Skip to content

GitLab

Commit b79df771 authored by carlushuang's avatar carlushuang
Browse files

Merge remote-tracking branch 'origin/develop' into cpu_avx2

parents 05d38218 63914743
Hide whitespace changes
Inline Side-by-side
Showing with 1035 additions and 425 deletions
example/01_gemm/gemm_dl_int8.cpp
View file @ b79df771
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream> #include <iostream>
#include <numeric> #include <numeric>
#include <initializer_list> #include <initializer_list>
#include <cstdlib> #include <cstdlib>
#include <stdlib.h>
#include <half.hpp> #include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "check_err.hpp" #include "ck/tensor_operation/gpu/device/device_gemm_dl.hpp"
#include "config.hpp" #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "device.hpp"
#include "host_tensor.hpp" #include "ck/library/utility/check_err.hpp"
#include "host_tensor_generator.hpp" #include "ck/library/host_tensor/device_memory.hpp"
#include "device_tensor.hpp" #include "ck/library/host_tensor/host_tensor.hpp"
#include "device_gemm_dl.hpp" #include "ck/library/host_tensor/host_tensor_generator.hpp"
#include "element_wise_operation.hpp" #include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "reference_gemm.hpp"
#include "gemm_specialization.hpp"
template <ck::index_t... Is> template <ck::index_t... Is>
using S = ck::Sequence<Is...>; using S = ck::Sequence<Is...>;
......
example/01_gemm/gemm_xdl_bf16.cpp
View file @ b79df771
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream> #include <iostream>
#include <numeric> #include <numeric>
#include <initializer_list> #include <initializer_list>
#include <cstdlib> #include <cstdlib>
#include <stdlib.h>
#include <half.hpp> #include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_xdl_cshuffle.hpp"
#include "check_err.hpp" #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "config.hpp" #include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "device.hpp"
#include "host_tensor.hpp" #include "ck/library/host_tensor/device_memory.hpp"
#include "host_tensor_generator.hpp" #include "ck/library/host_tensor/host_tensor.hpp"
#include "device_tensor.hpp" #include "ck/library/host_tensor/host_tensor_generator.hpp"
#include "device_gemm_xdl_cshuffle.hpp" #include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "element_wise_operation.hpp" #include "ck/library/utility/check_err.hpp"
#include "reference_gemm.hpp"
#include "gemm_specialization.hpp"
template <ck::index_t... Is> template <ck::index_t... Is>
using S = ck::Sequence<Is...>; using S = ck::Sequence<Is...>;
...@@ -83,8 +84,13 @@ using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle ...@@ -83,8 +84,13 @@ using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle
8>; // index_t CShuffleBlockTransferScalarPerVector_NPerBlock 8>; // index_t CShuffleBlockTransferScalarPerVector_NPerBlock
// clang-format on // clang-format on
using ReferenceGemmInstance = ck::tensor_operation::host:: using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
ReferenceGemm<float, float, float, float, PassThrough, PassThrough, PassThrough>; BDataType,
CDataType,
AccDataType,
PassThrough,
PassThrough,
PassThrough>;
int main(int argc, char* argv[]) int main(int argc, char* argv[])
{ {
...@@ -215,24 +221,17 @@ int main(int argc, char* argv[]) ...@@ -215,24 +221,17 @@ int main(int argc, char* argv[])
if(do_verification) if(do_verification)
{ {
Tensor<float> a_f32_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{})); Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<float> b_f32_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<float> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<float> c_m_n_device_f32_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
bf16_to_f32_(a_m_k, a_f32_m_k);
bf16_to_f32_(b_k_n, b_f32_k_n);
bf16_to_f32_(c_m_n_device_result, c_m_n_device_f32_result);
auto ref_gemm = ReferenceGemmInstance{}; auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker(); auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument( auto ref_argument = ref_gemm.MakeArgument(
a_f32_m_k, b_f32_k_n, c_m_n_host_result, a_element_op, b_element_op, c_element_op); a_m_k, b_k_n, c_m_n_host_result, a_element_op, b_element_op, c_element_op);
ref_invoker.Run(ref_argument); ref_invoker.Run(ref_argument);
return ck::utils::check_err(c_m_n_device_f32_result.mData, c_m_n_host_result.mData) ? 0 : 1; 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_fp16.cpp
View file @ b79df771
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream> #include <iostream>
#include <numeric> #include <numeric>
#include <initializer_list> #include <initializer_list>
#include <cstdlib> #include <cstdlib>
#include <stdlib.h>
#include <half.hpp> #include "ck/ck.hpp"
#include "check_err.hpp" #include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "config.hpp" #include "ck/tensor_operation/gpu/device/device_gemm_xdl.hpp"
#include "device.hpp" #include "ck/tensor_operation/gpu/device/device_gemm_xdl_cshuffle.hpp"
#include "host_tensor.hpp" #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "host_tensor_generator.hpp"
#include "device_tensor.hpp" #include "ck/library/utility/check_err.hpp"
#include "device_gemm_xdl.hpp" #include "ck/library/host_tensor/device_memory.hpp"
#include "device_gemm_xdl_cshuffle.hpp" #include "ck/library/host_tensor/host_tensor.hpp"
#include "element_wise_operation.hpp" #include "ck/library/host_tensor/host_tensor_generator.hpp"
#include "reference_gemm.hpp" #include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "gemm_specialization.hpp"
template <ck::index_t... Is> template <ck::index_t... Is>
using S = ck::Sequence<Is...>; using S = ck::Sequence<Is...>;
...@@ -27,30 +29,42 @@ using Col = ck::tensor_layout::gemm::ColumnMajor; ...@@ -27,30 +29,42 @@ using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough; using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ADataType = ck::half_t; using ADataType = F16;
using BDataType = ck::half_t; using BDataType = F16;
using CDataType = ck::half_t; using AccDataType = F32;
using AccDataType = float; using CShuffleDataType = F32;
using CDataType = F16;
using ALayout = ck::tensor_layout::gemm::RowMajor; using ALayout = Row;
using BLayout = ck::tensor_layout::gemm::ColumnMajor; using BLayout = Col;
using CLayout = ck::tensor_layout::gemm::RowMajor; using CLayout = Row;
using AElementOp = ck::tensor_operation::element_wise::PassThrough; using AElementOp = PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough; using BElementOp = PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough; using CElementOp = PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default; static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// clang-format off using DeviceGemmInstance0 = ck::tensor_operation::device::DeviceGemmXdl
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle // clang-format off
//######| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle| A| B| C| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| //######| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
//######| | | | Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| //######| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise|Spacialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar|
//######| | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| //######| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | //######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< Row, Col, Row, F16, F16, F16, F32, F32, AElementOp, BElementOp, CElementOp, GemmDefault, 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>; < ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout, AElementOp, BElementOp, CElementOp, GemmDefault, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 7, 1>;
// clang-format on
using DeviceGemmInstance1 = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle
// clang-format off
//######| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle| A| B| C| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//######| | | | Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//######| | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< ALayout, BLayout, CLayout, ADataType, BDataType, CDataType, AccDataType, CShuffleDataType, AElementOp, BElementOp, CElementOp, GemmDefault, 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>;
// clang-format on // clang-format on
using DeviceGemmInstance = DeviceGemmInstance0;
using ReferenceGemmInstance = ck::tensor_operation::host:: using ReferenceGemmInstance = ck::tensor_operation::host::
ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>; ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
...@@ -69,7 +83,11 @@ int main(int argc, char* argv[]) ...@@ -69,7 +83,11 @@ int main(int argc, char* argv[])
ck::index_t StrideB = 4096; ck::index_t StrideB = 4096;
ck::index_t StrideC = 4096; ck::index_t StrideC = 4096;
if(argc == 4) if(argc == 1)
{
// use default case
}
else if(argc == 4)
{ {
do_verification = std::stoi(argv[1]); do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]); init_method = std::stoi(argv[2]);
...@@ -93,7 +111,7 @@ int main(int argc, char* argv[]) ...@@ -93,7 +111,7 @@ int main(int argc, char* argv[])
{ {
printf("arg1: verification (0=no, 1=yes)\n"); printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"); printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=n0, 1=yes)\n"); printf("arg3: time kernel (0=no, 1=yes)\n");
printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n"); printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n");
exit(0); exit(0);
} }
......
example/01_gemm/gemm_xdl_fp64.cpp
View file @ b79df771
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream> #include <iostream>
#include <numeric> #include <numeric>
#include <initializer_list> #include <initializer_list>
#include <cstdlib> #include <cstdlib>
#include <stdlib.h>
#include <half.hpp> #include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "check_err.hpp" #include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "config.hpp" #include "ck/tensor_operation/gpu/device/device_gemm_xdl.hpp"
#include "device.hpp" #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp" #include "ck/library/host_tensor/device_memory.hpp"
#include "device_tensor.hpp" #include "ck/library/host_tensor/host_tensor.hpp"
#include "device_gemm_xdl.hpp" #include "ck/library/host_tensor/host_tensor_generator.hpp"
#include "device_gemm_xdl_cshuffle.hpp" #include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "element_wise_operation.hpp" #include "ck/library/utility/check_err.hpp"
#include "reference_gemm.hpp"
#include "gemm_specialization.hpp"
template <ck::index_t... Is> template <ck::index_t... Is>
using S = ck::Sequence<Is...>; using S = ck::Sequence<Is...>;
......
example/01_gemm/gemm_xdl_int8.cpp
View file @ b79df771
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream> #include <iostream>
#include <numeric> #include <numeric>
#include <initializer_list> #include <initializer_list>
#include <cstdlib> #include <cstdlib>
#include <stdlib.h>
#include <half.hpp> #include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "check_err.hpp" #include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "config.hpp" #include "ck/tensor_operation/gpu/device/device_gemm_xdl_cshuffle.hpp"
#include "device.hpp" #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp" #include "ck/library/utility/check_err.hpp"
#include "device_tensor.hpp" #include "ck/library/host_tensor/device_memory.hpp"
#include "device_gemm_xdl_cshuffle.hpp" #include "ck/library/host_tensor/host_tensor.hpp"
#include "element_wise_operation.hpp" #include "ck/library/host_tensor/host_tensor_generator.hpp"
#include "reference_gemm.hpp" #include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "gemm_specialization.hpp"
template <ck::index_t... Is> template <ck::index_t... Is>
using S = ck::Sequence<Is...>; using S = ck::Sequence<Is...>;
......
example/02_gemm_alpha_beta/CMakeLists.txt deleted 100644 → 0
View file @ 05d38218
add_example_executable(example_gemm_xdl_alpha_beta gemm_xdl_alpha_beta.cpp)
example/02_gemm_bilinear/CMakeLists.txt 0 → 100644
View file @ b79df771
add_example_executable(example_gemm_bilinear_xdl_fp16 gemm_bilinear_xdl_fp16.cpp)
example/02_gemm_alpha_beta/README.md → example/02_gemm_bilinear/README.md
View file @ b79df771
# Instructions for ```example_gemm_xdl_alpha_beta``` # Instructions for ```example_gemm_bilinear_xdl_fp16```
## Run ```example_gemm_xdl_alpha_beta``` ## Run ```example_gemm_bilinear_xdl_fp16```
```bash ```bash
#arg1: verification (0=no, 1=yes) #arg1: verification (0=no, 1=yes)
#arg2: initialization (0=no init, 1=integer value, 2=decimal value) #arg2: initialization (0=no init, 1=integer value, 2=decimal value)
#arg3: run kernel # of times (>1) #arg3: time kernel (0=no, 1=yes)
./bin/example_gemm_xdl_alpha_beta 1 1 1 0.5 0.5 #arg4 to 10: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD, StrideE
#arg11 to 12: alpha, beta
./bin/example_gemm_bilinear_xdl_fp16 1 1 1 3840 4096 4096 4096 4096 4096 4096 0.5 0.5
``` ```
Result (MI100 @ 1502Mhz, 184.6TFlops peak FP16) Result (MI100 @ 1502Mhz, 184.6TFlops peak FP16)
``` ```
......
