mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-16 10:59:55 +00:00
9df0a11a51
* ad gelu and fast_gelu
* added GeLU and fast GeLU
* clean up
* add gemm+fastgelu example
* add gemm+gelu instances
* update profiler
* clean up
* clean up
* adding gemm+bias+activation
* clean
* adding bias
* clean
* adding gemm multiple d
* debugging
* add gemm bias add fastgelu
* rename, clean
* refactoring; add readme
* refactor
* refactor
* refactor
* refactor
* refactor
* refactor
* fix
* fix
* update example
* update example
* rename
* update example
* add ckProfiler
* clean
* clean
* clean
* clean
* add client app example
* update readme
* delete obselete files
* remove old client app
* delete old file
* cleaning
* clean
* remove half
* fix header path
* fix header path
* fix header path
* fix header path
* fix header path
* fix header path for all examples
* fix header path
* fix header path
* fix header path
* fix header path
* fix header path
* fix header path
* fix header path
* fix header path
* fix header path
* revert client app example
* clean build
* fix build
* temporary disable client test on Jenkins
* clean
* clean
* clean
[ROCm/composable_kernel commit: d1db6a0c3e]
408 lines
15 KiB
C++
408 lines
15 KiB
C++
#include <iostream>
|
|
#include <numeric>
|
|
#include <initializer_list>
|
|
#include <cstdlib>
|
|
|
|
#include "profiler/include/profile_gemm_impl.hpp"
|
|
|
|
enum struct GemmMatrixLayout
|
|
{
|
|
MK_KN_MN, // 0
|
|
MK_NK_MN, // 1
|
|
KM_KN_MN, // 2
|
|
KM_NK_MN, // 3
|
|
MK_KN_NM, // 4
|
|
MK_NK_NM, // 5
|
|
KM_KN_NM, // 6
|
|
KM_NK_NM, // 7
|
|
};
|
|
|
|
enum struct GemmDataType
|
|
{
|
|
F32_F32_F32, // 0
|
|
F16_F16_F16, // 1
|
|
BF16_BF16_BF16, // 2
|
|
INT8_INT8_INT8, // 3
|
|
};
|
|
|
|
int profile_gemm(int argc, char* argv[])
|
|
{
|
|
if(!(argc == 14 || argc == 15))
|
|
{
|
|
printf("arg1: tensor operation (gemm: GEMM)\n");
|
|
printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n");
|
|
printf("arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n");
|
|
printf(" 1: A[m, k] * B[n, k] = C[m, n];\n");
|
|
printf(" 2: A[k, m] * B[k, n] = C[m, n];\n");
|
|
printf(" 3: A[k, m] * B[n, k] = C[m, n])\n");
|
|
printf("arg4: verification (0: no; 1: yes)\n");
|
|
printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
|
|
printf("arg6: print tensor value (0: no; 1: yes)\n");
|
|
printf("arg7: time kernel (0=n0, 1=yes)\n");
|
|
printf("arg8 to 13: M, N, K, StrideA, StrideB, StrideC\n");
|
|
printf("arg14: split k into mulitiple batch\n");
|
|
exit(1);
|
|
}
|
|
|
|
const auto data_type = static_cast<GemmDataType>(std::stoi(argv[2]));
|
|
const auto layout = static_cast<GemmMatrixLayout>(std::stoi(argv[3]));
|
|
const bool do_verification = std::stoi(argv[4]);
|
|
const int init_method = std::stoi(argv[5]);
|
|
const bool do_log = std::stoi(argv[6]);
|
|
const bool time_kernel = std::stoi(argv[7]);
|
|
|
|
const int M = std::stoi(argv[8]);
|
|
const int N = std::stoi(argv[9]);
|
|
const int K = std::stoi(argv[10]);
|
|
|
|
const int StrideA = std::stoi(argv[11]);
|
|
const int StrideB = std::stoi(argv[12]);
|
|
const int StrideC = std::stoi(argv[13]);
|
|
int KBatch = 1;
|
|
if(argc == 15)
|
|
KBatch = std::stoi(argv[14]);
|
|
|
|
if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
|
|
{
|
|
ck::profiler::profile_gemm_impl<ck::half_t,
|
|
ck::half_t,
|
|
ck::half_t,
|
|
float,
|
|
ck::tensor_layout::gemm::RowMajor,
|
|
ck::tensor_layout::gemm::RowMajor,
|
|
ck::tensor_layout::gemm::RowMajor>(
|
|
do_verification,
|
|
init_method,
|
|
do_log,
|
|
time_kernel,
|
|
M,
|
|
N,
|
|
K,
|
|
(StrideA < 0) ? K : StrideA,
|
|
(StrideB < 0) ? N : StrideB,
|
|
(StrideC < 0) ? N : StrideC,
|
|
KBatch);
|
|
}
|
|
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_NK_MN)
|
|
{
|
|
ck::profiler::profile_gemm_impl<ck::half_t,
|
|
ck::half_t,
|
|
ck::half_t,
|
|
float,
|
|
ck::tensor_layout::gemm::RowMajor,
|
|
ck::tensor_layout::gemm::ColumnMajor,
|
|
ck::tensor_layout::gemm::RowMajor>(
|
|
do_verification,
|
|
init_method,
|
|
do_log,
|
|
time_kernel,
|
|
M,
|
|
N,
|
|
K,
|
|
(StrideA < 0) ? K : StrideA,
|
|
(StrideB < 0) ? K : StrideB,
|
|
(StrideC < 0) ? N : StrideC,
|
|
KBatch);
|
|
}
|
|
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_KN_MN)
|
|
{
|
|
ck::profiler::profile_gemm_impl<ck::half_t,
|
|
ck::half_t,
|
|
ck::half_t,
|
|
float,
|
|
ck::tensor_layout::gemm::ColumnMajor,
|
|
ck::tensor_layout::gemm::RowMajor,
|
|
ck::tensor_layout::gemm::RowMajor>(
|
|
do_verification,
|
|
init_method,
|
|
do_log,
|
|
time_kernel,
|
|
M,
|
|
N,
|
|
K,
|
|
(StrideA < 0) ? M : StrideA,
|
|
(StrideB < 0) ? N : StrideB,
|
|
(StrideC < 0) ? N : StrideC,
|
|
KBatch);
|
|
}
|
|
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_NK_MN)
|
|
{
|
|
ck::profiler::profile_gemm_impl<ck::half_t,
|
|
ck::half_t,
|
|
ck::half_t,
|
|
float,
|
|
ck::tensor_layout::gemm::ColumnMajor,
|
|
ck::tensor_layout::gemm::ColumnMajor,
|
|
ck::tensor_layout::gemm::RowMajor>(
|
|
do_verification,
|
|
init_method,
|
|
do_log,
|
|
time_kernel,
|
|
M,
|
|
N,
|
|
K,
|
|
(StrideA < 0) ? M : StrideA,
|
|
(StrideB < 0) ? K : StrideB,
|
|
(StrideC < 0) ? N : StrideC,
|
|
KBatch);
|
|
}
|
|
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_KN_MN)
|
|
{
|
|
ck::profiler::profile_gemm_impl<float,
|
|
float,
|
|
float,
|
|
float,
|
|
ck::tensor_layout::gemm::RowMajor,
|
|
ck::tensor_layout::gemm::RowMajor,
|
|
ck::tensor_layout::gemm::RowMajor>(
|
|
do_verification,
|
|
init_method,
|
|
do_log,
|
|
time_kernel,
|
|
M,
|
|
N,
|
|
K,
|
|
(StrideA < 0) ? K : StrideA,
|
|
(StrideB < 0) ? N : StrideB,
|
|
(StrideC < 0) ? N : StrideC,
|
|
KBatch);
|
|
}
|
|
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_NK_MN)
|
|
{
|
|
ck::profiler::profile_gemm_impl<float,
|
|
float,
|
|
float,
|
|
float,
|
|
ck::tensor_layout::gemm::RowMajor,
|
|
ck::tensor_layout::gemm::ColumnMajor,
|
|
ck::tensor_layout::gemm::RowMajor>(
|
|
do_verification,
|
|
init_method,
|
|
do_log,
|
|
time_kernel,
|
|
M,
|
|
N,
|
|
K,
|
|
(StrideA < 0) ? K : StrideA,
|
|
(StrideB < 0) ? K : StrideB,
|
|
(StrideC < 0) ? N : StrideC,
|
|
KBatch);
|
|
}
|
|
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_KN_MN)
|
|
{
|
|
ck::profiler::profile_gemm_impl<float,
|
|
float,
|
|
float,
|
|
float,
|
|
ck::tensor_layout::gemm::ColumnMajor,
|
|
ck::tensor_layout::gemm::RowMajor,
|
|
ck::tensor_layout::gemm::RowMajor>(
|
|
do_verification,
|
|
init_method,
|
|
do_log,
|
|
time_kernel,
|
|
M,
|
|
N,
|
|
K,
|
|
(StrideA < 0) ? M : StrideA,
|
|
(StrideB < 0) ? N : StrideB,
|
|
(StrideC < 0) ? N : StrideC,
|
|
KBatch);
|
|
}
|
|
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_NK_MN)
|
|
{
|
|
ck::profiler::profile_gemm_impl<float,
|
|
float,
|
|
float,
|
|
float,
|
|
ck::tensor_layout::gemm::ColumnMajor,
|
|
ck::tensor_layout::gemm::ColumnMajor,
|
|
ck::tensor_layout::gemm::RowMajor>(
|
|
do_verification,
|
|
init_method,
|
|
do_log,
|
|
time_kernel,
|
|
M,
|
|
N,
|
|
K,
|
|
(StrideA < 0) ? M : StrideA,
|
|
(StrideB < 0) ? K : StrideB,
|
|
(StrideC < 0) ? N : StrideC,
|
|
KBatch);
