Skip to content

GitLab

"git@developer.sourcefind.cn:modelzoo/resnet50_tensorflow.git" did not exist on "84577d6de55fe9fdc5a8cd53916d3eccfc658735"
Unverified Commit 68b61504 authored by Po Yen Chen's avatar Po Yen Chen Committed by GitHub
Browse files

Add examples for GEMM + AddAddFastGelu (data type: int8, bf16, fp32) (#340) 

* Add always_false<> util to delay symbol resolution

* Use always_false<> to prevent trying instantiate unwanted method

* Add new specializations of AddAddFastGelu::operator() method

* Add GEMM + AddAddFastGelu examples for data types: int8, bf16, fp32

* Use floating point literal to simplify code

* Remove unnecessary capture in lambda expressions

* Extract fast GeLU calculation as standalone method

* Mark methods as 'constexpr'

* Add constraint for HostTensorDescriptor templated ctors

* Simplify HostTensorDescriptor ctor calls

* Add C++23 std::size_t literal suffix

* Use _uz suffix to shorten example code

* Remove unnecessary conversion to std::array<>

* Re-order include directives

* Remove C-style casting by literal suffix

* Remove unnecessary statements in main()

* Remove unused type parameter of always_false<>

* Remove unused include directive

* Exit main() by returning meaningful value

* Use 'if constexpr' to switch example flow

* Use std::is_same_v<> to shorten example code

* Add 'inline' specifier to literal functions

* Unify output methods in example

* Move common codes into .inc file

* Add type check in type_convert<>()

