mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-23 22:34:36 +00:00
Reorganize project folders (#6)
This commit is contained in:
Joseph Macaranas
18
example/22_cgemm/CMakeLists.txt
Normal file
18
example/22_cgemm/CMakeLists.txt
Normal file
@@ -0,0 +1,18 @@
|
||||
add_custom_target(example_cgemm_xdl)
|
||||
|
||||
add_example_executable(example_cgemm_xdl_bf16 cgemm_xdl_bf16.cpp)
|
||||
add_example_dependencies(example_cgemm_xdl example_cgemm_xdl_bf16)
|
||||
|
||||
add_example_executable(example_cgemm_xdl_fp16 cgemm_xdl_fp16.cpp)
|
||||
add_example_dependencies(example_cgemm_xdl example_cgemm_xdl_fp16)
|
||||
|
||||
add_example_executable(example_cgemm_xdl_fp32 cgemm_xdl_fp32.cpp)
|
||||
add_example_dependencies(example_cgemm_xdl example_cgemm_xdl_fp32)
|
||||
|
||||
add_example_executable(example_cgemm_xdl_int8 cgemm_xdl_int8.cpp)
|
||||
add_example_dependencies(example_cgemm_xdl example_cgemm_xdl_int8)
|
||||
|
||||
if(USE_BITINT_EXTENSION_INT4)
|
||||
add_example_executable(example_cgemm_xdl_int4 cgemm_xdl_int4.cpp)
|
||||
add_example_dependencies(example_cgemm_xdl example_cgemm_xdl_int4)
|
||||
endif()
|
||||
132
example/22_cgemm/cgemm_xdl_bf16.cpp
Normal file
132
example/22_cgemm/cgemm_xdl_bf16.cpp
Normal file
@@ -0,0 +1,132 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#include <iostream>
|
||||
|
||||
#include "cgemm_xdl_common.hpp"
|
||||
|
||||
#include "ck/library/reference_tensor_operation/cpu/reference_cgemm.hpp"
|
||||
|
||||
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/impl/device_cgemm_4gemm_xdl_cshuffle.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
|
||||
|
||||
using ADataType = BF16;
|
||||
using BDataType = BF16;
|
||||
using CDataType = BF16;
|
||||
using AccDataType = F32;
|
||||
|
||||
using ALayout = ck::tensor_layout::gemm::RowMajor;
|
||||
using BLayout = ck::tensor_layout::gemm::ColumnMajor;
|
||||
using CLayout = ck::tensor_layout::gemm::RowMajor;
|
||||
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
|
||||
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
|
||||
|
||||
using ReferenceCGemmInstance = ck::tensor_operation::host::
|
||||
ReferenceCGemm<ADataType, BDataType, CDataType, PassThrough, PassThrough, PassThrough>;
|
||||
|
||||
// clang-format off
|
||||
using DeviceCGemmInstance = ck::tensor_operation::device::DeviceCGemm_4Gemm_Xdl_CShuffle
|
||||
<ALayout, // typename ALayout
|
||||
BLayout, // typename BLayout
|
||||
CLayout, // typename CLayout
|
||||
ADataType, // typename ADataType
|
||||
BDataType, // typename BDataType
|
||||
CDataType, // typename CDataType
|
||||
AccDataType, // typename GemmAccDataType
|
||||
CDataType, // typename CShuffleDataType
|
||||
PassThrough, // typename AElementwiseOperation
|
||||
PassThrough, // typename BElementwiseOperation
|
||||
PassThrough, // typename CElementwiseOperation
|
||||
GemmDefault, // GemmSpecialization GemmSpec
|
||||
1, // index_t NumGemmKPrefetchStage
|
||||
256, // index_t BlockSize
|
||||
256, // index_t MPerBlock
|
||||
128, // index_t NPerBlock
|
||||
32, // index_t KPerBlock
|
||||
8, // index_t AK1
|
||||
8, // index_t BK1
|
||||
32, // index_t MPerXDL
|
||||
32, // index_t NPerXDL
|
||||
4, // index_t MXdlPerWave
|
||||
2, // index_t NXdlPerWave
|
||||
S<4, 64, 1>, // typename ABlockTransferThreadClusterLengths_AK0_M_AK1
|
||||
S<1, 0, 2>, // typename ABlockTransferThreadClusterArrangeOrder
|
||||
S<1, 0, 2>, // typename ABlockTransferSrcAccessOrder
|
||||
2, // index_t ABlockTransferSrcVectorDim
|
||||
8, // index_t ABlockTransferSrcScalarPerVector
|
||||
8, // index_t ABlockTransferDstScalarPerVector_AK1
|
||||
1, // index_t ABlockLdsExtraM
|
||||
S<4, 64, 1>, // typename BBlockTransferThreadClusterLengths_BK0_N_BK1
|
||||
S<1, 0, 2>, // typename BBlockTransferThreadClusterArrangeOrder
|
||||
S<1, 0, 2>, // typename BBlockTransferSrcAccessOrder
|
||||
2, // index_t BBlockTransferSrcVectorDim
|
||||
8, // index_t BBlockTransferSrcScalarPerVector
|
||||
8, // index_t BBlockTransferDstScalarPerVector_BK1
|
||||
1, // index_t BBlockLdsExtraN
|
||||
1, // index_t CShuffleMXdlPerWavePerShuffle
|
||||
1, // index_t CShuffleNXdlPerWavePerShuffle
|
||||
