Skip to content

GitLab

Commit c3de33fa authored by ltqin's avatar ltqin
Browse files

Merge branch 'develop' into bmatrix_skip_lds

parents b571256f 85fc91c3
Hide whitespace changes
Inline Side-by-side
Showing with 848 additions and 83 deletions
example/01_gemm/CMakeLists.txt
View file @ c3de33fa
...@@ -5,4 +5,5 @@ add_example_executable(example_gemm_xdl_fp16 gemm_xdl_fp16.cpp) ...@@ -5,4 +5,5 @@ add_example_executable(example_gemm_xdl_fp16 gemm_xdl_fp16.cpp)
add_example_executable(example_gemm_xdl_bf16 gemm_xdl_bf16.cpp) add_example_executable(example_gemm_xdl_bf16 gemm_xdl_bf16.cpp)
add_example_executable(example_gemm_xdl_int8 gemm_xdl_int8.cpp) add_example_executable(example_gemm_xdl_int8 gemm_xdl_int8.cpp)
add_example_executable(example_gemm_xdl_skip_b_lds_fp16 gemm_xdl_skip_b_lds_fp16.cpp) add_example_executable(example_gemm_xdl_skip_b_lds_fp16 gemm_xdl_skip_b_lds_fp16.cpp)
add_example_executable(example_gemm_xdl_fp64 gemm_xdl_fp64.cpp) # FIXME: re-enable this exampe as test when SWDEV-335738 is fixed
add_example_executable_no_testing(example_gemm_xdl_fp64 gemm_xdl_fp64.cpp)
example/01_gemm/gemm_dl_fp16.cpp
View file @ c3de33fa
...@@ -170,9 +170,7 @@ int main(int argc, char* argv[]) ...@@ -170,9 +170,7 @@ int main(int argc, char* argv[])
if(!gemm.IsSupportedArgument(argument)) if(!gemm.IsSupportedArgument(argument))
{ {
std::cout << "wrong! device_gemm with the specified compilation parameters does " std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
"not support this GEMM problem"
<< std::endl;
return 0; return 0;
} }
......
example/01_gemm/gemm_dl_fp32.cpp
View file @ c3de33fa
...@@ -169,9 +169,7 @@ int main(int argc, char* argv[]) ...@@ -169,9 +169,7 @@ int main(int argc, char* argv[])
if(!gemm.IsSupportedArgument(argument)) if(!gemm.IsSupportedArgument(argument))
{ {
std::cout << "wrong! device_gemm with the specified compilation parameters does " std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
"not support this GEMM problem"
<< std::endl;
return 0; return 0;
} }
......
example/01_gemm/gemm_dl_int8.cpp
View file @ c3de33fa
...@@ -167,9 +167,7 @@ int main(int argc, char* argv[]) ...@@ -167,9 +167,7 @@ int main(int argc, char* argv[])
if(!gemm.IsSupportedArgument(argument)) if(!gemm.IsSupportedArgument(argument))
{ {
std::cout << "wrong! device_gemm with the specified compilation parameters does " std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
"not support this GEMM problem"
<< std::endl;
return 0; return 0;
} }
......
example/01_gemm/gemm_xdl_bf16.cpp
View file @ c3de33fa
...@@ -193,9 +193,9 @@ int main(int argc, char* argv[]) ...@@ -193,9 +193,9 @@ int main(int argc, char* argv[])
if(!gemm.IsSupportedArgument(argument)) if(!gemm.IsSupportedArgument(argument))
{ {
throw std::runtime_error( std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
"wrong! device_gemm with the specified compilation parameters does "
"not support this GEMM problem"); return 0;
} }
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel}); float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
......
example/01_gemm/gemm_xdl_fp16.cpp
View file @ c3de33fa
...@@ -166,9 +166,9 @@ int main(int argc, char* argv[]) ...@@ -166,9 +166,9 @@ int main(int argc, char* argv[])
if(!gemm.IsSupportedArgument(argument)) if(!gemm.IsSupportedArgument(argument))
{ {
throw std::runtime_error( std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
"wrong! device_gemm with the specified compilation parameters does "
"not support this GEMM problem"); return 0;
} }
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel}); float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
......
example/01_gemm/gemm_xdl_fp64.cpp
View file @ c3de33fa
...@@ -21,8 +21,6 @@ template <ck::index_t... Is> ...@@ -21,8 +21,6 @@ template <ck::index_t... Is>
using S = ck::Sequence<Is...>; using S = ck::Sequence<Is...>;
using F64 = double; using F64 = double;
using F32 = float;
using F16 = ck::half_t;
using ADataType = double; using ADataType = double;
using BDataType = double; using BDataType = double;
...@@ -195,9 +193,9 @@ int main(int argc, char* argv[]) ...@@ -195,9 +193,9 @@ int main(int argc, char* argv[])
if(!gemm.IsSupportedArgument(argument)) if(!gemm.IsSupportedArgument(argument))
{ {
throw std::runtime_error( std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
"wrong! device_gemm with the specified compilation parameters does "
"not support this GEMM problem"); return 0;
} }
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel}); float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
...@@ -233,7 +231,7 @@ int main(int argc, char* argv[]) ...@@ -233,7 +231,7 @@ int main(int argc, char* argv[])
show_2d_matrix(std::cout << "c_host :", c_m_n_host_result) << std::endl; show_2d_matrix(std::cout << "c_host :", c_m_n_host_result) << std::endl;
} }
#endif #endif
ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData); return ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData) ? 0 : 1;
} }
return 0; return 0;
......
example/01_gemm/gemm_xdl_int8.cpp
View file @ c3de33fa
...@@ -194,9 +194,9 @@ int main(int argc, char* argv[]) ...@@ -194,9 +194,9 @@ int main(int argc, char* argv[])
if(!gemm.IsSupportedArgument(argument)) if(!gemm.IsSupportedArgument(argument))
{ {
throw std::runtime_error( std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
"wrong! device_gemm with the specified compilation parameters does "
"not support this GEMM problem"); return 0;
} }
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel}); float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
......
example/09_convnd_fwd/CMakeLists.txt
View file @ c3de33fa
add_example_executable(example_convnd_fwd_xdl_fp32 convnd_fwd_xdl_fp32.cpp) add_example_executable(example_convnd_fwd_xdl_fp32 convnd_fwd_xdl_fp32.cpp)
add_example_executable(example_convnd_fwd_xdl_int8 convnd_fwd_xdl_int8.cpp) add_example_executable(example_convnd_fwd_xdl_int8 convnd_fwd_xdl_int8.cpp)
add_example_executable(example_convnd_fwd_xdl_fp16 convnd_fwd_xdl_fp16.cpp) add_example_executable(example_convnd_fwd_xdl_fp16 convnd_fwd_xdl_fp16.cpp)
add_example_executable(example_convnd_fwd_xdl_fp64 convnd_fwd_xdl_fp64.cpp) # FIXME: re-enable this exampe as test when SWDEV-335738 is fixed
add_example_executable_no_testing(example_convnd_fwd_xdl_fp64 convnd_fwd_xdl_fp64.cpp)
target_link_libraries(example_convnd_fwd_xdl_fp64 PRIVATE conv_util) target_link_libraries(example_convnd_fwd_xdl_fp64 PRIVATE conv_util)
target_link_libraries(example_convnd_fwd_xdl_fp32 PRIVATE conv_util) target_link_libraries(example_convnd_fwd_xdl_fp32 PRIVATE conv_util)
target_link_libraries(example_convnd_fwd_xdl_int8 PRIVATE conv_util) target_link_libraries(example_convnd_fwd_xdl_int8 PRIVATE conv_util)
......
example/16_gemm_reduce/CMakeLists.txt
View file @ c3de33fa
add_example_executable(example_gemm_reduce_xdl_max_fp16 gemm_reduce_xdl_max_fp16.cpp) add_example_executable(example_gemm_reduce_xdl_max_fp16 gemm_reduce_xdl_max_fp16.cpp)
add_example_executable(example_gemm_reduce_xdl_sum_squaresum_fp16 gemm_reduce_xdl_sum_squaresum_fp16.cpp) add_example_executable(example_gemm_reduce_xdl_mean_squaremean_fp16 gemm_reduce_xdl_mean_squaremean_fp16.cpp)
example/16_gemm_reduce/gemm_reduce_xdl_max_fp16.cpp
View file @ c3de33fa
...@@ -29,10 +29,10 @@ using Col = ck::tensor_layout::gemm::ColumnMajor; ...@@ -29,10 +29,10 @@ using Col = ck::tensor_layout::gemm::ColumnMajor;
using ADataType = F16; using ADataType = F16;
using BDataType = F16; using BDataType = F16;
using CDataType = F16; using CDataType = F16;
using GemmAccDataType = F32;
using ReduceAccDataType = F32; using ReduceAccDataType = F32;
using DDataType = F64; using DDataType = F64;
using DPtrsGlobal = ck::Tuple<DDataType*>; using DPtrsGlobal = ck::Tuple<DDataType*>;
using AccDataType = F32;
using ALayout = ck::tensor_layout::gemm::RowMajor; using ALayout = ck::tensor_layout::gemm::RowMajor;
using BLayout = ck::tensor_layout::gemm::ColumnMajor; using BLayout = ck::tensor_layout::gemm::ColumnMajor;
...@@ -52,15 +52,34 @@ static constexpr auto GemmSpecialization = ...@@ -52,15 +52,34 @@ static constexpr auto GemmSpecialization =
// clang-format off // clang-format off
using DeviceGemmReduceInstance = ck::tensor_operation::device::DeviceGemmReduce_Xdl_CShuffle using DeviceGemmReduceInstance = ck::tensor_operation::device::DeviceGemmReduce_Xdl_CShuffle
//######| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| Dxs| DxsInEleOp| DxsOutEleOp| D| 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| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy| //######| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| Dxs| DxsInEleOp| DxsAccEleOp| D| 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| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
//######| | | | Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Reduce| | | MemoryData| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector| //######| | | | Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Reduce| | | MemoryData| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector|
//######| | | | | | | | | | | Operation| Operation| 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_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock| //######| | | | | | | | | | | Operation| Operation| 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_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | //######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< Row, Col, Row, F16, F16, F16, F32, F32, ReduceAccDataType, DPtrsGlobal, AElementOp, BElementOp, CElementOp, DsReduceOp, DsElementOp, DsElementOp, DGlobalMemOp, GemmSpecialization, 1, 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, S<64, 4>, 4, 1>; < Row, Col, Row, F16, F16, F16, F32, F32, ReduceAccDataType, DPtrsGlobal, AElementOp, BElementOp, CElementOp, DsReduceOp, DsElementOp, DsElementOp, DGlobalMemOp, GemmSpecialization, 1, 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, S<64, 4>, 4, 1>;
// clang-format on // clang-format on
using ReferenceGemmInstance = ck::tensor_operation::host:: using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>; BDataType,
CDataType,
GemmAccDataType,
AElementOp,
BElementOp,
CElementOp>;
template <typename ADataType, typename BDataType, typename CDataType, typename DDataType>
void DumpGemmLayerNormPerf(float gemm_reduce_time, int M, int N, int K)
{
std::size_t gemm_flop = std::size_t(2) * M * N * K;
std::size_t gemm_num_byte = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
sizeof(CDataType) * M * N + sizeof(DDataType) * M;
float tflops = static_cast<float>(gemm_flop) / 1.E9 / gemm_reduce_time;
float gemm_gb_per_sec = gemm_num_byte / 1.E6 / gemm_reduce_time;
std::cout << "gemm + reduceMax Perf: " << gemm_reduce_time << " ms, " << tflops << " TFlops, "
<< gemm_gb_per_sec << " GB/s, " << std::endl;
}
int main(int argc, char* argv[]) int main(int argc, char* argv[])
{ {
...@@ -193,21 +212,10 @@ int main(int argc, char* argv[]) ...@@ -193,21 +212,10 @@ int main(int argc, char* argv[])
"not support this GEMM problem"); "not support this GEMM problem");
} }
// init D // [CAUSION]: launch_and_time_kernel will not initialize D.
// If we evaluate kernel multiple time but without initialize D. Verification will fail
d_device_buf.SetValue(ck::NumericLimits<DDataType>::Lowest()); d_device_buf.SetValue(ck::NumericLimits<DDataType>::Lowest());
invoker.Run(argument, StreamConfig{nullptr, false});
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype =
sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(CDataType) * M * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
<< gemm.GetTypeString() << std::endl;
bool pass = true; bool pass = true;
...@@ -246,5 +254,13 @@ int main(int argc, char* argv[]) ...@@ -246,5 +254,13 @@ int main(int argc, char* argv[])
1e-3); 1e-3);
} }
if(time_kernel)
{
float gemm_reduceMax_ave_time = invoker.Run(argument, StreamConfig{nullptr, true});
DumpGemmLayerNormPerf<ADataType, BDataType, CDataType, DDataType>(
gemm_reduceMax_ave_time, M, N, K);
}
return pass ? 0 : 1; return pass ? 0 : 1;
} }
example/16_gemm_reduce/gemm_reduce_xdl_sum_squaresum_fp16.cpp → example/16_gemm_reduce/gemm_reduce_xdl_mean_squaremean_fp16.cpp
View file @ c3de33fa
...@@ -29,10 +29,10 @@ using Col = ck::tensor_layout::gemm::ColumnMajor; ...@@ -29,10 +29,10 @@ using Col = ck::tensor_layout::gemm::ColumnMajor;
using ADataType = F16; using ADataType = F16;
using BDataType = F16; using BDataType = F16;
using CDataType = F16; using CDataType = F16;
using GemmAccDataType = F32;
using ReduceAccDataType = F32; using ReduceAccDataType = F32;
using DDataType = F32; using DDataType = F32;
using DPtrsGlobal = ck::Tuple<DDataType*, DDataType*>; using DPtrsGlobal = ck::Tuple<DDataType*, DDataType*>;
using AccDataType = F32;
using ALayout = ck::tensor_layout::gemm::RowMajor; using ALayout = ck::tensor_layout::gemm::RowMajor;
using BLayout = ck::tensor_layout::gemm::ColumnMajor; using BLayout = ck::tensor_layout::gemm::ColumnMajor;
...@@ -47,10 +47,12 @@ using DxsReduceOp = ck::Tuple<D0ReduceOp, D1ReduceOp>; ...@@ -47,10 +47,12 @@ using DxsReduceOp = ck::Tuple<D0ReduceOp, D1ReduceOp>;
using UnaryIdenticElementOp = using UnaryIdenticElementOp =
ck::tensor_operation::element_wise::UnaryIdentic<ReduceAccDataType, ReduceAccDataType, false>; ck::tensor_operation::element_wise::UnaryIdentic<ReduceAccDataType, ReduceAccDataType, false>;
using UnaryDivElementOp =
ck::tensor_operation::element_wise::UnaryIdentic<ReduceAccDataType, ReduceAccDataType, true>;
using UnarySquareElementOp = using UnarySquareElementOp =
ck::tensor_operation::element_wise::UnarySquare<ReduceAccDataType, ReduceAccDataType, false>; ck::tensor_operation::element_wise::UnarySquare<ReduceAccDataType, ReduceAccDataType, false>;
using DxsInElementOp = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>; using DxsInElementOp = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>;
using DxsOutElementOp = ck::Tuple<UnaryIdenticElementOp, UnaryIdenticElementOp>; using DxsOutElementOp = ck::Tuple<UnaryDivElementOp, UnaryDivElementOp>;
using DGlobalMemOp = using DGlobalMemOp =
ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd, ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
...@@ -61,15 +63,35 @@ static constexpr auto GemmSpecialization = ...@@ -61,15 +63,35 @@ static constexpr auto GemmSpecialization =
// clang-format off // clang-format off
using DeviceGemmReduceInstance = ck::tensor_operation::device::DeviceGemmReduce_Xdl_CShuffle using DeviceGemmReduceInstance = ck::tensor_operation::device::DeviceGemmReduce_Xdl_CShuffle
//######| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| Dxs| DxsInEleOp| DxsOutEleOp| D| 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| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy| //######| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| Dxs| DxsInEleOp| DxsAccEleOp| D| 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| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
//######| | | | Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Reduce| | | MemoryData| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector| //######| | | | Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Reduce| | | MemoryData| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector|
//######| | | | | | | | | | | Operation| Operation| 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_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock| //######| | | | | | | | | | | Operation| Operation| 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_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | //######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, AElementOp, BElementOp, CElementOp, DxsReduceOp, DxsInElementOp, DxsOutElementOp, DGlobalMemOp, GemmSpecialization, 1, 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, S<64, 4>, 4, 1>; < Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, AElementOp, BElementOp, CElementOp, DxsReduceOp, DxsInElementOp, DxsOutElementOp, DGlobalMemOp, GemmSpecialization, 1, 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, S<64, 4>, 4, 1>;
// clang-format on // clang-format on
using ReferenceGemmInstance = ck::tensor_operation::host:: using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>; BDataType,
CDataType,
GemmAccDataType,
AElementOp,
BElementOp,
CElementOp>;
template <typename ADataType, typename BDataType, typename CDataType, typename DDataType>
void DumpGemmLayerNormPerf(float gemm_reduce_time, int M, int N, int K)
{
std::size_t gemm_flop = std::size_t(2) * M * N * K;
std::size_t gemm_num_byte = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
sizeof(CDataType) * M * N + sizeof(DDataType) * M +
sizeof(DDataType) * M;
float tflops = static_cast<float>(gemm_flop) / 1.E9 / gemm_reduce_time;
float gemm_gb_per_sec = gemm_num_byte / 1.E6 / gemm_reduce_time;
std::cout << "gemm + reduce_mean + reduce_mean_square Perf: " << gemm_reduce_time << " ms, "
<< tflops << " TFlops, " << gemm_gb_per_sec << " GB/s, " << std::endl;
}
int main(int argc, char* argv[]) int main(int argc, char* argv[])
{ {
...@@ -182,6 +204,9 @@ int main(int argc, char* argv[]) ...@@ -182,6 +204,9 @@ int main(int argc, char* argv[])
auto dxs_global = ck::make_tuple(static_cast<DDataType*>(d0_device_buf.GetDeviceBuffer()), auto dxs_global = ck::make_tuple(static_cast<DDataType*>(d0_device_buf.GetDeviceBuffer()),
static_cast<DDataType*>(d1_device_buf.GetDeviceBuffer())); static_cast<DDataType*>(d1_device_buf.GetDeviceBuffer()));
auto dxs_in_element_op = DxsInElementOp{};
auto dxs_out_element_op = DxsOutElementOp{M, M};
// do GEMM // do GEMM
auto gemm = DeviceGemmReduceInstance{}; auto gemm = DeviceGemmReduceInstance{};
auto invoker = gemm.MakeInvoker(); auto invoker = gemm.MakeInvoker();
...@@ -198,8 +223,8 @@ int main(int argc, char* argv[]) ...@@ -198,8 +223,8 @@ int main(int argc, char* argv[])
a_element_op, a_element_op,
b_element_op, b_element_op,
c_element_op, c_element_op,
DxsInElementOp{}, dxs_in_element_op,
DxsOutElementOp{}); dxs_out_element_op);
if(!gemm.IsSupportedArgument(argument)) if(!gemm.IsSupportedArgument(argument))
{ {
...@@ -214,19 +239,7 @@ int main(int argc, char* argv[]) ...@@ -214,19 +239,7 @@ int main(int argc, char* argv[])
// if time_kernel == true, kernel will run multiple times. This kernel use atomic-add so result // if time_kernel == true, kernel will run multiple times. This kernel use atomic-add so result
// will not be correct. need to set time_kernel = false for correctness test // will not be correct. need to set time_kernel = false for correctness test
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel}); invoker.Run(argument, StreamConfig{nullptr, false});
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype =
sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(CDataType) * M * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
<< gemm.GetTypeString() << std::endl;
bool pass = true; bool pass = true;
if(do_verification) if(do_verification)
...@@ -257,12 +270,14 @@ int main(int argc, char* argv[]) ...@@ -257,12 +270,14 @@ int main(int argc, char* argv[])
float d0_val = 0; float d0_val = 0;
float d1_val = 0; float d1_val = 0;
UnaryIdenticElementOp{}(d0_val, c_val); dxs_in_element_op(ck::Number<0>{})(d0_val, c_val);
UnarySquareElementOp{}(d1_val, c_val); dxs_in_element_op(ck::Number<1>{})(d1_val, c_val);
d0_reduce_op(d0_acc, d0_val); d0_reduce_op(d0_acc, d0_val);
d1_reduce_op(d1_acc, d1_val); d1_reduce_op(d1_acc, d1_val);
} }
dxs_out_element_op(ck::Number<0>{})(d0_acc, d0_acc);
dxs_out_element_op(ck::Number<1>{})(d1_acc, d1_acc);
d0_m_host_result(m) = ck::type_convert<DDataType>(d0_acc); d0_m_host_result(m) = ck::type_convert<DDataType>(d0_acc);
d1_m_host_result(m) = ck::type_convert<DDataType>(d1_acc); d1_m_host_result(m) = ck::type_convert<DDataType>(d1_acc);
} }
...@@ -282,5 +297,12 @@ int main(int argc, char* argv[]) ...@@ -282,5 +297,12 @@ int main(int argc, char* argv[])
1e-5); 1e-5);
} }
if(time_kernel)
{
float ave_time = invoker.Run(argument, StreamConfig{nullptr, true});
DumpGemmLayerNormPerf<ADataType, BDataType, CDataType, DDataType>(ave_time, M, N, K);
}
return pass ? 0 : 1; return pass ? 0 : 1;
} }
example/18_batched_gemm_reduce/batched_gemm_reduce_xdl_fp16.cpp
View file @ c3de33fa
...@@ -59,7 +59,7 @@ static constexpr auto GemmSpecialization = ...@@ -59,7 +59,7 @@ static constexpr auto GemmSpecialization =
// clang-format off // clang-format off
using DeviceBatchedGemmReduceInstance = ck::tensor_operation::device::DeviceBatchedGemmReduce_Xdl_CShuffle using DeviceBatchedGemmReduceInstance = ck::tensor_operation::device::DeviceBatchedGemmReduce_Xdl_CShuffle
//######| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| Dxs| DxsInEleOp| DxsOutEleOp| D| 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| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy| //######| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| Dxs| DxsInEleOp| DxsAccEleOp| D| 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| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
//######| | | | Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Reduce| | | MemoryData| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector| //######| | | | Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Reduce| | | MemoryData| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector|
//######| | | | | | | | | | | Operation| Operation| 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_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock| //######| | | | | | | | | | | Operation| Operation| 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_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | //######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
......
example/21_gemm_layernorm/CMakeLists.txt 0 → 100644
View file @ c3de33fa
add_example_executable(example_gemm_layernorm_xdl_fp16 gemm_layernorm_xdl_fp16.cpp)
example/21_gemm_layernorm/gemm_layernorm_xdl_fp16.cpp 0 → 100644
View file @ c3de33fa
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <stdlib.h>
#include "check_err.hpp"
#include "config.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "device_tensor.hpp"
#include "device_5ary_elementwise.hpp"
#include "device_gemm_reduce_xdl_cshuffle.hpp"
#include "element_wise_operation.hpp"
#include "reference_gemm.hpp"
#include "gemm_specialization.hpp"
#include "element_wise_reduce_operation.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using ADataType = F16;
using BDataType = F16;
using CDataType = F16;
using GemmAccDataType = F32;
using ReduceAccDataType = F32;
using DDataType = F32;
using DPtrsGlobal = ck::Tuple<DDataType*, DDataType*>;
using GammaDataType = F16;
using BetaDataType = F16;
using LayerNormOutDataType = F16;
using NormalizeComputeDataType = F32;
using ALayout = ck::tensor_layout::gemm::RowMajor;
using BLayout = ck::tensor_layout::gemm::ColumnMajor;
using CLayout = ck::tensor_layout::gemm::RowMajor;
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
using ReduceSumOp = ck::reduce::Add<ReduceAccDataType>;
using DxsReduceOp = ck::Tuple<ReduceSumOp, ReduceSumOp>;
using UnaryIdenticElementOp =
ck::tensor_operation::element_wise::UnaryIdentic<ReduceAccDataType, ReduceAccDataType, false>;
using UnaryDivElementOp =
ck::tensor_operation::element_wise::UnaryIdentic<ReduceAccDataType, ReduceAccDataType, true>;
using UnarySquareElementOp =
ck::tensor_operation::element_wise::UnarySquare<ReduceAccDataType, ReduceAccDataType, false>;
using DxsInElementOp = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>;
using DxsOutElementOp = ck::Tuple<UnaryDivElementOp, UnaryDivElementOp>;
using DxsGlobalMemOp =
ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
ck::InMemoryDataOperationEnum::AtomicAdd>;
static constexpr auto GemmSpecialization =
ck::tensor_operation::device::GemmSpecialization::Default;
// clang-format off
using DeviceGemmReduceInstance = ck::tensor_operation::device::DeviceGemmReduce_Xdl_CShuffle
//######| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| Dxs| DxsInEleOp| DxsAccEleOp| D| 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| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
//######| | | | Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Reduce| | | MemoryData| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector|
//######| | | | | | | | | | | Operation| Operation| 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_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, AElementOp, BElementOp, CElementOp, DxsReduceOp, DxsInElementOp, DxsOutElementOp, DxsGlobalMemOp, GemmSpecialization, 1, 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, S<64, 4>, 4, 1>;
// clang-format on
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
CDataType,
GemmAccDataType,
AElementOp,
BElementOp,
CElementOp>;
using NormalizeFunctor = ck::tensor_operation::element_wise::Normalize;
// A:x, B:E[x], C:E[x^2], D:Gamma, E:Beta , F:y
using DeviceNormalizeInstance =
ck::tensor_operation::device::Device5AryElementwise<CDataType,
DDataType,
DDataType,
GammaDataType,
BetaDataType,
LayerNormOutDataType,
NormalizeComputeDataType,
NormalizeFunctor,
2,
8,
8, // scalarPerVector: gemm_out
1, // scalarPerVector: reduce_mean
1, // scalarPerVector: reduce_mean_square
8, // scalarPerVector: Gamma
8, // scalarPerVector: Beta
8>; // scalarPerVector: LayerNorm_out
auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
return HostTensorDescriptor(std::vector<std::size_t>({len}),
std::vector<std::size_t>({stride}));
};
auto f_host_tensor_descriptor2d =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({stride, 1}));
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({1, stride}));
}
};
template <typename CDataType,
typename DDataType,
typename A_functor,
typename B_functor,
typename C_functor>
void host_gemm_layernorm(Tensor<LayerNormOutDataType>& out_m_n,
const Tensor<ADataType>& a_m_k,
const Tensor<ADataType>& b_k_n,
const Tensor<GammaDataType>& gamma_n,
const Tensor<GammaDataType>& beta_n,
A_functor a_element_op,
B_functor b_element_op,
C_functor c_element_op,
int M,
int N)
{
using out_type = ck::remove_reference_t<decltype(out_m_n(0, 0))>;
int StrideC = N;
Tensor<CDataType> c_m_n(f_host_tensor_descriptor2d(M, N, StrideC, CLayout{}));
Tensor<DDataType> mean_m(f_host_tensor_descriptor1d(M, 1));
Tensor<DDataType> meanSquare_m(f_host_tensor_descriptor1d(M, 1));
auto averageOpInst = UnaryDivElementOp{M};
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument =
ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, c_element_op);
ref_invoker.Run(ref_argument);
// reduce_mean and reduce_square_mean
auto reduceSumOpInst = ReduceSumOp{};
for(int m = 0; m < M; ++m)
{
float mean_acc = reduceSumOpInst.GetReductionZeroVal();
float square_mean_acc = reduceSumOpInst.GetReductionZeroVal();
for(int n = 0; n < N; ++n)
{
ReduceAccDataType c_val = ck::type_convert<float>(c_m_n(m, n));
ReduceAccDataType square_c_val = 0;
UnarySquareElementOp{}(square_c_val, c_val);
reduceSumOpInst(mean_acc, c_val);
reduceSumOpInst(square_mean_acc, square_c_val);
}
averageOpInst(mean_acc, mean_acc);
averageOpInst(square_mean_acc, square_mean_acc);
mean_m(m) = ck::type_convert<DDataType>(mean_acc);
meanSquare_m(m) = ck::type_convert<DDataType>(square_mean_acc);
}
// LayerNorm
auto layerNormInst = NormalizeFunctor{};
for(int m = 0; m < M; ++m)
{
for(int n = 0; n < N; ++n)
{
float out_f32 = 0;
layerNormInst(out_f32, c_m_n(m, n), mean_m(m), meanSquare_m(m), gamma_n(n), beta_n(n));
out_m_n(m, n) = static_cast<out_type>(out_f32);
}
}
}
template <typename ADataType,
typename BDataType,
typename CDataType,
typename DDataType,
typename GammaDataType,
typename BetaDataType,
typename NormalizeDataType>
void DumpGemmLayerNormPerf(float gemm_reduce_time, float normalize_time, int M, int N, int K)
{
std::size_t gemm_flop = std::size_t(2) * M * N * K;
std::size_t gemm_num_byte = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
sizeof(CDataType) * M * N + sizeof(DDataType) * M +
sizeof(DDataType) * M;
std::size_t normalize_num_btye = sizeof(CDataType) * M * N + sizeof(DDataType) * M +
sizeof(DDataType) * M + sizeof(GammaDataType) * N +
sizeof(BetaDataType) * N + sizeof(NormalizeDataType) * M * N;
float tflops = static_cast<float>(gemm_flop) / 1.E9 / gemm_reduce_time;
float gemm_gb_per_sec = gemm_num_byte / 1.E6 / gemm_reduce_time;
float normalize_gb_per_sec = normalize_num_btye / 1.E6 / normalize_time;
std::cout << "gemm + reduce_mean + reduce_square_mean Perf: " << gemm_reduce_time << " ms, "
<< tflops << " TFlops, " << gemm_gb_per_sec << " GB/s, " << std::endl;
std::cout << "5-ary elementwise Perf: " << normalize_time << " ms, " << normalize_gb_per_sec
<< " GB/s, " << std::endl;
}
int main()
{
// GEMM shape
ck::index_t M = 1024;
ck::index_t N = 1024;
ck::index_t K = 1024;
ck::index_t StrideA = 1024;
ck::index_t StrideB = 1024;
ck::index_t StrideC = 1024;
Tensor<ADataType> a_m_k(f_host_tensor_descriptor2d(M, K, StrideA, ALayout{}));
Tensor<BDataType> b_k_n(f_host_tensor_descriptor2d(K, N, StrideB, BLayout{}));
Tensor<CDataType> c_m_n(f_host_tensor_descriptor2d(M, N, StrideC, CLayout{}));
Tensor<DDataType> reduceMean_m(f_host_tensor_descriptor1d(M, 1));
Tensor<DDataType> reduceMeanSquare_m(f_host_tensor_descriptor1d(M, 1));
Tensor<GammaDataType> gamma_n(f_host_tensor_descriptor1d(N, 1));
Tensor<BetaDataType> beta_n(f_host_tensor_descriptor1d(N, 1));
Tensor<LayerNormOutDataType> layerNorm_m_n(
f_host_tensor_descriptor2d(M, N, StrideC, CLayout{}));
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{-1, 1});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-1, 1});
gamma_n.GenerateTensorValue(GeneratorTensor_3<GammaDataType>{-1, 1});
beta_n.GenerateTensorValue(GeneratorTensor_3<BetaDataType>{-1, 1});
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
DeviceMem c_device_buf(sizeof(CDataType) * c_m_n.mDesc.GetElementSpace());
DeviceMem reduceMean_device_buf(sizeof(DDataType) * reduceMean_m.mDesc.GetElementSpace());
DeviceMem reduceMeanSquare_device_buf(sizeof(DDataType) *
reduceMeanSquare_m.mDesc.GetElementSpace());
DeviceMem gamma_device_buf(sizeof(GammaDataType) * gamma_n.mDesc.GetElementSpace());
DeviceMem beta_device_buf(sizeof(BetaDataType) * beta_n.mDesc.GetElementSpace());
DeviceMem layerNorm_device_buf(sizeof(LayerNormOutDataType) *
layerNorm_m_n.mDesc.GetElementSpace());
a_device_buf.ToDevice(a_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
gamma_device_buf.ToDevice(gamma_n.mData.data());
beta_device_buf.ToDevice(beta_n.mData.data());
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto c_element_op = CElementOp{};
auto dxs_global =
ck::make_tuple(static_cast<DDataType*>(reduceMean_device_buf.GetDeviceBuffer()),
static_cast<DDataType*>(reduceMeanSquare_device_buf.GetDeviceBuffer()));
auto dxs_in_element_op = DxsInElementOp{};
auto dxs_out_element_op = DxsOutElementOp{M, M};
// Prepare GEMM, reduce_mean, reduce_mean_square
auto gemmReduce = DeviceGemmReduceInstance{};
auto gemmReduce_invoker = gemmReduce.MakeInvoker();
auto gemmReduce_argument =
gemmReduce.MakeArgument(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
dxs_global,
M,
N,
K,
StrideA,
StrideB,
StrideC,
a_element_op,
b_element_op,
c_element_op,
dxs_in_element_op,
dxs_out_element_op);
if(!gemmReduce.IsSupportedArgument(gemmReduce_argument))
{
throw std::runtime_error(
"wrong! device_gemm with the specified compilation parameters does "
"not support this GEMM problem");
}
reduceMean_device_buf.SetZero();
reduceMeanSquare_device_buf.SetZero();
// Prepare LayerNorm
auto normalize = DeviceNormalizeInstance{};
auto normalize_invoker = normalize.MakeInvoker();
auto normalize_argument = normalize.MakeArgument(
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
static_cast<DDataType*>(reduceMean_device_buf.GetDeviceBuffer()),
static_cast<DDataType*>(reduceMeanSquare_device_buf.GetDeviceBuffer()),
static_cast<GammaDataType*>(gamma_device_buf.GetDeviceBuffer()),
static_cast<BetaDataType*>(beta_device_buf.GetDeviceBuffer()),
static_cast<LayerNormOutDataType*>(layerNorm_device_buf.GetDeviceBuffer()),
{M, N},
{StrideC, 1},
{1, 0},
{1, 0},
{0, 1},
{0, 1},
{StrideC, 1},
NormalizeFunctor{});
if(!normalize.IsSupportedArgument(normalize_argument))
{
throw std::runtime_error("The runtime parameters seems not supported by the "
"Device5AryElementwise instance, exiting!");
}
// run kernel
gemmReduce_invoker.Run(gemmReduce_argument, StreamConfig{nullptr, false});
normalize_invoker.Run(normalize_argument, StreamConfig{nullptr, false});
bool pass = true;
{
// verification
Tensor<LayerNormOutDataType> host_layerNorm_m_n(
f_host_tensor_descriptor2d(M, N, StrideC, CLayout{}));
host_gemm_layernorm<CDataType, DDataType>(host_layerNorm_m_n,
a_m_k,
b_k_n,
gamma_n,
beta_n,
a_element_op,
b_element_op,
c_element_op,
M,
N);
layerNorm_device_buf.FromDevice(layerNorm_m_n.mData.data());
pass &= ck::utils::check_err(layerNorm_m_n.mData,
host_layerNorm_m_n.mData,
"Error: Incorrect results d1",
1e-3,
1e-3);
}
{
// evaluate kernel perf
bool time_kernel = true;
float gemm_reduce_mean_reduce_square_mean_ave_time =
gemmReduce_invoker.Run(gemmReduce_argument, StreamConfig{nullptr, time_kernel});
float normalize_ave_time =
normalize_invoker.Run(normalize_argument, StreamConfig{nullptr, time_kernel});
if(time_kernel)
DumpGemmLayerNormPerf<ADataType,
BDataType,
CDataType,
DDataType,
GammaDataType,
BetaDataType,
LayerNormOutDataType>(
gemm_reduce_mean_reduce_square_mean_ave_time, normalize_ave_time, M, N, K);
}
return pass ? 0 : 1;
}
example/CMakeLists.txt
View file @ c3de33fa
...@@ -54,3 +54,4 @@ add_subdirectory(16_gemm_reduce) ...@@ -54,3 +54,4 @@ add_subdirectory(16_gemm_reduce)
add_subdirectory(18_batched_gemm_reduce) add_subdirectory(18_batched_gemm_reduce)
add_subdirectory(19_binary_elementwise) add_subdirectory(19_binary_elementwise)
add_subdirectory(20_convnd_bwd_weight_xdl) add_subdirectory(20_convnd_bwd_weight_xdl)
add_subdirectory(21_gemm_layernorm)
include/ck/tensor_operation/gpu/device/device_5ary_elementwise.hpp 0 → 100644
View file @ c3de33fa
#pragma once
#include <iostream>
#include <sstream>
#include "device.hpp"
#include "device_base.hpp"
#include "common_header.hpp"
#include "gridwise_5ary_Elementwise_1d.hpp"
#include "tensor_layout.hpp"
#include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
template <typename ADataType,
typename BDataType,
typename CDataType,
typename DDataType,
typename EDataType,
typename FDataType,
typename ComputeDataType,
typename ElementwiseFunctor,
index_t NDim,
index_t MPerThread,
index_t AScalarPerVector,
index_t BScalarPerVector,
index_t CScalarPerVector,
index_t DScalarPerVector,
index_t EScalarPerVector,
index_t FScalarPerVector>
struct Device5AryElementwise : public BaseOperator
{
static constexpr auto I0 = Number<0>{};
template <typename Desc_M>
static auto PadDescriptor_M_1d(Desc_M desc_m, index_t gridSize, index_t blockSize)
{
const auto m = desc_m.GetLength(I0);
const index_t loop_step = gridSize * blockSize * MPerThread;
const auto pad = math::integer_least_multiple(m, loop_step) - m;
const auto desc_m_pad =
transform_tensor_descriptor(desc_m,
make_tuple(make_right_pad_transform(m, pad)),
make_tuple(Sequence<0>{}),
make_tuple(Sequence<0>{}));
return desc_m_pad;
}
static auto MakeDescriptor_M(const std::vector<index_t>& lengths,
const std::vector<index_t>& stride,
index_t gridSize,
index_t blockSize)
{
auto tupleOfShape = generate_tuple([&](auto I) { return lengths[I]; }, Number<NDim>{});
auto tupleOfStride = generate_tuple([&](auto I) { return stride[I]; }, Number<NDim>{});
// nd desc - [s0, s1, s2, ...]
const auto desc = make_naive_tensor_descriptor(tupleOfShape, tupleOfStride);
// merge nd to 1d desc - [s0 * s1 * ...]
if constexpr(NDim > 1)
{
const auto desc_m = transform_tensor_descriptor(
desc,
make_tuple(make_merge_transform(tupleOfShape)),
make_tuple(generate_sequence_v2([&](auto I) { return I; }, Number<NDim>{})),
make_tuple(Sequence<0>{}));
return PadDescriptor_M_1d(desc_m, gridSize, blockSize);
}
else
return PadDescriptor_M_1d(desc, gridSize, blockSize);
}
using AGridDesc_M = decltype(MakeDescriptor_M({1, 1}, {1, 1}, 1, 1));
using BGridDesc_M = decltype(MakeDescriptor_M({1, 1}, {1, 1}, 1, 1));
using CGridDesc_M = decltype(MakeDescriptor_M({1, 1}, {1, 1}, 1, 1));
using DGridDesc_M = decltype(MakeDescriptor_M({1, 1}, {1, 1}, 1, 1));
using EGridDesc_M = decltype(MakeDescriptor_M({1, 1}, {1, 1}, 1, 1));
using FGridDesc_M = decltype(MakeDescriptor_M({1, 1}, {1, 1}, 1, 1));
using Gridwise5AryEltwise = Gridwise5AryElementwise_1D<ADataType,
BDataType,
CDataType,
DDataType,
EDataType,
FDataType,
ComputeDataType,
AGridDesc_M,
BGridDesc_M,
CGridDesc_M,
DGridDesc_M,
EGridDesc_M,
FGridDesc_M,
ElementwiseFunctor,
MPerThread,
AScalarPerVector,
BScalarPerVector,
CScalarPerVector,
DScalarPerVector,
EScalarPerVector,
FScalarPerVector>;
struct Argument : public BaseArgument
{
Argument(const ADataType* p_a,
const BDataType* p_b,
const CDataType* p_c,
const DDataType* p_d,
const EDataType* p_e,
FDataType* p_f,
const std::vector<index_t>& lengths,
const std::vector<index_t>& a_strides,
const std::vector<index_t>& b_strides,
const std::vector<index_t>& c_strides,
const std::vector<index_t>& d_strides,
const std::vector<index_t>& e_strides,
const std::vector<index_t>& f_strides,
ElementwiseFunctor functor)
: p_a_(p_a),
p_b_(p_b),
p_c_(p_c),
p_d_(p_d),
p_e_(p_e),
p_f_(p_f),
lengths_(lengths),
a_strides_(a_strides),
b_strides_(b_strides),
c_strides_(c_strides),
d_strides_(d_strides),
e_strides_(e_strides),
f_strides_(f_strides),
functor_(functor),
blockSize_(256),
gridSize_(120) // FIXME - Calculate the grid size by number of CU in the future
{
a_grid_desc_m_ = MakeDescriptor_M(lengths, a_strides, gridSize_, blockSize_);
b_grid_desc_m_ = MakeDescriptor_M(lengths, b_strides, gridSize_, blockSize_);
c_grid_desc_m_ = MakeDescriptor_M(lengths, c_strides, gridSize_, blockSize_);
d_grid_desc_m_ = MakeDescriptor_M(lengths, d_strides, gridSize_, blockSize_);
e_grid_desc_m_ = MakeDescriptor_M(lengths, e_strides, gridSize_, blockSize_);
f_grid_desc_m_ = MakeDescriptor_M(lengths, f_strides, gridSize_, blockSize_);
}
const ADataType* p_a_;
const BDataType* p_b_;
const CDataType* p_c_;
const DDataType* p_d_;
const EDataType* p_e_;
FDataType* p_f_;
std::vector<index_t> lengths_;
AGridDesc_M a_grid_desc_m_;
BGridDesc_M b_grid_desc_m_;
CGridDesc_M c_grid_desc_m_;
DGridDesc_M d_grid_desc_m_;
EGridDesc_M e_grid_desc_m_;
FGridDesc_M f_grid_desc_m_;
std::vector<index_t> a_strides_;
std::vector<index_t> b_strides_;
std::vector<index_t> c_strides_;
std::vector<index_t> d_strides_;
std::vector<index_t> e_strides_;
std::vector<index_t> f_strides_;
ElementwiseFunctor functor_;
index_t blockSize_;
index_t gridSize_;
};
struct Invoker : public BaseInvoker
{
float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
{
const auto kernel = kernel_5ary_elementwise_1d<Gridwise5AryEltwise,
ADataType,
BDataType,
CDataType,
DDataType,
EDataType,
FDataType,
AGridDesc_M,
BGridDesc_M,
CGridDesc_M,
DGridDesc_M,
EGridDesc_M,
FGridDesc_M,
ElementwiseFunctor>;
float elapsed_time = launch_and_time_kernel(stream_config,
kernel,
dim3(arg.gridSize_),
dim3(arg.blockSize_),
0,
arg.p_a_,
arg.p_b_,
arg.p_c_,
arg.p_d_,
arg.p_e_,
arg.p_f_,
arg.a_grid_desc_m_,
arg.b_grid_desc_m_,
arg.c_grid_desc_m_,
arg.d_grid_desc_m_,
arg.e_grid_desc_m_,
arg.f_grid_desc_m_,
arg.functor_);
return elapsed_time;
}
// polymorphic
float Run(const BaseArgument* p_arg,
const StreamConfig& stream_config = StreamConfig{}) override
{
return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
}
};
bool IsSupportedArgument(const BaseArgument& p_arg) { return IsSupportedArgument(&p_arg); }