* Add type_convert<float>() before computation

* Merge AddAddFastGelu method specializations

* Remove always_false<>

* Add constraint to AddAddFastGelu::operator() parameter types
parent fdfd7eb5
Hide whitespace changes
Inline Side-by-side
Showing with 470 additions and 226 deletions
example/04_gemm_add_add_fastgelu/CMakeLists.txt
View file @ 68b61504
add_example_executable(example_gemm_add_add_fastgelu_xdl_bf16 gemm_add_add_fastgelu_xdl_bf16.cpp)
add_example_executable(example_gemm_add_add_fastgelu_xdl_fp16 gemm_add_add_fastgelu_xdl_fp16.cpp)
add_example_executable(example_gemm_add_add_fastgelu_xdl_fp32 gemm_add_add_fastgelu_xdl_fp32.cpp)
add_example_executable(example_gemm_add_add_fastgelu_xdl_int8 gemm_add_add_fastgelu_xdl_int8.cpp)
example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_bf16.cpp 0 → 100644
View file @ 68b61504
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <cstddef>
#include <iostream>
#include <stdexcept>
#include <string>
#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/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using BF16 = ck::bhalf_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 = BF16;
using BDataType = BF16;
using AccDataType = F32;
using CShuffleDataType = F32;
using D0DataType = BF16;
using D1DataType = BF16;
using DsDataType = ck::Tuple<D0DataType, D1DataType>;
using EDataType = BF16;
using ALayout = Row;
using BLayout = Col;
using D0Layout = Row;
using D1Layout = Row;
using DsLayout = ck::Tuple<D0Layout, D1Layout>;
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| DsLayout| 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, DsLayout, 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
#include "run_gemm_add_add_fastgelu_example.inc"
int main(int argc, char* argv[]) { return !run_gemm_add_add_fastgelu_example(argc, argv); }
example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_fp16.cpp
View file @ 68b61504
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <cstddef>
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <stdexcept>
#include <string>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
......@@ -12,11 +12,12 @@
#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/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/literals.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
......@@ -61,189 +62,6 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleD_Xdl_C
< ALayout, BLayout, DsLayout, 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.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
DeviceMem d0_device_buf(sizeof(D0DataType) * d0_m_n.mDesc.GetElementSpaceSize());
DeviceMem d1_device_buf(sizeof(D1DataType) * d1_m_n.mDesc.GetElementSpaceSize());
DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
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;
}
#include "run_gemm_add_add_fastgelu_example.inc"
return 0;
}
int main(int argc, char* argv[]) { return !run_gemm_add_add_fastgelu_example(argc, argv); }
example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_fp32.cpp 0 → 100644
View file @ 68b61504
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <cstddef>
#include <iostream>
#include <stdexcept>
#include <string>
#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/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.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 = F32;
using BDataType = F32;
using AccDataType = F32;
using CShuffleDataType = F32;
using D0DataType = F32;
using D1DataType = F32;
using DsDataType = ck::Tuple<D0DataType, D1DataType>;
using EDataType = F32;
using ALayout = Row;
using BLayout = Col;
using D0Layout = Row;
using D1Layout = Row;
using DsLayout = ck::Tuple<D0Layout, D1Layout>;
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| DsLayout| 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, DsLayout, 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, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 32, 1, 8>, 4>;
// clang-format on
#include "run_gemm_add_add_fastgelu_example.inc"
int main(int argc, char* argv[]) { return !run_gemm_add_add_fastgelu_example(argc, argv); }
example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_int8.cpp 0 → 100644
View file @ 68b61504
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <cstddef>
#include <iostream>
#include <stdexcept>
#include <string>
#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/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using I8 = int8_t;
using I32 = int32_t;
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 = I8;
using BDataType = I8;
using AccDataType = I32;
using CShuffleDataType = I32;
using D0DataType = I8;
using D1DataType = I8;
using DsDataType = ck::Tuple<D0DataType, D1DataType>;
using EDataType = I8;
using ALayout = Row;
using BLayout = Col;
using D0Layout = Row;
using D1Layout = Row;
using DsLayout = ck::Tuple<D0Layout, D1Layout>;
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| DsLayout| 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, DsLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementOp, BElementOp, CDEElementOp, GemmDefault, 1, 256, 256, 128, 64, 16, 16, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 4>, 16>;
// clang-format on
#include "run_gemm_add_add_fastgelu_example.inc"
int main(int argc, char* argv[]) { return !run_gemm_add_add_fastgelu_example(argc, argv); }
example/04_gemm_add_add_fastgelu/run_gemm_add_add_fastgelu_example.inc 0 → 100644
View file @ 68b61504
#pragma once
struct ProblemSize final
{
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;
};
struct ExecutionConfig final
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
};
bool run_gemm_add_add_fastgelu(const ProblemSize& problem_size, const ExecutionConfig& config)
{
auto& [M, N, K, StrideA, StrideB, StrideD0, StrideD1, StrideE] = problem_size;
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
if constexpr(std::is_same_v<decltype(layout), ck::tensor_layout::gemm::RowMajor>)
{
return HostTensorDescriptor({row, col}, {stride, 1_uz});
}
else
{
return HostTensorDescriptor({row, col}, {1_uz, 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(config.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.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
DeviceMem d0_device_buf(sizeof(D0DataType) * d0_m_n.mDesc.GetElementSpaceSize());