example/02_gemm_bilinear/gemm_bilinear_xdl_fp16.cpp 0 → 100644
View file @ b79df771
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/host_tensor/device_memory.hpp"
#include "ck/library/host_tensor/host_tensor.hpp"
#include "ck/library/host_tensor/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "ck/library/utility/check_err.hpp"
struct AlphaBetaAdd
{
AlphaBetaAdd(float alpha, float beta) : alpha_(alpha), beta_(beta){};
template <typename E, typename C, typename D>
__host__ __device__ constexpr void operator()(E& e, const C& c, const D& d) const;
template <>
__host__ __device__ constexpr void operator()<ck::half_t, float, ck::half_t>(
ck::half_t& e, const float& c, const ck::half_t& d) const
{
e = ck::type_convert<ck::half_t>(alpha_ * c + beta_ * ck::type_convert<float>(d));
};
float alpha_;
float beta_;
};
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 PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ADataType = F16;
using BDataType = F16;
using AccDataType = F32;
using CShuffleDataType = F32;
using DDataType = F16;
using DsDataType = ck::Tuple<DDataType>;
using EDataType = F16;
using ALayout = Row;
using BLayout = Col;
using DELayout = Row;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = AlphaBetaAdd;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
using DeviceOpInstance =
ck::tensor_operation::device::DeviceGemmMultipleD_Xdl_CShuffle<ALayout,
BLayout,
DELayout,
ADataType,
BDataType,
AccDataType,
CShuffleDataType,
DsDataType,
EDataType,
AElementOp,
BElementOp,
CDEElementOp,
GemmDefault,
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>;
int main(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
// GEMM shape
ck::index_t M = 3840;
ck::index_t N = 4096;
ck::index_t K = 4096;
ck::index_t StrideA = 4096;
ck::index_t StrideB = 4096;
ck::index_t StrideD = 4096;
ck::index_t StrideE = 4096;
float alpha = 1.0f;
float beta = 1.0f;
if(argc == 1)
{
// use default case
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 6)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
alpha = std::stof(argv[4]);
beta = std::stof(argv[5]);
}
else if(argc == 13)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
M = std::stoi(argv[4]);
N = std::stoi(argv[5]);
K = std::stoi(argv[6]);
StrideA = std::stoi(argv[7]);
StrideB = std::stoi(argv[8]);
StrideD = std::stoi(argv[9]);
StrideE = std::stoi(argv[10]);
alpha = std::stof(argv[11]);
beta = std::stof(argv[12]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=no, 1=yes)\n");
printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD, StrideE, alpha, "
"beta\n");
exit(0);
}
auto f_host_tensor_descriptor =
[](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}));
}
};
Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<DDataType> d_m_n(f_host_tensor_descriptor(M, N, StrideD, DELayout{}));
Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, DELayout{}));
Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, DELayout{}));
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "d_m_n: " << d_m_n.mDesc << std::endl;
std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
d_m_n.GenerateTensorValue(GeneratorTensor_2<DDataType>{-5, 5});
break;
default:
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
d_m_n.GenerateTensorValue(GeneratorTensor_3<DDataType>{-0.5, 0.5});
}
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
DeviceMem d_device_buf(sizeof(DDataType) * d_m_n.mDesc.GetElementSpace());
DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpace());
a_device_buf.ToDevice(a_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
d_device_buf.ToDevice(d_m_n.mData.data());
e_device_buf.ToDevice(e_m_n_device_result.mData.data());
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto cde_element_op = CDEElementOp{alpha, beta};
// do GEMM
auto device_op = DeviceOpInstance{};
auto invoker = device_op.MakeInvoker();
auto argument =
device_op.MakeArgument(a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
std::array<const void*, 1>{d_device_buf.GetDeviceBuffer()},
e_device_buf.GetDeviceBuffer(),
M,
N,
K,
StrideA,
StrideB,
std::array<ck::index_t, 1>{StrideD},
StrideE,
a_element_op,
b_element_op,
cde_element_op);
if(!device_op.IsSupportedArgument(argument))
{
throw std::runtime_error(
"wrong! device_gemm with the specified compilation parameters does "
"not support this GEMM problem");
}
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(EDataType) * 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"
<< std::endl;
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
if(do_verification)
{
Tensor<CShuffleDataType> c_m_n(HostTensorDescriptor(
std::vector<std::size_t>{static_cast<std::size_t>(M), static_cast<std::size_t>(N)}));
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
CShuffleDataType,
AccDataType,
AElementOp,
BElementOp,
PassThrough>;
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, PassThrough{});
ref_invoker.Run(ref_argument);
for(int m = 0; m < M; ++m)
{
for(int n = 0; n < N; ++n)
{
cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d_m_n(m, n));
}
}
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
return ck::utils::check_err(e_m_n_device_result.mData, e_m_n_host_result.mData) ? 0 : 1;
}
return 0;
}
example/03_gemm_bias_relu/CMakeLists.txt
View file @ b79df771
add_example_executable(example_gemm_xdl_bias_relu gemm_xdl_bias_relu.cpp) add_example_executable(example_gemm_bias_relu_xdl_fp16 gemm_bias_relu_xdl_fp16.cpp)
example/03_gemm_bias_relu/README.md
View file @ b79df771
# Instructions for ```example_gemm_xdl_bias_relu_add``` # Instructions for ```example_gemm_bias_relu_xdl_fp16```
## Run ```example_gemm_xdl_bias_relu_add``` ## Run ```example_gemm_bias_relu_xdl_fp16```
```bash ```bash
#arg1: verification (0=no, 1=yes) #arg1: verification (0=no, 1=yes)
#arg2: initialization (0=no init, 1=integer value, 2=decimal value) #arg2: initialization (0=no init, 1=integer value, 2=decimal value)
#arg3: run kernel # of times (>1) #arg3: time kernel (0=no, 1=yes)
#arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC #arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideE
./bin/example_gemm_xdl_bias_relu_add 0 1 5 3840 4096 4096 4096 4096 4096 ./bin/example_gemm_bias_relu_xdl_fp16 1 1 1 3840 4096 4096 4096 4096 4096
```
Result (MI100 @ 1087Mhz, 133.5TFlops peak FP16)
```
a_m_k: dim 2, lengths {3840, 4096}, strides {4096, 1}
b_k_n: dim 2, lengths {4096, 4096}, strides {1, 4096}
c_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
c0_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
c1_m_n: dim 2, lengths {3840, 4096}, strides {1, 0}
arg.a_grid_desc_k0_m_k1_{512, 3840, 8}
arg.b_grid_desc_k0_n_k1_{512, 4096, 8}
arg.c_grid_desc_m_n_{ 3840, 4096}
arg.c0_grid_desc_m_n_{ 3840, 4096}
arg.c1_grid_desc_m_n_{ 3840, 4096}
launch_and_time_kernel: grid_dim {480, 1, 1}, block_dim {256, 1, 1}
Warm up
Start running 5 times...
Perf: 1.27583 ms, 100.992 TFlops, 73.9688 GB/s
``` ```
example/03_gemm_bias_relu/gemm_xdl_bias_relu.cpp → example/03_gemm_bias_relu/gemm_bias_relu_xdl_fp16.cpp
View file @ b79df771
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream> #include <iostream>
#include <numeric> #include <numeric>
#include <initializer_list> #include <initializer_list>
#include <cstdlib> #include <cstdlib>
#include <stdlib.h>
#include <half.hpp> #include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "check_err.hpp" #include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "config.hpp" #include "ck/tensor_operation/gpu/device/device_gemm_multiple_d_xdl_cshuffle.hpp"
#include "print.hpp" #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "device.hpp"
#include "host_tensor.hpp" #include "ck/library/host_tensor/device_memory.hpp"
#include "host_tensor_generator.hpp" #include "ck/library/host_tensor/host_tensor.hpp"
#include "host_gemm.hpp" #include "ck/library/host_tensor/host_tensor_generator.hpp"
#include "device_tensor.hpp" #include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "element_wise_operation.hpp" #include "ck/library/utility/check_err.hpp"
#include "device_gemm_xdl_c_shuffle_bias_activation.hpp"
#include "reference_gemm_bias_activation.hpp"
template <ck::index_t... Is> template <ck::index_t... Is>
using S = ck::Sequence<Is...>; using S = ck::Sequence<Is...>;
using ADataType = ck::half_t; using F16 = ck::half_t;
using BDataType = ck::half_t; using F32 = float;
using CDataType = ck::half_t;
using AccDataType = float; using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using ALayout = ck::tensor_layout::gemm::RowMajor;
using BLayout = ck::tensor_layout::gemm::ColumnMajor; using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using CLayout = ck::tensor_layout::gemm::RowMajor;
// C = A * B
using AElementOp = ck::tensor_operation::element_wise::PassThrough; // E = Relu(C + D);
using BElementOp = ck::tensor_operation::element_wise::PassThrough; struct AddRelu
using CElementOp = ck::tensor_operation::element_wise::AddRelu; {
__host__ __device__ void
// clang-format off operator()(ck::half_t& e, const ck::half_t& c, const ck::half_t& d) const
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdl_C_Shuffle_Bias_Activation< {
ADataType, // ADataType const ck::half_t x = c + d;
BDataType, // BDataType
CDataType, // CDataType e = x > 0 ? x : 0;
AccDataType, // AccDataType }
ALayout, // ALayout };
BLayout, // BLayout
CLayout, // CLayout using ADataType = F16;
AElementOp, // AElementwiseOperation using BDataType = F16;
BElementOp, // BElementwiseOperation using AccDataType = F32;
CElementOp, // CElementwiseOperation using CShuffleDataType = F16;
256, // BlockSize using DDataType = F16;
256, // MPerBlock using DsDataType = ck::Tuple<DDataType>;
128, // NPerBlock using EDataType = F16;
4, // K0PerBlock
8, // K1 using ALayout = Row;
32, // MPerXDL using BLayout = Col;
32, // NPerXDL using ELayout = Row;
4, // MXdlPerWave
2, // NXdlPerWave using AElementOp = PassThrough;
S<4, 64, 1>, // ABlockTransferThreadClusterLengths_K0_M_K1 using BElementOp = PassThrough;
S<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder using CDEElementOp = AddRelu;
S<1, 0, 2>, // ABlockTransferSrcAccessOrder
2, // ABlockTransferSrcVectorDim static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
8, // ABlockTransferSrcScalarPerVector
8, // ABlockTransferDstScalarPerVector_K1 using DeviceOpInstance =
true, // ABlockLdsAddExtraM ck::tensor_operation::device::DeviceGemmMultipleD_Xdl_CShuffle<ALayout,
S<4, 64, 1>, // BBlockTransferThreadClusterLengths_K0_N_K1 BLayout,
S<1, 0, 2>, // BBlockTransferThreadClusterArrangeOrder ELayout,
S<1, 0, 2>, // BBlockTransferSrcAccessOrder ADataType,
2, // BBlockTransferSrcVectorDim BDataType,
8, // BBlockTransferSrcScalarPerVector AccDataType,
8, // BBlockTransferDstScalarPerVector_K1 CShuffleDataType,
true, // BBlockLdsAddExtraN DsDataType,
1, // CShuffleMXdlPerWavePerShuffle EDataType,
1, // CShuffleNXdlPerWavePerShuffle AElementOp,
S<1, 1, 32, 1, 1, 8>, // CBlockTransferClusterLengths_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl BElementOp,
8>; // CBlockTransferScalarPerVector_NWaveNPerXdl CDEElementOp,
// clang-format on GemmDefault,
1,
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemmBiasActivation<ADataType, 256,
BDataType, 256,
CDataType, 128,
AElementOp, 32,
BElementOp, 8,
CElementOp>; 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>;
int main(int argc, char* argv[]) int main(int argc, char* argv[])
{ {
...@@ -94,9 +117,13 @@ int main(int argc, char* argv[]) ...@@ -94,9 +117,13 @@ int main(int argc, char* argv[])
ck::index_t StrideA = 4096; ck::index_t StrideA = 4096;
ck::index_t StrideB = 4096; ck::index_t StrideB = 4096;
ck::index_t StrideC = 4096; ck::index_t StrideE = 4096;
if(argc == 4) if(argc == 1)
{
// use default case
}
else if(argc == 4)
{ {
do_verification = std::stoi(argv[1]); do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]); init_method = std::stoi(argv[2]);
...@@ -114,14 +141,14 @@ int main(int argc, char* argv[]) ...@@ -114,14 +141,14 @@ int main(int argc, char* argv[])
StrideA = std::stoi(argv[7]); StrideA = std::stoi(argv[7]);
StrideB = std::stoi(argv[8]); StrideB = std::stoi(argv[8]);
StrideC = std::stoi(argv[9]); StrideE = std::stoi(argv[9]);
} }
else else
{ {
printf("arg1: verification (0=no, 1=yes)\n"); printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"); printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=n0, 1=yes)\n"); printf("arg3: time kernel (0=no, 1=yes)\n");
printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n"); printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideE\n");
exit(0); exit(0);
} }
...@@ -141,17 +168,14 @@ int main(int argc, char* argv[]) ...@@ -141,17 +168,14 @@ int main(int argc, char* argv[])
Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{})); Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{})); Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{})); Tensor<DDataType> d_m_n(f_host_tensor_descriptor(M, N, 0, ELayout{}));
Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{})); Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
// c0_n[n]
Tensor<CDataType> c0_n(HostTensorDescriptor(
std::vector<std::size_t>({static_cast<std::size_t>(N)}), std::vector<std::size_t>({1})));
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl; std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl; std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl; std::cout << "d_m_n: " << d_m_n.mDesc << std::endl;
std::cout << "c0_n: " << c0_n.mDesc << std::endl; std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
switch(init_method) switch(init_method)
{ {
...@@ -159,59 +183,59 @@ int main(int argc, char* argv[]) ...@@ -159,59 +183,59 @@ int main(int argc, char* argv[])
case 1: case 1:
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5}); a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}); b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
c0_n.GenerateTensorValue(GeneratorTensor_2<CDataType>{-5, 5}); d_m_n.GenerateTensorValue(GeneratorTensor_2<DDataType>{-5, 5});
break; break;
default: default:
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0}); a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}); b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
c0_n.GenerateTensorValue(GeneratorTensor_3<CDataType>{0.0, 1.0}); d_m_n.GenerateTensorValue(GeneratorTensor_3<DDataType>{0.0, 1.0});
} }
DeviceMem a_m_k_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace()); DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
DeviceMem b_k_n_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace()); DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
DeviceMem c_m_n_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpace()); DeviceMem d_device_buf(sizeof(DDataType) * d_m_n.mDesc.GetElementSpace());
DeviceMem c0_n_device_buf(sizeof(CDataType) * c0_n.mDesc.GetElementSpace()); DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpace());
a_m_k_device_buf.ToDevice(a_m_k.mData.data()); a_device_buf.ToDevice(a_m_k.mData.data());
b_k_n_device_buf.ToDevice(b_k_n.mData.data()); b_device_buf.ToDevice(b_k_n.mData.data());
c_m_n_device_buf.ToDevice(c_m_n_device_result.mData.data()); d_device_buf.ToDevice(d_m_n.mData.data());
c0_n_device_buf.ToDevice(c0_n.mData.data());
auto a_element_op = AElementOp{}; auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{}; auto b_element_op = BElementOp{};
auto c_element_op = CElementOp{}; auto cde_element_op = CDEElementOp{};
// do GEMM // do GEMM
auto gemm = DeviceGemmInstance{}; auto device_op = DeviceOpInstance{};
auto invoker = gemm.MakeInvoker(); auto invoker = device_op.MakeInvoker();
auto argument = gemm.MakeArgument(static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()), auto argument =
static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()), device_op.MakeArgument(a_device_buf.GetDeviceBuffer(),
static_cast<CDataType*>(c0_n_device_buf.GetDeviceBuffer()), b_device_buf.GetDeviceBuffer(),
M, std::array<const void*, 1>{d_device_buf.GetDeviceBuffer()},
N, e_device_buf.GetDeviceBuffer(),
K, M,
StrideA, N,
StrideB, K,
StrideC, StrideA,
a_element_op, StrideB,
b_element_op, std::array<ck::index_t, 1>{0},
c_element_op); StrideE,
a_element_op,
if(!gemm.IsSupportedArgument(argument)) b_element_op,
cde_element_op);
if(!device_op.IsSupportedArgument(argument))
{ {
throw std::runtime_error( throw std::runtime_error("wrong! this device_op instance does not support this problem");
"wrong! device_gemm with the specified compilation parameters does "
"not support this GEMM problem");
} }
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel}); float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * M * N * K; std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * M + std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
sizeof(CDataType) * M * N + sizeof(CDataType) * N; sizeof(EDataType) * M * N + sizeof(EDataType) * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time; float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
...@@ -220,19 +244,37 @@ int main(int argc, char* argv[]) ...@@ -220,19 +244,37 @@ int main(int argc, char* argv[])
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s" std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
<< std::endl; << std::endl;
c_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data());
if(do_verification) if(do_verification)
{ {
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
Tensor<AccDataType> c_m_n(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
AccDataType,
AccDataType,
AElementOp,
BElementOp,
PassThrough>;
auto ref_gemm = ReferenceGemmInstance{}; auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker(); auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument( auto ref_argument =
a_m_k, b_k_n, c_m_n_host_result, c0_n, a_element_op, b_element_op, c_element_op); ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, PassThrough{});
ref_invoker.Run(ref_argument); ref_invoker.Run(ref_argument);
return ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData) ? 0 : 1; for(int m = 0; m < M; ++m)
{
for(int n = 0; n < N; ++n)
{
cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d_m_n(m, n));
}
}
return ck::utils::check_err(e_m_n_device_result.mData, e_m_n_host_result.mData) ? 0 : 1;
} }
return 0; return 0;
......