S<1, 32, 1, 8>, // typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
|
||||
8>; // index_t CShuffleBlockTransferScalarPerVector_NPerBlock
|
||||
// clang-format on
|
||||
|
||||
int main(int argc, char* argv[])
|
||||
{
|
||||
bool do_verification = true;
|
||||
int init_method = 1;
|
||||
bool time_kernel = false;
|
||||
|
||||
// CGEMM shape
|
||||
ck::index_t M = 3840;
|
||||
ck::index_t N = 4096;
|
||||
ck::index_t K = 416;
|
||||
|
||||
ck::index_t StrideA = 4096;
|
||||
ck::index_t StrideB = 4096;
|
||||
ck::index_t StrideC = 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 == 10)
|
||||
{
|
||||
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]);
|
||||
}
|
||||
else
|
||||
{
|
||||
std::cout << "arg1: verification (0=no, 1=yes)\n"
|
||||
<< "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
|
||||
<< "arg3: run kernel # of times (>1)\n"
|
||||
<< "arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n"
|
||||
<< std::endl;
|
||||
exit(0);
|
||||
}
|
||||
|
||||
return !run_cgemm_xdl<ADataType,
|
||||
BDataType,
|
||||
CDataType,
|
||||
ALayout,
|
||||
BLayout,
|
||||
CLayout,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
DeviceCGemmInstance,
|
||||
ReferenceCGemmInstance>(
|
||||
M, N, K, StrideA, StrideB, StrideC, do_verification, init_method, time_kernel);
|
||||
}
|
||||
254
example/22_cgemm/cgemm_xdl_common.hpp
Normal file
254
example/22_cgemm/cgemm_xdl_common.hpp
Normal file
@@ -0,0 +1,254 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#include <numeric>
|
||||
#include <initializer_list>
|
||||
#include <cstdlib>
|
||||
|
||||
#include "ck/ck.hpp"
|
||||
#include "ck/stream_config.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"
|
||||
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
|
||||
|
||||
template <ck::index_t... Is>
|
||||
using S = ck::Sequence<Is...>;
|
||||
|
||||
using F16 = ck::half_t;
|
||||
using F32 = float;
|
||||
using BF16 = ck::bhalf_t;
|
||||
using INT8 = std::int8_t;
|
||||
using INT32 = std::int32_t;
|
||||
#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
|
||||
using INT4 = ck::int4_t;
|
||||
#endif
|
||||
|
||||
template <typename ADataType,
|
||||
typename BDataType,
|
||||
typename CDataType,
|
||||
typename ALayout,
|
||||
typename BLayout,
|
||||
typename CLayout,
|
||||
typename AElementwiseOperation,
|
||||
typename BElementwiseOperation,
|
||||
typename CElementwiseOperation,
|
||||
typename DeviceCGemmInstance,
|
||||
typename ReferenceCGemmInstance,
|
||||
typename KernelADataType = ADataType,
|
||||
typename KernelBDataType = BDataType,
|
||||
typename KernelCDataType = CDataType>
|
||||
bool run_cgemm_xdl(ck::index_t M,
|
||||
ck::index_t N,
|
||||
ck::index_t K,
|
||||
ck::index_t StrideA,
|
||||
ck::index_t StrideB,
|
||||
ck::index_t StrideC,
|
||||
bool do_verification,
|
||||
int init_method,
|
||||
bool time_kernel)
|
||||
{
|
||||
#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
|
||||
static_assert(sizeof(ck::int4_t) == sizeof(int8_t),
|
||||
"sizeof ck::int4_t and int8_t is different!");
|
||||
static_assert(sizeof(ADataType) == sizeof(KernelADataType),
|
||||
"sizeof ADataType and KernelADataType is different!");
|
||||
static_assert(sizeof(BDataType) == sizeof(KernelBDataType),
|
||||
"sizeof BDataType and KernelBDataType is different!");
|
||||
static_assert(sizeof(CDataType) == sizeof(KernelCDataType),
|
||||
"sizeof CDataType and KernelCDataType is different!");
|
||||
#endif
|
||||
|
||||
auto f_host_tensor_descriptor =
|
||||
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
|
||||
using namespace ck::literals;
|
||||
|
||||
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
|
||||
{
|
||||
return HostTensorDescriptor({row, col}, {stride, 1_uz});
|
||||
}
|
||||
else
|
||||
{
|
||||
return HostTensorDescriptor({row, col}, {1_uz, stride});
|
||||
}
|
||||
};
|
||||
|
||||
Tensor<ADataType> a_m_k_real(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
|
||||
Tensor<ADataType> a_m_k_imag(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
|
||||
Tensor<BDataType> b_k_n_real(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
|
||||
Tensor<BDataType> b_k_n_imag(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
|
||||
Tensor<KernelCDataType> c_m_n_real_device_result(
|
||||
f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
|
||||
Tensor<KernelCDataType> c_m_n_imag_device_result(
|
||||
f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
|
||||
|
||||
std::cout << "a_m_k_real: " << a_m_k_real.mDesc << std::endl;
|
||||
std::cout << "a_m_k_imag: " << a_m_k_imag.mDesc << std::endl;
|
||||
std::cout << "b_k_n_real: " << b_k_n_real.mDesc << std::endl;
|
||||
std::cout << "b_k_n_imag: " << b_k_n_imag.mDesc << std::endl;
|
||||
std::cout << "c_m_n_real: " << c_m_n_real_device_result.mDesc << std::endl;
|
||||
std::cout << "c_m_n_imag: " << c_m_n_imag_device_result.mDesc << std::endl;
|
||||
|
||||
switch(init_method)
|
||||
{
|
||||
case 0: break;
|
||||
case 1:
|
||||
a_m_k_real.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
|
||||
a_m_k_imag.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
|
||||
b_k_n_real.GenerateTensorValue(GeneratorTensor_2<BDataType>{-2, 2});
|
||||
b_k_n_imag.GenerateTensorValue(GeneratorTensor_2<BDataType>{-2, 2});
|
||||
break;
|
||||
default:
|
||||
a_m_k_real.GenerateTensorValue(GeneratorTensor_3<ADataType>{-0.5, 0.5});
|
||||
a_m_k_imag.GenerateTensorValue(GeneratorTensor_3<ADataType>{-0.5, 0.5});
|
||||
b_k_n_real.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
|
||||
b_k_n_imag.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
|
||||
}
|
||||
|
||||
auto cgemm = DeviceCGemmInstance{};
|
||||
|
||||
DeviceMem a_m_k_real_device_buf(sizeof(KernelADataType) *
|
||||
a_m_k_real.mDesc.GetElementSpaceSize());
|
||||
DeviceMem a_m_k_imag_device_buf(sizeof(KernelADataType) *
|
||||
a_m_k_imag.mDesc.GetElementSpaceSize());
|
||||
DeviceMem b_k_n_real_device_buf(sizeof(KernelBDataType) *
|
||||
b_k_n_real.mDesc.GetElementSpaceSize());
|
||||
DeviceMem b_k_n_imag_device_buf(sizeof(KernelBDataType) *
|
||||
b_k_n_imag.mDesc.GetElementSpaceSize());
|
||||
DeviceMem c_m_n_real_device_buf(sizeof(KernelCDataType) *
|
||||
c_m_n_real_device_result.mDesc.GetElementSpaceSize());
|
||||
DeviceMem c_m_n_imag_device_buf(sizeof(KernelCDataType) *
|
||||
c_m_n_imag_device_result.mDesc.GetElementSpaceSize());
|
||||
DeviceMem workspace_device_buf(cgemm.GetWorkspaceSize(M, N, K, StrideA, StrideB, StrideC));
|
||||
|
||||
#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
|
||||
if constexpr(std::is_same_v<ADataType, ck::int4_t>)
|
||||
{
|
||||
Tensor<KernelADataType> a_m_k_real_converted(a_m_k_real);
|
||||
Tensor<KernelADataType> a_m_k_imag_converted(a_m_k_imag);
|
||||
Tensor<KernelBDataType> b_k_n_real_converted(b_k_n_real);
|
||||
Tensor<KernelBDataType> b_k_n_imag_converted(b_k_n_imag);
|
||||
|
||||
a_m_k_real_device_buf.ToDevice(a_m_k_real_converted.mData.data());
|
||||
a_m_k_imag_device_buf.ToDevice(a_m_k_imag_converted.mData.data());
|
||||
b_k_n_real_device_buf.ToDevice(b_k_n_real_converted.mData.data());
|
||||
b_k_n_imag_device_buf.ToDevice(b_k_n_imag_converted.mData.data());
|
||||
}
|
||||
else
|
||||
#endif // CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
|
||||
{
|
||||
a_m_k_real_device_buf.ToDevice(a_m_k_real.mData.data());
|
||||
a_m_k_imag_device_buf.ToDevice(a_m_k_imag.mData.data());
|
||||
b_k_n_real_device_buf.ToDevice(b_k_n_real.mData.data());
|
||||
b_k_n_imag_device_buf.ToDevice(b_k_n_imag.mData.data());
|
||||
}
|
||||
|
||||
auto a_element_op = AElementwiseOperation{};
|
||||
auto b_element_op = BElementwiseOperation{};
|
||||
auto c_element_op = CElementwiseOperation{};
|
||||
|
||||
// do GEMM
|
||||
auto invoker = cgemm.MakeInvoker();
|
||||
auto argument =
|
||||
cgemm.MakeArgument(static_cast<KernelADataType*>(a_m_k_real_device_buf.GetDeviceBuffer()),
|
||||
static_cast<KernelADataType*>(a_m_k_imag_device_buf.GetDeviceBuffer()),
|
||||
static_cast<KernelBDataType*>(b_k_n_real_device_buf.GetDeviceBuffer()),
|
||||
static_cast<KernelBDataType*>(b_k_n_imag_device_buf.GetDeviceBuffer()),
|
||||
static_cast<KernelCDataType*>(c_m_n_real_device_buf.GetDeviceBuffer()),
|
||||
static_cast<KernelCDataType*>(c_m_n_imag_device_buf.GetDeviceBuffer()),
|
||||
static_cast<KernelCDataType*>(workspace_device_buf.GetDeviceBuffer()),
|
||||
M,
|
