise| Elementwise| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Prefetch| //#####################| | | | Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//#####################| | | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| | //#####################| | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//#####################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | //#####################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, F16, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8, 2>, DeviceGemm_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 2, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, F16, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8, 2>, DeviceGemm_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 2, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, F16, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8, 2>, DeviceGemm_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 2, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, F16, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8, 2>, DeviceGemm_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 2, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, F16, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 4>, 8, 2>, DeviceGemm_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 2, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, F16, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8, 2>, DeviceGemm_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 2, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, F16, Row, Col, Row, PassThrough, PassThrough, PassThrough, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8, 2>, DeviceGemm_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 2, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, F16, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8, 2>, DeviceGemm_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 2, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, F16, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8, 2>, DeviceGemm_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 2, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, F16, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 4>, 8, 2>, DeviceGemm_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 2, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, F16, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8, 2>, DeviceGemm_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 2, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, F16, Row, Col, Row, PassThrough, PassThrough, PassThrough, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8, 2>, DeviceGemm_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 2, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, F16, Row, Col, Row, PassThrough, PassThrough, PassThrough, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8, 2> DeviceGemm_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 2, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>
// clang-format on // clang-format on
>; >;
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment