mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-17 19:40:04 +00:00
gemm + layernorm (#261)
* Implement reduction meand and reduction square mean
* Refine file name
* Add reduce mean and square mean
* Fix parameter name
* Add normalize device op (not implement invoker::run())
* Remove epislon
* Refine deviceop
* Add 5ary elementwise for normalization
* Add layernorm example
* layerNorm verication
* Fix compiler error due to merge from develop
* Fix typo
* Fix compile error
* Refine naming
* [What] Suport non pointer for invoker and argument
[Why] Snyc coding style with gemm
* Refine folder name
* Refine class name
* Evaluate perf of the kernel
* Fix compile error
* [What] Refine perf evaluation in example of gemm + reduction
[Why] evaluation of gemm + reduction may cause verification fail. Because evaluation will not initial global memory
* clang-format
[ROCm/composable_kernel commit: d32a67a9b6]
This commit is contained in:
rocking5566
@@ -1,2 +1,2 @@
|
||||
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)
|
||||
|
||||
@@ -29,10 +29,10 @@ using Col = ck::tensor_layout::gemm::ColumnMajor;
|
||||
using ADataType = F16;
|
||||
using BDataType = F16;
|
||||
using CDataType = F16;
|
||||
using GemmAccDataType = F32;
|
||||
using ReduceAccDataType = F32;
|
||||
using DDataType = F64;
|
||||
using DPtrsGlobal = ck::Tuple<DDataType*>;
|
||||
using AccDataType = F32;
|
||||
|
||||
using ALayout = ck::tensor_layout::gemm::RowMajor;
|
||||
using BLayout = ck::tensor_layout::gemm::ColumnMajor;
|
||||
@@ -52,15 +52,34 @@ static constexpr auto GemmSpecialization =
|
||||
|
||||
// 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| 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|
|
||||
//######| | | | | | | | | | | 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>;
|
||||
// clang-format on
|
||||
|
||||
using ReferenceGemmInstance = ck::tensor_operation::host::
|
||||
ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
|
||||
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
|
||||
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[])
|
||||
{
|
||||
@@ -193,21 +212,10 @@ int main(int argc, char* argv[])
|
||||
"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());
|
||||
|
||||
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;
|
||||
invoker.Run(argument, StreamConfig{nullptr, false});
|
||||
|
||||
bool pass = true;
|
||||
|
||||
@@ -246,5 +254,13 @@ int main(int argc, char* argv[])
|
||||
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;
|
||||
}
|
||||
|
||||
@@ -29,10 +29,10 @@ 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 AccDataType = F32;
|
||||
|
||||
using ALayout = ck::tensor_layout::gemm::RowMajor;
|
||||
using BLayout = ck::tensor_layout::gemm::ColumnMajor;
|
||||
@@ -47,10 +47,12 @@ using DxsReduceOp = ck::Tuple<D0ReduceOp, D1ReduceOp>;
|
||||
|
||||
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<UnaryIdenticElementOp, UnaryIdenticElementOp>;
|
||||
using DxsOutElementOp = ck::Tuple<UnaryDivElementOp, UnaryDivElementOp>;
|
||||
|
||||
using DGlobalMemOp =
|
||||
ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
|
||||
@@ -61,15 +63,35 @@ static constexpr auto GemmSpecialization =
|
||||
|
||||
// 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| 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|
|
||||
//######| | | | | | | | | | | 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>;
|
||||
// clang-format on
|
||||
|
||||
using ReferenceGemmInstance = ck::tensor_operation::host::
|
||||
ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
|
||||
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
|
||||
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[])
|
||||
{
|
||||
@@ -182,6 +204,9 @@ int main(int argc, char* argv[])
|
||||
auto dxs_global = ck::make_tuple(static_cast<DDataType*>(d0_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
|
||||
auto gemm = DeviceGemmReduceInstance{};
|
||||
auto invoker = gemm.MakeInvoker();
|
||||
@@ -198,8 +223,8 @@ int main(int argc, char* argv[])
|
||||
a_element_op,
|
||||
b_element_op,
|
||||
c_element_op,
|
||||
DxsInElementOp{},
|
||||
DxsOutElementOp{});
|
||||
dxs_in_element_op,
|
||||
dxs_out_element_op);
|
||||
|
||||
if(!gemm.IsSupportedArgument(argument))
|
||||
{
|
||||
@@ -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
|
||||
// will not be correct. need to set time_kernel = false for correctness test
|
||||
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;
|
||||
|
||||
invoker.Run(argument, StreamConfig{nullptr, false});
|
||||
bool pass = true;
|
||||
|
||||
if(do_verification)
|
||||
@@ -257,12 +270,14 @@ int main(int argc, char* argv[])
|
||||
float d0_val = 0;
|
||||
float d1_val = 0;
|
||||
|
||||
UnaryIdenticElementOp{}(d0_val, c_val);
|
||||
UnarySquareElementOp{}(d1_val, c_val);
|
||||
dxs_in_element_op(ck::Number<0>{})(d0_val, c_val);
|
||||
dxs_in_element_op(ck::Number<1>{})(d1_val, c_val);
|
||||
d0_reduce_op(d0_acc, d0_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);
|
||||
d1_m_host_result(m) = ck::type_convert<DDataType>(d1_acc);
|
||||
}
|
||||
@@ -282,5 +297,12 @@ int main(int argc, char* argv[])
|
||||
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;
|
||||
}
|
||||
@@ -59,7 +59,7 @@ static constexpr auto GemmSpecialization =
|
||||
|
||||
// clang-format off
|
||||
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|
|
||||
//######| | | | | | | | | | | 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|
|
||||
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
|
||||
1
example/21_gemm_layernorm/CMakeLists.txt
Normal file
1
example/21_gemm_layernorm/CMakeLists.txt
Normal file
@@ -0,0 +1 @@
|
||||
add_example_executable(example_gemm_layernorm_xdl_fp16 gemm_layernorm_xdl_fp16.cpp)
|
||||
378
example/21_gemm_layernorm/gemm_layernorm_xdl_fp16.cpp
Normal file
378
example/21_gemm_layernorm/gemm_layernorm_xdl_fp16.cpp
Normal file
@@ -0,0 +1,378 @@
|
||||
#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;
|
||||
}
|
||||
@@ -54,3 +54,4 @@ add_subdirectory(16_gemm_reduce)
|
||||
add_subdirectory(18_batched_gemm_reduce)
|
||||
add_subdirectory(19_binary_elementwise)
|
||||
add_subdirectory(20_convnd_bwd_weight_xdl)
|
||||
add_subdirectory(21_gemm_layernorm)
|
||||
|
||||
Reference in New Issue
Block a user