||||
N,
|
||||
K,
|
||||
StrideA,
|
||||
StrideB,
|
||||
StrideC,
|
||||
a_element_op,
|
||||
b_element_op,
|
||||
c_element_op);
|
||||
|
||||
if(!cgemm.IsSupportedArgument(argument))
|
||||
{
|
||||
throw std::runtime_error(
|
||||
"wrong! device_cgemm with the specified compilation parameters does "
|
||||
"not support this CGEMM problem");
|
||||
}
|
||||
|
||||
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
|
||||
|
||||
std::size_t flop = std::size_t(8) * M * N * K;
|
||||
std::size_t num_btype =
|
||||
std::size_t(2) *
|
||||
(sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(CDataType) * M * N);
|
||||
|
||||
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
|
||||
float gb_per_sec = num_btype / 1.E6 / ave_time;
|
||||
|
||||
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
|
||||
<< cgemm.GetTypeString() << std::endl;
|
||||
|
||||
if(do_verification)
|
||||
{
|
||||
Tensor<CDataType> c_m_n_real_host_result(
|
||||
f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
|
||||
Tensor<CDataType> c_m_n_imag_host_result(
|
||||
f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
|
||||
|
||||
auto ref_cgemm = ReferenceCGemmInstance{};
|
||||
auto ref_invoker = ref_cgemm.MakeInvoker();
|
||||
auto ref_argument = ref_cgemm.MakeArgument(a_m_k_real,
|
||||
a_m_k_imag,
|
||||
b_k_n_real,
|
||||
b_k_n_imag,
|
||||
c_m_n_real_host_result,
|
||||
c_m_n_imag_host_result,
|
||||
a_element_op,
|
||||
b_element_op,
|
||||
c_element_op);
|
||||
|
||||
ref_invoker.Run(ref_argument);
|
||||
|
||||
c_m_n_real_device_buf.FromDevice(c_m_n_real_device_result.mData.data());
|
||||
c_m_n_imag_device_buf.FromDevice(c_m_n_imag_device_result.mData.data());
|
||||
|
||||
bool result = true;
|
||||
#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
|
||||
if constexpr(std::is_same_v<ADataType, ck::int4_t>)
|
||||
{
|
||||
const Tensor<CDataType> c_m_n_real_device_result_converted(c_m_n_real_device_result);
|
||||
const Tensor<CDataType> c_m_n_imag_device_result_converted(c_m_n_imag_device_result);
|
||||
|
||||
result = ck::utils::check_err(c_m_n_real_device_result_converted,
|
||||
c_m_n_real_host_result,
|
||||
"Verification error: incorrect results in real part!",
|
||||
1e-2f,
|
||||
1e-1f);
|
||||
result = result && ck::utils::check_err(
|
||||
c_m_n_imag_device_result_converted,
|
||||
c_m_n_imag_host_result,
|
||||
"Verification error: incorrect results in imaginary part!",
|
||||
1e-2f,
|
||||
1e-1f);
|
||||
}
|
||||
else
|
||||
#endif // CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
|
||||
{
|
||||
result = ck::utils::check_err(c_m_n_real_device_result,
|
||||
c_m_n_real_host_result,
|
||||
"Verification error: incorrect results in real part!",
|
||||
1e-2f,
|
||||
1e-1f);
|
||||
result = result && ck::utils::check_err(
|
||||
c_m_n_imag_device_result,
|
||||
c_m_n_imag_host_result,
|
||||
"Verification error: incorrect results in imaginary part!",
|
||||
1e-2f,
|
||||
1e-1f);
|
||||
}
|
||||
|
||||
return result;
|
||||
}
|
||||
return true;
|
||||
}
|
||||
131
example/22_cgemm/cgemm_xdl_fp16.cpp
Normal file
131
example/22_cgemm/cgemm_xdl_fp16.cpp
Normal file
@@ -0,0 +1,131 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#include <iostream>
|
||||
|
||||
#include "cgemm_xdl_common.hpp"
|
||||
#include "ck/library/reference_tensor_operation/cpu/reference_cgemm.hpp"
|
||||
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/impl/device_cgemm_4gemm_xdl_cshuffle.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
|
||||
|
||||
using ADataType = F16;
|
||||
using BDataType = F16;
|
||||
using CDataType = F16;
|
||||
using AccDataType = F32;
|
||||
using CShuffleDataType = F32;
|
||||
|
||||
using ALayout = ck::tensor_layout::gemm::RowMajor;
|
||||
using BLayout = ck::tensor_layout::gemm::ColumnMajor;
|
||||
using CLayout = ck::tensor_layout::gemm::RowMajor;
|
||||
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
|
||||
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
|
||||
|
||||
using ReferenceCGemmInstance = ck::tensor_operation::host::
|
||||
ReferenceCGemm<ADataType, BDataType, CDataType, PassThrough, PassThrough, PassThrough>;
|
||||
|
||||
// clang-format off
|
||||