example/04_gemm_add_add_fastgelu/CMakeLists.txt 0 → 100644
View file @ b79df771
add_example_executable(example_gemm_add_add_fastgelu_xdl_fp16 gemm_add_add_fastgelu_xdl_fp16.cpp)
example/04_gemm_bias_relu_add/README.md → example/04_gemm_add_add_fastgelu/README.md
View file @ b79df771
# Instructions for ```example_gemm_xdl_bias_relu_add``` # Instructions for ```example_gemm_add_add_fastgelu_xdl_fp16```
## Run ```example_gemm_xdl_bias_relu_add``` ## Run ```example_gemm_add_add_fastgelu_xdl_fp16```
```bash ```bash
#arg1: verification (0=no, 1=yes) #arg1: verification (0=no, 1=yes)
#arg2: initialization (0=no init, 1=integer value, 2=decimal value) #arg2: initialization (0=no init, 1=integer value, 2=decimal value)
#arg3: run kernel # of times (>1) #arg3: time kernel (0=no, 1=yes)
#arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC #arg4 to 11: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD0, StrideD1, StrideE"
./bin/example_gemm_xdl_bias_relu_add 0 1 5 3840 4096 4096 4096 4096 4096 ./bin/example_gemm_add_add_fastgelu_xdl_fp16 1 1 1
``` ```
Result (MI100 @ 1087Mhz, 133.5TFlops peak FP16) Result (MI100 @ 1087Mhz, 133.5TFlops peak FP16)
``` ```
a_m_k: dim 2, lengths {3840, 4096}, strides {4096, 1} a_m_k: dim 2, lengths {3840, 4096}, strides {4096, 1}
b_k_n: dim 2, lengths {4096, 4096}, strides {1, 4096} b_k_n: dim 2, lengths {4096, 4096}, strides {1, 4096}
c_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1} d0_m_n: dim 2, lengths {3840, 4096}, strides {0, 1}
c0_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1} d1_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
c1_m_n: dim 2, lengths {3840, 4096}, strides {1, 0} e_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
arg.a_grid_desc_k0_m_k1_{512, 3840, 8}
arg.b_grid_desc_k0_n_k1_{512, 4096, 8}
arg.c_grid_desc_m_n_{ 3840, 4096}
arg.c0_grid_desc_m_n_{ 3840, 4096}
arg.c1_grid_desc_m_n_{ 3840, 4096}
launch_and_time_kernel: grid_dim {480, 1, 1}, block_dim {256, 1, 1} launch_and_time_kernel: grid_dim {480, 1, 1}, block_dim {256, 1, 1}
Warm up Warm up 1 time
Start running 5 times... Start running 10 times...
Perf: 1.27583 ms, 100.992 TFlops, 73.9688 GB/s Perf: 1.26914 ms, 101.525 TFlops, 100.804 GB/s, DeviceGemmMultipleD_Xdl_CShuffle<256, 256, 128, 32, 8, 8>
``` ```
example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_fp16.cpp 0 → 100644
View file @ b79df771
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/host_tensor/device_memory.hpp"
#include "ck/library/host_tensor/host_tensor.hpp"
#include "ck/library/host_tensor/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "ck/library/utility/check_err.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 PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AddAddFastGelu = ck::tensor_operation::element_wise::AddAddFastGelu;
using ADataType = F16;
using BDataType = F16;
using AccDataType = F32;
using CShuffleDataType = F32;
using D0DataType = F16;
using D1DataType = F16;
using DsDataType = ck::Tuple<D0DataType, D1DataType>;
using EDataType = F16;
using ALayout = Row;
using BLayout = Col;
using D0Layout = Row;
using D1Layout = Row;
using ELayout = Row;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = AddAddFastGelu;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// clang-format off
using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleD_Xdl_CShuffle
//######| ALayout| BLayout| ELayout| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//######| | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//######| | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< ALayout, BLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementOp, BElementOp, CDEElementOp, GemmDefault, 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>;
// clang-format on
int main(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
// GEMM shape
ck::index_t M = 3840;
ck::index_t N = 4096;
ck::index_t K = 4096;
ck::index_t StrideA = 4096;
ck::index_t StrideB = 4096;
ck::index_t StrideD0 = 0;
ck::index_t StrideD1 = 4096;
ck::index_t StrideE = 4096;
if(argc == 1)
{
// use default case
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 12)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
M = std::stoi(argv[4]);
N = std::stoi(argv[5]);
K = std::stoi(argv[6]);
StrideA = std::stoi(argv[7]);
StrideB = std::stoi(argv[8]);
StrideD0 = std::stoi(argv[9]);
StrideD1 = std::stoi(argv[10]);
StrideE = std::stoi(argv[11]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=no, 1=yes)\n");
printf("arg4 to 10: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD0, StrideD1, "
"StrideE\n");
exit(0);
}
auto f_host_tensor_descriptor =
[](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}));
}
};
Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<D0DataType> d0_m_n(f_host_tensor_descriptor(M, N, StrideD0, D0Layout{}));
Tensor<D1DataType> d1_m_n(f_host_tensor_descriptor(M, N, StrideD1, D1Layout{}));
Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "d0_m_n: " << d0_m_n.mDesc << std::endl;
std::cout << "d1_m_n: " << d1_m_n.mDesc << std::endl;
std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
d0_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-5, 5});
d1_m_n.GenerateTensorValue(GeneratorTensor_2<D1DataType>{-5, 5});
break;
default:
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
d0_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{0.0, 1.0});
d1_m_n.GenerateTensorValue(GeneratorTensor_3<D1DataType>{0.0, 1.0});
}
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
DeviceMem d0_device_buf(sizeof(D0DataType) * d0_m_n.mDesc.GetElementSpace());
DeviceMem d1_device_buf(sizeof(D1DataType) * d1_m_n.mDesc.GetElementSpace());
DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpace());
a_device_buf.ToDevice(a_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
d0_device_buf.ToDevice(d0_m_n.mData.data());