bool IsSupportedArgument(const BaseArgument* p_arg) override
{
const Argument* pArg = dynamic_cast<const Argument*>(p_arg);
if(pArg == nullptr)
return false;
if(pArg->lengths_.size() != NDim)
return false;
if(pArg->lengths_.back() % MPerThread != 0)
return false;
auto IsScalarPerVectorValid = [](bool isLastDimensionCoalesced, int scalarPerVector) {
bool ret = true;
if(!isLastDimensionCoalesced)
ret = scalarPerVector == 1;
else
ret = MPerThread % scalarPerVector == 0;
return ret;
};
if(!IsScalarPerVectorValid(pArg->a_strides_.back() == 1, AScalarPerVector))
return false;
if(!IsScalarPerVectorValid(pArg->b_strides_.back() == 1, BScalarPerVector))
return false;
if(!IsScalarPerVectorValid(pArg->c_strides_.back() == 1, CScalarPerVector))
return false;
if(!IsScalarPerVectorValid(pArg->d_strides_.back() == 1, DScalarPerVector))
return false;
if(!IsScalarPerVectorValid(pArg->e_strides_.back() == 1, EScalarPerVector))
return false;
if(!IsScalarPerVectorValid(pArg->f_strides_.back() == 1, FScalarPerVector))
return false;
return true;
};
static auto MakeArgument(const ADataType* p_a,
const BDataType* p_b,
const CDataType* p_c,
const DDataType* p_d,
const EDataType* p_e,
FDataType* p_f,
std::vector<index_t> lengths,
std::vector<index_t> a_strides,
std::vector<index_t> b_strides,
std::vector<index_t> c_strides,
std::vector<index_t> d_strides,
std::vector<index_t> e_strides,
std::vector<index_t> f_strides,
ElementwiseFunctor functor)
{
return Argument{p_a,
p_b,
p_c,
p_d,
p_e,
p_f,
lengths,
a_strides,
b_strides,
c_strides,
d_strides,
e_strides,
f_strides,
functor};
}
std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
const void* p_b,
const void* p_c,
const void* p_d,
const void* p_e,
void* p_f,
std::vector<index_t> lengths,
std::vector<index_t> a_strides,
std::vector<index_t> b_strides,
std::vector<index_t> c_strides,
std::vector<index_t> d_strides,
std::vector<index_t> e_strides,
std::vector<index_t> f_strides,
ElementwiseFunctor functor)
{
return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
static_cast<const BDataType*>(p_b),
static_cast<const CDataType*>(p_c),
static_cast<const DDataType*>(p_d),
static_cast<const EDataType*>(p_e),
static_cast<FDataType*>(p_f),
lengths,
a_strides,
b_strides,
c_strides,
d_strides,
e_strides,
f_strides,
functor);
}
static auto MakeInvoker() { return Invoker{}; }
std::unique_ptr<BaseInvoker> MakeInvokerPointer() { return std::make_unique<Invoker>(); }
}; // namespace device
} // namespace device
} // namespace tensor_operation
} // namespace ck
include/ck/tensor_operation/gpu/device/device_batched_gemm_reduce_xdl_cshuffle.hpp
View file @ c3de33fa
...@@ -22,7 +22,7 @@ template <typename GridwiseGemm, ...@@ -22,7 +22,7 @@ template <typename GridwiseGemm,
typename BElementwiseOperation, typename BElementwiseOperation,
typename CElementwiseOperation, typename CElementwiseOperation,
typename DxsInElementwiseOperation, typename DxsInElementwiseOperation,
typename DxsOutElementwiseOperation, typename DxsAccElementwiseOperation,
typename AGridDesc_AK0_M_AK1, typename AGridDesc_AK0_M_AK1,
typename BGridDesc_BK0_N_BK1, typename BGridDesc_BK0_N_BK1,
typename CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock, typename CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
...@@ -44,7 +44,7 @@ __global__ void ...@@ -44,7 +44,7 @@ __global__ void
const BElementwiseOperation b_element_op, const BElementwiseOperation b_element_op,
const CElementwiseOperation c_element_op, const CElementwiseOperation c_element_op,
const DxsInElementwiseOperation dxs_in_element_op, const DxsInElementwiseOperation dxs_in_element_op,
const DxsOutElementwiseOperation dxs_out_element_op, const DxsAccElementwiseOperation dxs_out_element_op,
const AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1, const AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1,
const BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1, const BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1,
const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
...@@ -126,7 +126,7 @@ template <typename ALayout, ...@@ -126,7 +126,7 @@ template <typename ALayout,
typename CElementwiseOperation, typename CElementwiseOperation,
typename DxsReduceOperation, typename DxsReduceOperation,
typename DxsInElementwiseOperation, typename DxsInElementwiseOperation,
typename DxsOutElementwiseOperation, typename DxsAccElementwiseOperation,
typename DGlobalMemoryDataOperation, typename DGlobalMemoryDataOperation,
GemmSpecialization GemmSpec, GemmSpecialization GemmSpec,
index_t NumGemmKPrefetchStage, index_t NumGemmKPrefetchStage,
...@@ -167,7 +167,7 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<DPtrsGloba ...@@ -167,7 +167,7 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<DPtrsGloba
BElementwiseOperation, BElementwiseOperation,
CElementwiseOperation, CElementwiseOperation,
DxsInElementwiseOperation, DxsInElementwiseOperation,
DxsOutElementwiseOperation> DxsAccElementwiseOperation>
{ {
using DeviceOp = DeviceBatchedGemmReduce_Xdl_CShuffle; using DeviceOp = DeviceBatchedGemmReduce_Xdl_CShuffle;
...@@ -527,7 +527,7 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<DPtrsGloba ...@@ -527,7 +527,7 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<DPtrsGloba
CElementwiseOperation, CElementwiseOperation,
DxsReduceOperation, DxsReduceOperation,
DxsInElementwiseOperation, DxsInElementwiseOperation,
DxsOutElementwiseOperation, DxsAccElementwiseOperation,
InMemoryDataOperationEnum::Set, InMemoryDataOperationEnum::Set,
DGlobalMemoryDataOperation, DGlobalMemoryDataOperation,
AGridDesc_AK0_M_AK1, AGridDesc_AK0_M_AK1,
...@@ -587,7 +587,7 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<DPtrsGloba ...@@ -587,7 +587,7 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<DPtrsGloba
BElementwiseOperation b_element_op, BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op, CElementwiseOperation c_element_op,
DxsInElementwiseOperation dxs_in_element_op, DxsInElementwiseOperation dxs_in_element_op,