DeviceMem d1_device_buf(sizeof(D1DataType) * d1_m_n.mDesc.GetElementSpaceSize());
DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
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(),
{d0_device_buf.GetDeviceBuffer(), d1_device_buf.GetDeviceBuffer()},
e_device_buf.GetDeviceBuffer(),
M,
N,
K,
StrideA,
StrideB,
{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, config.time_kernel});
std::size_t flop = 2_uz * 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(config.do_verification)
{
Tensor<AccDataType> c_m_n(HostTensorDescriptor{M, 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);
}
return true;
}
bool run_gemm_add_add_fastgelu_example(int argc, char* argv[])
{
ProblemSize problem_size;
ExecutionConfig config;
if(argc == 1)
{
// use default case
}
else if(argc == 4)
{
config.do_verification = std::stoi(argv[1]);
config.init_method = std::stoi(argv[2]);
config.time_kernel = std::stoi(argv[3]);
}
else if(argc == 12)
{
config.do_verification = std::stoi(argv[1]);
config.init_method = std::stoi(argv[2]);
config.time_kernel = std::stoi(argv[3]);
problem_size.M = std::stoi(argv[4]);
problem_size.N = std::stoi(argv[5]);
problem_size.K = std::stoi(argv[6]);
problem_size.StrideA = std::stoi(argv[7]);
problem_size.StrideB = std::stoi(argv[8]);
problem_size.StrideD0 = std::stoi(argv[9]);
problem_size.StrideD1 = std::stoi(argv[10]);
problem_size.StrideE = std::stoi(argv[11]);
}
else
{
std::cerr << "arg1: verification (0=no, 1=yes)" << std::endl
<< "arg2: initialization (0=no init, 1=integer value, 2=decimal value)"
<< std::endl
<< "arg3: time kernel (0=no, 1=yes)" << std::endl
<< "arg4 to 10: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD0, StrideD1, "
"StrideE"
<< std::endl;
return true;
}
return run_gemm_add_add_fastgelu(problem_size, config);
}
include/ck/tensor_operation/gpu/element/element_wise_operation.hpp
View file @ 68b61504
......@@ -114,28 +114,33 @@ struct AddHardswishAdd
// E = FastGelu(C + D0 + D1)
struct AddAddFastGelu
{
template <typename E, typename C, typename D0, typename D1>
__host__ __device__ void operator()(E&, const C&, const D0&, const D1&) const;
// Fast GeLU
// https://paperswithcode.com/method/gelu
// y = 0.5*x*(1+tanh(sqrt(2/pi)*(x+0.044715*x^3)))
__host__ __device__ static constexpr float GetFastGeLU(float x)
{
const float u = 2.f * x * (0.035677f * x * x + 0.797885f);
const float emu = exp(-u);
const float cdf = 0.5f + 0.5f * (2.f / (1.f + emu) - 1.f);
return x * cdf;
}
template <>
__host__ __device__ void operator()<half_t, float, half_t, half_t>(half_t& e,
const float& c,
const half_t& d0,
const half_t& d1) const
template <typename T>
static inline constexpr bool is_valid_param_type_v =
std::is_same_v<T, float> || std::is_same_v<T, half_t> || std::is_same_v<T, bhalf_t> ||
std::is_same_v<T, int32_t> || std::is_same_v<T, int8_t>;
template <typename E, typename C, typename D0, typename D1>
__host__ __device__ constexpr void
operator()(E& e, const C& c, const D0& d0, const D1& d1) const
{
// Fast GeLU
// https://paperswithcode.com/method/gelu
// y = 0.5*x*(1+tanh(sqrt(2/pi)*(x+0.044715*x^3)))
const auto fast_gelu = [&](float x) {
const float u = float(2) * x * (float(0.035677) * x * x + float(0.797885));
const float emu = exp(-u);
const float cdf = float(0.5) + float(0.5) * (float(2) / (float(1) + emu) - float(1));
return x * cdf;
};
const float y = fast_gelu(c + float(d0) + float(d1));
e = type_convert<half_t>(y);
static_assert(is_valid_param_type_v<E> && is_valid_param_type_v<C> &&
is_valid_param_type_v<D0> && is_valid_param_type_v<D1>);
const float y =
GetFastGeLU(type_convert<float>(c) + type_convert<float>(d0) + type_convert<float>(d1));
e = type_convert<E>(y);
}
};
......
include/ck/utility/data_type.hpp
View file @ 68b61504
......@@ -934,6 +934,8 @@ using int8x64_t = typename vector_type<int8_t, 64>::type;
template <typename Y, typename X>
__host__ __device__ constexpr Y type_convert(X x)
{
static_assert(!std::is_reference_v<Y> && !std::is_reference_v<X>);
return static_cast<Y>(x);
}
......
library/include/ck/library/utility/host_tensor.hpp
View file @ 68b61504
......@@ -77,38 +77,36 @@ struct HostTensorDescriptor
void CalculateStrides();
template <typename X>
template <typename X, typename = std::enable_if_t<std::is_convertible_v<X, std::size_t>>>
HostTensorDescriptor(const std::initializer_list<X>& lens) : mLens(lens.begin(), lens.end())
{
this->CalculateStrides();
}
template <typename X>
HostTensorDescriptor(const std::vector<X>& lens) : mLens(lens.begin(), lens.end())
{
this->CalculateStrides();
}
template <typename Range>
template <typename Range,
typename = std::enable_if_t<
std::is_convertible_v<decltype(*std::begin(std::declval<Range>())), std::size_t>>>
HostTensorDescriptor(const Range& lens) : mLens(lens.begin(), lens.end())
{
this->CalculateStrides();
}
template <typename X, typename Y>
template <typename X,
typename Y,
typename = std::enable_if_t<std::is_convertible_v<X, std::size_t> &&
std::is_convertible_v<Y, std::size_t>>>
HostTensorDescriptor(const std::initializer_list<X>& lens,
const std::initializer_list<Y>& strides)
: mLens(lens.begin(), lens.end()), mStrides(strides.begin(), strides.end())
{
}
template <typename X, typename Y>
HostTensorDescriptor(const std::vector<X>& lens, const std::vector<Y>& strides)
: mLens(lens.begin(), lens.end()), mStrides(strides.begin(), strides.end())
{
}
template <typename Range1, typename Range2>
template <
typename Range1,
typename Range2,
typename = std::enable_if_t<
std::is_convertible_v<decltype(*std::begin(std::declval<Range1>())), std::size_t> &&
std::is_convertible_v<decltype(*std::begin(std::declval<Range2>())), std::size_t>>>
HostTensorDescriptor(const Range1& lens, const Range2& strides)
: mLens(lens.begin(), lens.end()), mStrides(strides.begin(), strides.end())
{
......
library/include/ck/library/utility/literals.hpp 0 → 100644
View file @ 68b61504
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
// [P0330] Literal Suffix for (signed) size_t (C++23)
// ref: https://wg21.link/p0330r8
inline constexpr std::size_t operator""_uz(unsigned long long size)
{
return static_cast<std::size_t>(size);
}
inline constexpr std::size_t operator""_zu(unsigned long long size)
{
return static_cast<std::size_t>(size);
}
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