using DeviceCGemmInstance = ck::tensor_operation::device::DeviceCGemm_4Gemm_Xdl_CShuffle
|
||||
<ALayout, // typename ALayout
|
||||
BLayout, // typename BLayout
|
||||
CLayout, // typename CLayout
|
||||
ADataType, // typename ADataType
|
||||
BDataType, // typename BDataType
|
||||
CDataType, // typename CDataType
|
||||
AccDataType, // typename GemmAccDataType
|
||||
CShuffleDataType, // typename CShuffleDataType
|
||||
PassThrough, // typename AElementwiseOperation
|
||||
PassThrough, // typename BElementwiseOperation
|
||||
PassThrough, // typename CElementwiseOperation
|
||||
GemmDefault, // GemmSpecialization GemmSpec
|
||||
1, // index_t NumGemmKPrefetchStage
|
||||
256, // index_t BlockSize
|
||||
256, // index_t MPerBlock
|
||||
128, // index_t NPerBlock
|
||||
32, // index_t KPerBlock
|
||||
8, // index_t AK1
|
||||
8, // index_t BK1
|
||||
32, // index_t MPerXDL
|
||||
32, // index_t NPerXDL
|
||||
4, // index_t MXdlPerWave
|
||||
2, // index_t NXdlPerWave
|
||||
S<4, 64, 1>, // typename ABlockTransferThreadClusterLengths_AK0_M_AK1
|
||||
S<1, 0, 2>, // typename ABlockTransferThreadClusterArrangeOrder
|
||||
S<1, 0, 2>, // typename ABlockTransferSrcAccessOrder
|
||||
2, // index_t ABlockTransferSrcVectorDim
|
||||
8, // index_t ABlockTransferSrcScalarPerVector
|
||||
8, // index_t ABlockTransferDstScalarPerVector_AK1
|
||||
1, // index_t ABlockLdsExtraM
|
||||
S<4, 64, 1>, // typename BBlockTransferThreadClusterLengths_BK0_N_BK1
|
||||
S<1, 0, 2>, // typename BBlockTransferThreadClusterArrangeOrder
|
||||
S<1, 0, 2>, // typename BBlockTransferSrcAccessOrder
|
||||
2, // index_t BBlockTransferSrcVectorDim
|
||||
8, // index_t BBlockTransferSrcScalarPerVector
|
||||
8, // index_t BBlockTransferDstScalarPerVector_BK1
|
||||
1, // index_t BBlockLdsExtraN
|
||||
1, // index_t CShuffleMXdlPerWavePerShuffle
|
||||
1, // index_t CShuffleNXdlPerWavePerShuffle
|
||||
S<1, 32, 1, 8>, // typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
|
||||
8>; // index_t CShuffleBlockTransferScalarPerVector_NPerBlock
|
||||
// clang-format on
|
||||
|
||||
int main(int argc, char* argv[])
|
||||
{
|
||||
bool do_verification = true;
|
||||
int init_method = 1;
|
||||
bool time_kernel = false;
|
||||
|
||||
// CGEMM 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;
|
||||
|
||||
if(argc == 4)
|
||||
{
|
||||
do_verification = std::stoi(argv[1]);
|
||||
init_method = std::stoi(argv[2]);
|
||||
time_kernel = std::stoi(argv[3]);
|
||||
}
|
||||
else if(argc == 10)
|
||||
{
|
||||
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]);
|
||||
}
|
||||
else
|
||||
{
|
||||
std::cout << "arg1: verification (0=no, 1=yes)\n"
|
||||
<< "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
|
||||
<< "arg3: run kernel # of times (>1)\n"
|
||||
<< "arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n"
|
||||
<< std::endl;
|
||||
exit(0);
|
||||
}
|
||||
|
||||
return !run_cgemm_xdl<ADataType,
|
||||
BDataType,
|
||||
CDataType,
|
||||
ALayout,
|
||||
BLayout,
|
||||
CLayout,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
DeviceCGemmInstance,
|
||||
ReferenceCGemmInstance>(
|
||||
M, N, K, StrideA, StrideB, StrideC, do_verification, init_method, time_kernel);
|
||||
}
|
||||
132
example/22_cgemm/cgemm_xdl_fp32.cpp
Normal file
132
example/22_cgemm/cgemm_xdl_fp32.cpp
Normal file
@@ -0,0 +1,132 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#include <iostream>
|
||||
|
||||
#include "cgemm_xdl_common.hpp"
|
||||
|
||||
#include "ck/library/reference_tensor_operation/cpu/reference_cgemm.hpp"
|
||||
|
||||
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/impl/device_cgemm_4gemm_xdl_cshuffle.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
|
||||
|
||||
using ADataType = F32;
|
||||
using BDataType = F32;
|
||||
using CDataType = F32;
|
||||
using AccDataType = F32;
|
||||
|
||||
using ALayout = ck::tensor_layout::gemm::RowMajor;
|
||||
using BLayout = ck::tensor_layout::gemm::ColumnMajor;
|
||||
using CLayout = ck::tensor_layout::gemm::RowMajor;
|
||||
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
|
||||
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
|
||||
|
||||