DxsOutElementwiseOperation dxs_out_element_op, DxsAccElementwiseOperation dxs_out_element_op,
index_t BatchCount) index_t BatchCount)
: p_a_grid_{p_a_grid}, : p_a_grid_{p_a_grid},
p_b_grid_{p_b_grid}, p_b_grid_{p_b_grid},
...@@ -645,7 +645,7 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<DPtrsGloba ...@@ -645,7 +645,7 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<DPtrsGloba
BElementwiseOperation b_element_op_; BElementwiseOperation b_element_op_;
CElementwiseOperation c_element_op_; CElementwiseOperation c_element_op_;
DxsInElementwiseOperation dxs_in_element_op_; DxsInElementwiseOperation dxs_in_element_op_;
DxsOutElementwiseOperation dxs_out_element_op_; DxsAccElementwiseOperation dxs_out_element_op_;
}; };
// Invoker // Invoker
...@@ -703,7 +703,7 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<DPtrsGloba ...@@ -703,7 +703,7 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<DPtrsGloba
BElementwiseOperation, BElementwiseOperation,
CElementwiseOperation, CElementwiseOperation,
DxsInElementwiseOperation, DxsInElementwiseOperation,
DxsOutElementwiseOperation, DxsAccElementwiseOperation,
DeviceOp::AGridDesc_AK0_M_AK1, DeviceOp::AGridDesc_AK0_M_AK1,
DeviceOp::BGridDesc_BK0_N_BK1, DeviceOp::BGridDesc_BK0_N_BK1,
typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock, typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
...@@ -746,7 +746,7 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<DPtrsGloba ...@@ -746,7 +746,7 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<DPtrsGloba
BElementwiseOperation, BElementwiseOperation,
CElementwiseOperation, CElementwiseOperation,
DxsInElementwiseOperation, DxsInElementwiseOperation,
DxsOutElementwiseOperation, DxsAccElementwiseOperation,
DeviceOp::AGridDesc_AK0_M_AK1, DeviceOp::AGridDesc_AK0_M_AK1,
DeviceOp::BGridDesc_BK0_N_BK1, DeviceOp::BGridDesc_BK0_N_BK1,
typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock, typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
...@@ -832,7 +832,7 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<DPtrsGloba ...@@ -832,7 +832,7 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<DPtrsGloba
BElementwiseOperation b_element_op, BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op, CElementwiseOperation c_element_op,
DxsInElementwiseOperation dxs_in_element_op, DxsInElementwiseOperation dxs_in_element_op,
DxsOutElementwiseOperation dxs_out_element_op, DxsAccElementwiseOperation dxs_out_element_op,
index_t BatchCount) index_t BatchCount)
{ {
return Argument{p_a, return Argument{p_a,
...@@ -870,7 +870,7 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<DPtrsGloba ...@@ -870,7 +870,7 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<DPtrsGloba
BElementwiseOperation b_element_op, BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op, CElementwiseOperation c_element_op,
DxsInElementwiseOperation dxs_in_element_op, DxsInElementwiseOperation dxs_in_element_op,
DxsOutElementwiseOperation dxs_out_element_op, DxsAccElementwiseOperation dxs_out_element_op,
index_t BatchCount) override index_t BatchCount) override
{ {
return std::make_unique<Argument>(static_cast<const ADataType*>(p_a), return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
......
include/ck/tensor_operation/gpu/device/device_convnd_fwd_xdl_nhwc_kyxc_nhwk.hpp
View file @ c3de33fa
#ifndef DEVICE_CONVND_FWD_XDL_NHWC_KYXC_NHWK_HPP #pragma once
#define DEVICE_CONVND_FWD_XDL_NHWC_KYXC_NHWK_HPP
#include <functional> #include <functional>
#include <iostream> #include <iostream>
...@@ -8,6 +7,7 @@ ...@@ -8,6 +7,7 @@
#include <sstream> #include <sstream>
#include "device.hpp" #include "device.hpp"
#include "device_prop.hpp"
#include "device_base.hpp" #include "device_base.hpp"
#include "device_conv_fwd.hpp" #include "device_conv_fwd.hpp"
#include "convolution_forward_specialization.hpp" #include "convolution_forward_specialization.hpp"
...@@ -858,6 +858,27 @@ struct DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K ...@@ -858,6 +858,27 @@ struct DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
static bool IsSupportedArgument(const Argument& arg) static bool IsSupportedArgument(const Argument& arg)
{ {
if(ck::get_device_name() == "gfx908")
{
if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, float> ||
is_same_v<AccDataType, int32_t>))
{
return false;
}
}
else if(ck::get_device_name() == "gfx90a")
{
if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, float> ||
is_same_v<AccDataType, int32_t> || is_same_v<AccDataType, double>))
{
return false;
}
}
else
{
return false;
}
// Input tensors can't be bigger than 2GB each. // Input tensors can't be bigger than 2GB each.
constexpr ck::long_index_t GB2 = (ck::long_index_t{1} << 31); constexpr ck::long_index_t GB2 = (ck::long_index_t{1} << 31);
...@@ -1021,4 +1042,3 @@ struct DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K ...@@ -1021,4 +1042,3 @@ struct DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
} // namespace device } // namespace device
} // namespace tensor_operation } // namespace tensor_operation
} // namespace ck } // namespace ck
#endif
include/ck/tensor_operation/gpu/device/device_gemm_dl.hpp
View file @ c3de33fa
...@@ -4,6 +4,7 @@ ...@@ -4,6 +4,7 @@
#include <sstream> #include <sstream>
#include "device.hpp" #include "device.hpp"
#include "device_prop.hpp"
#include "device_base.hpp" #include "device_base.hpp"
#include "device_gemm.hpp" #include "device_gemm.hpp"
#include "common_header.hpp" #include "common_header.hpp"
...@@ -13,7 +14,6 @@ ...@@ -13,7 +14,6 @@
#include "gemm_specialization.hpp" #include "gemm_specialization.hpp"
#include "element_wise_operation.hpp" #include "element_wise_operation.hpp"
#include "gridwise_gemm_dl_v1r3.hpp" #include "gridwise_gemm_dl_v1r3.hpp"
#include "device_prop.hpp"
namespace ck { namespace ck {
namespace tensor_operation { namespace tensor_operation {
......
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