d1_device_buf.ToDevice(d1_m_n.mData.data());
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto cde_element_op = CDEElementOp{};
// do GEMM
auto device_op = DeviceOpInstance{};
auto invoker = device_op.MakeInvoker();
auto argument =
device_op.MakeArgument(a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
std::array<const void*, 2>{d0_device_buf.GetDeviceBuffer(),
d1_device_buf.GetDeviceBuffer()},
e_device_buf.GetDeviceBuffer(),
M,
N,
K,
StrideA,
StrideB,
std::array<ck::index_t, 2>{StrideD0, StrideD1},
StrideE,
a_element_op,
b_element_op,
cde_element_op);
if(!device_op.IsSupportedArgument(argument))
{
throw std::runtime_error("wrong! this device_op instance does not support this problem");
}
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(D0DataType) * N + sizeof(D1DataType) * M * N +
sizeof(EDataType) * 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, "
<< device_op.GetTypeString() << std::endl;
if(do_verification)
{
Tensor<AccDataType> c_m_n(HostTensorDescriptor(
std::vector<std::size_t>{static_cast<std::size_t>(M), static_cast<std::size_t>(N)}));
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
AccDataType,
AccDataType,
AElementOp,
BElementOp,
PassThrough>;
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, PassThrough{});
ref_invoker.Run(ref_argument);
for(int m = 0; m < M; ++m)
{
for(int n = 0; n < N; ++n)
{
cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d0_m_n(m, n), d1_m_n(m, n));
}
}
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
return ck::utils::check_err(e_m_n_device_result.mData, e_m_n_host_result.mData) ? 0 : 1;
}
return 0;
}
example/04_gemm_bias_relu_add/CMakeLists.txt deleted 100644 → 0
View file @ 05d38218
add_example_executable(example_gemm_xdl_bias_relu_add gemm_xdl_bias_relu_add.cpp)
example/04_gemm_bias_relu_add/gemm_xdl_bias_relu_add.cpp deleted 100644 → 0
View file @ 05d38218
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <stdlib.h>
#include <half.hpp>
#include "check_err.hpp"
#include "config.hpp"
#include "print.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "host_gemm.hpp"
#include "device_tensor.hpp"
#include "element_wise_operation.hpp"
#include "device_gemm_xdl_c_shuffle_bias_activation_add.hpp"
#include "reference_gemm_bias_activation_add.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using ADataType = ck::half_t;
using BDataType = ck::half_t;
using CDataType = ck::half_t;
using AccDataType = float;
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::AddReluAdd;
// clang-format off
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdl_C_Shuffle_Bias_Activation_Add<
ADataType, // ADataType
BDataType, // BDataType
CDataType, // CDataType
AccDataType, // AccDataType
ALayout, // ALayout
BLayout, // BLayout
CLayout, // CLayout
AElementOp, // AElementwiseOperation
BElementOp, // BElementwiseOperation
CElementOp, // CElementwiseOperation
256, // BlockSize
256, // MPerBlock
128, // NPerBlock
4, // K0PerBlock
8, // K1
32, // MPerXDL
32, // NPerXDL
4, // MXdlPerWave
2, // NXdlPerWave
S<4, 64, 1>, // ABlockTransferThreadClusterLengths_K0_M_K1
S<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // ABlockTransferSrcAccessOrder
2, // ABlockTransferSrcVectorDim
8, // ABlockTransferSrcScalarPerVector
8, // ABlockTransferDstScalarPerVector_K1
true, // ABlockLdsAddExtraM
S<4, 64, 1>, // BBlockTransferThreadClusterLengths_K0_N_K1
S<1, 0, 2>, // BBlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // BBlockTransferSrcAccessOrder
2, // BBlockTransferSrcVectorDim
8, // BBlockTransferSrcScalarPerVector
8, // BBlockTransferDstScalarPerVector_K1
true, // BBlockLdsAddExtraN
1, // CShuffleMXdlPerWavePerShuffle
1, // CShuffleNXdlPerWavePerShuffle
S<1, 1, 32, 1, 1, 8>, // CBlockTransferClusterLengths_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl
8>; // CBlockTransferScalarPerVector_NWaveNPerXdl
// clang-format on
using ReferenceGemmInstance =
ck::tensor_operation::host::ReferenceGemmBiasActivationAdd<ADataType,
BDataType,
CDataType,
AElementOp,
BElementOp,
CElementOp>;
int main(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
// GEMM shape
ck::index_t M = 3840;
ck::index_t N = 4096;
ck::index_t K = 4096;
ck::index_t StrideA = 4096;
ck::index_t StrideB = 4096;
ck::index_t StrideC = 4096;
ck::index_t StrideC1 = 4096;
if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 11)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
M = std::stoi(argv[4]);
N = std::stoi(argv[5]);
K = std::stoi(argv[6]);
StrideA = std::stoi(argv[7]);
StrideB = std::stoi(argv[8]);
StrideC = std::stoi(argv[9]);
StrideC1 = std::stoi(argv[10]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=n0, 1=yes)\n");
printf("arg4 to 10: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC, StrideC1\n");
exit(0);
}
auto f_host_tensor_descriptor =
[](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}));
}
};
Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
// c0_n[n]
Tensor<CDataType> c0_n(HostTensorDescriptor(
std::vector<std::size_t>({static_cast<std::size_t>(N)}), std::vector<std::size_t>({1})));
// c1_m_n[m ,n]
Tensor<CDataType> c1_m_n(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl;
std::cout << "c0_n: " << c0_n.mDesc << std::endl;
std::cout << "c1_m_n: " << c1_m_n.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
c0_n.GenerateTensorValue(GeneratorTensor_2<CDataType>{-5, 5});
c1_m_n.GenerateTensorValue(GeneratorTensor_2<CDataType>{-5, 5});
break;
default:
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
c0_n.GenerateTensorValue(GeneratorTensor_3<CDataType>{0.0, 1.0});
c1_m_n.GenerateTensorValue(GeneratorTensor_3<CDataType>{0.0, 1.0});
}
DeviceMem a_m_k_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
DeviceMem b_k_n_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
DeviceMem c_m_n_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpace());