using ReferenceCGemmInstance = ck::tensor_operation::host::
|
||||
ReferenceCGemm<ADataType, BDataType, CDataType, PassThrough, PassThrough, PassThrough>;
|
||||
|
||||
// clang-format off
|
||||
using DeviceCGemmInstance = ck::tensor_operation::device::DeviceCGemm_4Gemm_Xdl_CShuffle
|
||||
<ALayout, // typename ALayout
|
||||
BLayout, // typename BLayout
|
||||
CLayout, // typename CLayout
|
||||
ADataType, // typename ADataType
|
||||
BDataType, // typename BDataType
|
||||
CDataType, // typename CDataType
|
||||
AccDataType, // typename GemmAccDataType
|
||||
CDataType, // typename CShuffleDataType
|
||||
PassThrough, // typename AElementwiseOperation
|
||||
PassThrough, // typename BElementwiseOperation
|
||||
PassThrough, // typename CElementwiseOperation
|
||||
GemmDefault, // GemmSpecialization GemmSpec
|
||||
1, // index_t NumGemmKPrefetchStage
|
||||
256, // index_t BlockSize
|
||||
256, // index_t MPerBlock
|
||||
128, // index_t NPerBlock
|
||||
16, // index_t KPerBlock
|
||||
4, // index_t AK1
|
||||
4, // index_t BK1
|
||||
32, // index_t MPerXDL
|
||||
32, // index_t NPerXDL
|
||||
4, // index_t MXdlPerWave
|
||||
2, // index_t NXdlPerWave
|
||||
S<4, 64, 1>, // typename ABlockTransferThreadClusterLengths_AK0_M_AK1
|
||||
S<1, 0, 2>, // typename ABlockTransferThreadClusterArrangeOrder
|
||||
S<1, 0, 2>, // typename ABlockTransferSrcAccessOrder
|
||||
2, // index_t ABlockTransferSrcVectorDim
|
||||
4, // index_t ABlockTransferSrcScalarPerVector
|
||||
4, // index_t ABlockTransferDstScalarPerVector_AK1
|
||||
1, // index_t ABlockLdsExtraM
|
||||
S<4, 64, 1>, // typename BBlockTransferThreadClusterLengths_BK0_N_BK1
|
||||
S<1, 0, 2>, // typename BBlockTransferThreadClusterArrangeOrder
|
||||
S<1, 0, 2>, // typename BBlockTransferSrcAccessOrder
|
||||
2, // index_t BBlockTransferSrcVectorDim
|
||||
4, // index_t BBlockTransferSrcScalarPerVector
|
||||
4, // index_t BBlockTransferDstScalarPerVector_BK1
|
||||
1, // index_t BBlockLdsExtraN
|
||||
1, // index_t CShuffleMXdlPerWavePerShuffle
|
||||
1, // index_t CShuffleNXdlPerWavePerShuffle
|
||||
S<1, 16, 1, 16>, // typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
|
||||
4>; // index_t CShuffleBlockTransferScalarPerVector_NPerBlock
|
||||
// clang-format on
|
||||
|
||||
int main(int argc, char* argv[])
|
||||
{
|
||||
bool do_verification = true;
|
||||
int init_method = 1;
|
||||
bool time_kernel = false;
|
||||
|
||||
// CGEMM 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;
|
||||
|
||||
if(argc == 4)
|
||||
{
|
||||
do_verification = std::stoi(argv[1]);
|
||||
init_method = std::stoi(argv[2]);
|
||||
time_kernel = std::stoi(argv[3]);
|
||||
}
|
||||
else if(argc == 10)
|
||||
{
|
||||
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]);
|
||||
}
|
||||
else
|
||||
{
|
||||
std::cout << "arg1: verification (0=no, 1=yes)\n"
|
||||
<< "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
|
||||
<< "arg3: run kernel # of times (>1)\n"
|
||||
<< "arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n"
|
||||
<< std::endl;
|
||||
exit(0);
|
||||
}
|
||||
|
||||
return !run_cgemm_xdl<ADataType,
|
||||
BDataType,
|
||||
CDataType,
|
||||
ALayout,
|
||||
BLayout,
|
||||
CLayout,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
DeviceCGemmInstance,
|
||||
ReferenceCGemmInstance>(
|
||||
M, N, K, StrideA, StrideB, StrideC, do_verification, init_method, time_kernel);
|
||||
}
|
||||
140
example/22_cgemm/cgemm_xdl_int4.cpp
Normal file
140
example/22_cgemm/cgemm_xdl_int4.cpp
Normal file
@@ -0,0 +1,140 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#include <iostream>
|
||||
|
||||
#include "cgemm_xdl_common.hpp"
|
||||
|
||||
#include "ck/library/reference_tensor_operation/cpu/reference_cgemm.hpp"
|
||||
|
||||
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/impl/device_cgemm_4gemm_xdl_cshuffle.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
|
||||
|
||||
using ADataType = INT4;
|
||||
using BDataType = INT4;
|
||||
using CDataType = INT4;
|
||||
using AccDataType = INT32;
|
||||
using CShuffleDataType = INT32;
|
||||
|
||||
using KernelADataType = INT8;
|
||||
using KernelBDataType = INT8;
|
||||
using KernelCDataType = INT8;
|
||||
|
||||
using ALayout = ck::tensor_layout::gemm::RowMajor;
|
||||
using BLayout = ck::tensor_layout::gemm::ColumnMajor;
|
||||
using CLayout = ck::tensor_layout::gemm::RowMajor;
|
||||
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
|
||||
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
|
||||
|
||||
using ReferenceCGemmInstance = ck::tensor_operation::host::
|
||||
ReferenceCGemm<ADataType, BDataType, CDataType, PassThrough, PassThrough, PassThrough>;
|
||||
|
||||
// clang-format off
|
||||
using DeviceCGemmInstance = ck::tensor_operation::device::DeviceCGemm_4Gemm_Xdl_CShuffle
|
||||
<ALayout, // typename ALayout
|
||||
BLayout, // typename BLayout
|
||||
CLayout, // typename CLayout
|
||||
KernelADataType, // typename ADataType
|
||||
KernelBDataType, // typename BDataType
|
||||
KernelCDataType, // typename CDataType
|
||||
AccDataType, // typename GemmAccDataType
|
||||
CShuffleDataType, // typename CShuffleDataType
|
||||
PassThrough, // typename AElementwiseOperation
|
||||
PassThrough, // typename BElementwiseOperation
|
||||
PassThrough, // typename CElementwiseOperation
|
||||
GemmDefault, // GemmSpecialization GemmSpec
|
||||
1, // index_t NumGemmKPrefetchStage
|
||||
256, // index_t BlockSize
|
||||
256, // index_t MPerBlock
|
||||
128, // index_t NPerBlock
|
||||
64, // index_t KPerBlock
|
||||
16, // index_t AK1
|
||||
16, // index_t BK1
|
||||
32, // index_t MPerXDL
|
||||
32, // index_t NPerXDL
|
||||
4, // index_t MXdlPerWave
|
||||
2, // index_t NXdlPerWave
|
||||
S<4, 64, 1>, // typename ABlockTransferThreadClusterLengths_AK0_M_AK1
|
||||
S<1, 0, 2>, // typename ABlockTransferThreadClusterArrangeOrder
|
||||
S<1, 0, 2>, // typename ABlockTransferSrcAccessOrder
|
||||
2, // index_t ABlockTransferSrcVectorDim
|
||||
16, // index_t ABlockTransferSrcScalarPerVector
|
||||
16, // index_t ABlockTransferDstScalarPerVector_AK1
|
||||
1, // index_t ABlockLdsExtraM
|
||||
S<4, 64, 1>, // typename BBlockTransferThreadClusterLengths_BK0_N_BK1
|
||||
S<1, 0, 2>, // typename BBlockTransferThreadClusterArrangeOrder
|
||||
S<1, 0, 2>, // typename BBlockTransferSrcAccessOrder
|
||||
2, // index_t BBlockTransferSrcVectorDim
|
||||
8, // index_t BBlockTransferSrcScalarPerVector
|
||||
8, // index_t BBlockTransferDstScalarPerVector_BK1
|
||||
1, // index_t BBlockLdsExtraN
|
||||
1, // index_t CShuffleMXdlPerWavePerShuffle
|
||||
1, // index_t CShuffleNXdlPerWavePerShuffle
|
||||
S<1, 64, 1, 4>, // typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
|
||||
16>; // index_t CShuffleBlockTransferScalarPerVector_NPerBlock
|
||||
// clang-format on
|
||||
|
||||
int main(int argc, char* argv[])
|
||||
{
|
||||
bool do_verification = true;
|
||||
int init_method = 1;
|
||||
bool time_kernel = true;
|
||||
|
||||
// CGEMM shape
|
||||
ck::index_t M = 1024;
|
||||
ck::index_t N = 1152;
|
||||
ck::index_t K = 512;
|
||||
|
||||
ck::index_t StrideA = K;
|
||||
ck::index_t StrideB = K;
|
||||
ck::index_t StrideC = N;
|
||||
|
||||
if(argc == 4)
|
||||
{
|
||||
do_verification = std::stoi(argv[1]);
|
||||
init_method = std::stoi(argv[2]);
|
||||
time_kernel = std::stoi(argv[3]);
|
||||
}
|
||||
else if(argc == 10)
|
||||
{
|
||||
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]);
|
||||
}
|
||||
else
|
||||
{
|
||||
std::cout << "arg1: verification (0=no, 1=yes)\n"
|
||||
<< "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
|
||||
<< "arg3: time kernel (0=no, 1=yes)\n"
|
||||
<< "arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n"
|
||||
<< std::endl;
|
||||
exit(EXIT_SUCCESS);
|
||||
}
|
||||
|
||||
return !run_cgemm_xdl<ADataType,
|
||||
BDataType,
|
||||
CDataType,
|
||||
ALayout,
|
||||
BLayout,
|
||||
CLayout,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
DeviceCGemmInstance,
|
||||
ReferenceCGemmInstance,
|
||||
KernelADataType,
|
||||
KernelBDataType,
|
||||
KernelCDataType>(
|
||||
M, N, K, StrideA, StrideB, StrideC, do_verification, init_method, time_kernel);
|
||||
}
|
||||
132
example/22_cgemm/cgemm_xdl_int8.cpp
Normal file
132
example/22_cgemm/cgemm_xdl_int8.cpp
Normal file
@@ -0,0 +1,132 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#include <iostream>
|
||||
|
||||