DeviceMem c0_n_device_buf(sizeof(CDataType) * c0_n.mDesc.GetElementSpace());
DeviceMem c1_m_n_device_buf(sizeof(CDataType) * c1_m_n.mDesc.GetElementSpace());
a_m_k_device_buf.ToDevice(a_m_k.mData.data());
b_k_n_device_buf.ToDevice(b_k_n.mData.data());
c_m_n_device_buf.ToDevice(c_m_n_device_result.mData.data());
c0_n_device_buf.ToDevice(c0_n.mData.data());
c1_m_n_device_buf.ToDevice(c1_m_n.mData.data());
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto c_element_op = CElementOp{};
// do GEMM
auto gemm = DeviceGemmInstance{};
auto invoker = gemm.MakeInvoker();
auto argument = gemm.MakeArgument(static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c0_n_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c1_m_n_device_buf.GetDeviceBuffer()),
M,
N,
K,
StrideA,
StrideB,
StrideC,
StrideC1,
a_element_op,
b_element_op,
c_element_op);
if(!gemm.IsSupportedArgument(argument))
{
throw std::runtime_error(
"wrong! device_gemm with the specified compilation parameters does "
"not support this GEMM problem");
}
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 * M +
sizeof(CDataType) * M * N + sizeof(CDataType) * 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"
<< std::endl;
c_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data());
if(do_verification)
{
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_host_result,
c0_n,
c1_m_n,
a_element_op,
b_element_op,
c_element_op);
ref_invoker.Run(ref_argument);
return ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData) ? 0 : 1;
}
return 0;
}
example/06_conv2d_fwd_bias_relu/conv2d_fwd_xdl_bias_relu.cpp
View file @ b79df771
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream> #include <iostream>
#include <numeric> #include <numeric>
#include <initializer_list> #include <initializer_list>
#include <cstdlib> #include <cstdlib>
#include <stdlib.h>
#include <half.hpp> #include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "check_err.hpp" #include "ck/tensor_operation/gpu/device/device_conv2d_fwd_xdl_c_shuffle_bias_activation_nhwc_kyxc_nhwk.hpp"
#include "config.hpp" #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "conv_util.hpp"
#include "device.hpp" #include "ck/library/utility/check_err.hpp"
#include "device_conv2d_fwd_xdl_c_shuffle_bias_activation_nhwc_kyxc_nhwk.hpp" #include "ck/library/utility/conv_util.hpp"
#include "device_tensor.hpp" #include "ck/library/host_tensor/device_memory.hpp"
#include "element_wise_operation.hpp" #include "ck/library/host_tensor/host_tensor.hpp"
#include "host_tensor.hpp" #include "ck/library/host_tensor/host_tensor_generator.hpp"
#include "host_tensor_generator.hpp" #include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd_bias_activation.hpp"
#include "reference_conv_fwd_bias_activation.hpp"
#include "tensor_layout.hpp"
namespace { namespace {
......
example/07_conv2d_fwd_bias_relu_add/conv2d_fwd_xdl_bias_relu_add.cpp
View file @ b79df771
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream> #include <iostream>
#include <numeric> #include <numeric>
#include <initializer_list> #include <initializer_list>
#include <cstdlib> #include <cstdlib>
#include <stdlib.h>
#include <half.hpp> #include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "check_err.hpp" #include "ck/tensor_operation/gpu/device/device_conv2d_fwd_xdl_c_shuffle_bias_activation_add_nhwc_kyxc_nhwk.hpp"
#include "config.hpp" #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "conv_util.hpp"
#include "device.hpp" #include "ck/library/utility/check_err.hpp"
#include "device_conv2d_fwd_xdl_c_shuffle_bias_activation_add_nhwc_kyxc_nhwk.hpp" #include "ck/library/utility/conv_util.hpp"
#include "device_tensor.hpp" #include "ck/library/host_tensor/device_memory.hpp"
#include "element_wise_operation.hpp" #include "ck/library/host_tensor/host_tensor.hpp"
#include "host_tensor.hpp" #include "ck/library/host_tensor/host_tensor_generator.hpp"
#include "host_tensor_generator.hpp" #include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd_bias_activation_add.hpp"
#include "reference_conv_fwd_bias_activation_add.hpp"
#include "tensor_layout.hpp"
namespace { namespace {
......
example/09_convnd_fwd/convnd_fwd_xdl_fp16.cpp
View file @ b79df771
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib> #include <cstdlib>
#include <iostream> #include <iostream>
#include <numeric> #include <numeric>
#include <type_traits> #include <type_traits>
#include "check_err.hpp" #include "ck/ck.hpp"
#include "config.hpp" #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "conv_util.hpp" #include "ck/tensor_operation/gpu/device/device_convnd_fwd_xdl_nhwc_kyxc_nhwk.hpp"
#include "device.hpp" #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "device_tensor.hpp"
#include "device_convnd_fwd_xdl_nhwc_kyxc_nhwk.hpp" #include "ck/library/utility/check_err.hpp"
#include "element_wise_operation.hpp" #include "ck/library/utility/conv_util.hpp"
#include "host_tensor.hpp" #include "ck/library/host_tensor/device_memory.hpp"
#include "host_tensor_generator.hpp" #include "ck/library/host_tensor/host_tensor.hpp"
#include "reference_conv_fwd.hpp" #include "ck/library/host_tensor/host_tensor_generator.hpp"
#include "tensor_layout.hpp" #include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
namespace { namespace {
...@@ -291,8 +294,8 @@ int main(int argc, char* argv[]) ...@@ -291,8 +294,8 @@ int main(int argc, char* argv[])
float tflops = static_cast<float>(flop) / 1.E9 / ave_time; float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / 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, " << conv->GetTypeString() std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
<< std::endl; << conv->GetTypeString() << std::endl;
if(do_verification) if(do_verification)
{ {
......
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