#include "cgemm_xdl_common.hpp"
|
||||
|
||||
#include "ck/library/reference_tensor_operation/cpu/reference_cgemm.hpp"
|
||||
|
||||
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/impl/device_cgemm_4gemm_xdl_cshuffle.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
|
||||
|
||||
using ADataType = INT8;
|
||||
using BDataType = INT8;
|
||||
using CDataType = INT8;
|
||||
using AccDataType = INT32;
|
||||
|
||||
using ALayout = ck::tensor_layout::gemm::RowMajor;
|
||||
using BLayout = ck::tensor_layout::gemm::ColumnMajor;
|
||||
using CLayout = ck::tensor_layout::gemm::RowMajor;
|
||||
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
|
||||
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
|
||||
|
||||
using ReferenceCGemmInstance = ck::tensor_operation::host::
|
||||
ReferenceCGemm<ADataType, BDataType, CDataType, PassThrough, PassThrough, PassThrough>;
|
||||
|
||||
// clang-format off
|
||||
using DeviceCGemmInstance = ck::tensor_operation::device::DeviceCGemm_4Gemm_Xdl_CShuffle
|
||||
<ALayout, // typename ALayout
|
||||
BLayout, // typename BLayout
|
||||
CLayout, // typename CLayout
|
||||
ADataType, // typename ADataType
|
||||
BDataType, // typename BDataType
|
||||
CDataType, // typename CDataType
|
||||
AccDataType, // typename GemmAccDataType
|
||||
CDataType, // typename CShuffleDataType
|
||||
PassThrough, // typename AElementwiseOperation
|
||||
PassThrough, // typename BElementwiseOperation
|
||||
PassThrough, // typename CElementwiseOperation
|
||||
GemmDefault, // GemmSpecialization GemmSpec
|
||||
1, // index_t NumGemmKPrefetchStage
|
||||
256, // index_t BlockSize
|
||||
256, // index_t MPerBlock
|
||||
128, // index_t NPerBlock
|
||||
64, // index_t KPerBlock
|
||||
16, // index_t AK1
|
||||
16, // index_t BK1
|
||||
32, // index_t MPerXDL
|
||||
32, // index_t NPerXDL
|
||||
4, // index_t MXdlPerWave
|
||||
2, // index_t NXdlPerWave
|
||||
S<4, 64, 1>, // typename ABlockTransferThreadClusterLengths_AK0_M_AK1
|
||||
S<1, 0, 2>, // typename ABlockTransferThreadClusterArrangeOrder
|
||||
S<1, 0, 2>, // typename ABlockTransferSrcAccessOrder
|
||||
2, // index_t ABlockTransferSrcVectorDim
|
||||
16, // index_t ABlockTransferSrcScalarPerVector
|
||||
16, // index_t ABlockTransferDstScalarPerVector_AK1
|
||||
1, // index_t ABlockLdsExtraM
|
||||
S<4, 64, 1>, // typename BBlockTransferThreadClusterLengths_BK0_N_BK1
|
||||
S<1, 0, 2>, // typename BBlockTransferThreadClusterArrangeOrder
|
||||
S<1, 0, 2>, // typename BBlockTransferSrcAccessOrder
|
||||
2, // index_t BBlockTransferSrcVectorDim
|
||||
8, // index_t BBlockTransferSrcScalarPerVector
|
||||
8, // index_t BBlockTransferDstScalarPerVector_BK1
|
||||
1, // index_t BBlockLdsExtraN
|
||||
1, // index_t CShuffleMXdlPerWavePerShuffle
|
||||
1, // index_t CShuffleNXdlPerWavePerShuffle
|
||||
S<1, 64, 1, 4>, // typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
|
||||
16>; // index_t CShuffleBlockTransferScalarPerVector_NPerBlock
|
||||
// clang-format on
|
||||
|
||||
int main(int argc, char* argv[])
|
||||
{
|
||||
bool do_verification = true;
|
||||
int init_method = 1;
|
||||
bool time_kernel = false;
|
||||
|
||||
// CGEMM 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;
|
||||
|
||||
if(argc == 4)
|
||||
{
|
||||
do_verification = std::stoi(argv[1]);
|
||||
init_method = std::stoi(argv[2]);
|
||||
time_kernel = std::stoi(argv[3]);
|
||||
}
|
||||
else if(argc == 10)
|
||||
{
|
||||
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]);
|
||||
}
|
||||
else
|
||||
{
|
||||
std::cout << "arg1: verification (0=no, 1=yes)\n"
|
||||
<< "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
|
||||
<< "arg3: run kernel # of times (>1)\n"
|
||||
<< "arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n"
|
||||
<< std::endl;
|
||||
exit(0);
|
||||
}
|
||||
|
||||
return !run_cgemm_xdl<ADataType,
|
||||
BDataType,
|
||||
CDataType,
|
||||
ALayout,
|
||||
BLayout,
|
||||
CLayout,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
DeviceCGemmInstance,
|
||||
ReferenceCGemmInstance>(
|
||||
M, N, K, StrideA, StrideB, StrideC, do_verification, init_method, time_kernel);
|
||||
}
|
||||
Reference in New Issue
Block a user