|
|
}
|
|
else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::MK_KN_MN)
|
|
{
|
|
ck::profiler::profile_gemm_impl<int8_t,
|
|
int8_t,
|
|
int8_t,
|
|
int32_t,
|
|
ck::tensor_layout::gemm::RowMajor,
|
|
ck::tensor_layout::gemm::RowMajor,
|
|
ck::tensor_layout::gemm::RowMajor>(
|
|
do_verification,
|
|
init_method,
|
|
do_log,
|
|
time_kernel,
|
|
M,
|
|
N,
|
|
K,
|
|
(StrideA < 0) ? K : StrideA,
|
|
(StrideB < 0) ? N : StrideB,
|
|
(StrideC < 0) ? N : StrideC,
|
|
KBatch);
|
|
}
|
|
else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::MK_NK_MN)
|
|
{
|
|
ck::profiler::profile_gemm_impl<int8_t,
|
|
int8_t,
|
|
int8_t,
|
|
int32_t,
|
|
ck::tensor_layout::gemm::RowMajor,
|
|
ck::tensor_layout::gemm::ColumnMajor,
|
|
ck::tensor_layout::gemm::RowMajor>(
|
|
do_verification,
|
|
init_method,
|
|
do_log,
|
|
time_kernel,
|
|
M,
|
|
N,
|
|
K,
|
|
(StrideA < 0) ? M : StrideA,
|
|
(StrideB < 0) ? K : StrideB,
|
|
(StrideC < 0) ? N : StrideC,
|
|
KBatch);
|
|
}
|
|
else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::KM_KN_MN)
|
|
{
|
|
ck::profiler::profile_gemm_impl<int8_t,
|
|
int8_t,
|
|
int8_t,
|
|
int32_t,
|
|
ck::tensor_layout::gemm::ColumnMajor,
|
|
ck::tensor_layout::gemm::RowMajor,
|
|
ck::tensor_layout::gemm::RowMajor>(
|
|
do_verification,
|
|
init_method,
|
|
do_log,
|
|
time_kernel,
|
|
M,
|
|
N,
|
|
K,
|
|
(StrideA < 0) ? M : StrideA,
|
|
(StrideB < 0) ? N : StrideB,
|
|
(StrideC < 0) ? N : StrideC,
|
|
KBatch);
|
|
}
|
|
else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::KM_NK_MN)
|
|
{
|
|
ck::profiler::profile_gemm_impl<int8_t,
|
|
int8_t,
|
|
int8_t,
|
|
int32_t,
|
|
ck::tensor_layout::gemm::ColumnMajor,
|
|
ck::tensor_layout::gemm::ColumnMajor,
|
|
ck::tensor_layout::gemm::RowMajor>(
|
|
do_verification,
|
|
init_method,
|
|
do_log,
|
|
time_kernel,
|
|
M,
|
|
N,
|
|
K,
|
|
(StrideA < 0) ? M : StrideA,
|
|
(StrideB < 0) ? K : StrideB,
|
|
(StrideC < 0) ? N : StrideC,
|
|
KBatch);
|
|
}
|
|
else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::MK_KN_MN)
|
|
{
|
|
ck::profiler::profile_gemm_impl<ck::bhalf_t,
|
|
ck::bhalf_t,
|
|
ck::bhalf_t,
|
|
float,
|
|
ck::tensor_layout::gemm::RowMajor,
|
|
ck::tensor_layout::gemm::RowMajor,
|
|
ck::tensor_layout::gemm::RowMajor>(
|
|
do_verification,
|
|
init_method,
|
|
do_log,
|
|
time_kernel,
|
|
M,
|
|
N,
|
|
K,
|
|
(StrideA < 0) ? K : StrideA,
|
|
(StrideB < 0) ? N : StrideB,
|
|
(StrideC < 0) ? N : StrideC,
|
|
KBatch);
|
|
}
|
|
else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::MK_NK_MN)
|
|
{
|
|
ck::profiler::profile_gemm_impl<ck::bhalf_t,
|
|
ck::bhalf_t,
|
|
ck::bhalf_t,
|
|
float,
|
|
ck::tensor_layout::gemm::RowMajor,
|
|
ck::tensor_layout::gemm::ColumnMajor,
|
|
ck::tensor_layout::gemm::RowMajor>(
|
|
do_verification,
|
|
init_method,
|
|
do_log,
|
|
time_kernel,
|
|
M,
|
|
N,
|
|
K,
|
|
(StrideA < 0) ? M : StrideA,
|
|
(StrideB < 0) ? K : StrideB,
|
|
(StrideC < 0) ? N : StrideC,
|
|
KBatch);
|
|
}
|
|
else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::KM_KN_MN)
|
|
{
|
|
ck::profiler::profile_gemm_impl<ck::bhalf_t,
|
|
ck::bhalf_t,
|
|
ck::bhalf_t,
|
|
float,
|
|
ck::tensor_layout::gemm::ColumnMajor,
|
|
ck::tensor_layout::gemm::RowMajor,
|
|
ck::tensor_layout::gemm::RowMajor>(
|
|
do_verification,
|
|
init_method,
|
|
do_log,
|
|
time_kernel,
|
|
M,
|
|
N,
|
|
K,
|
|
(StrideA < 0) ? M : StrideA,
|
|
(StrideB < 0) ? N : StrideB,
|
|
(StrideC < 0) ? N : StrideC,
|
|
KBatch);
|
|
}
|
|
else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::KM_NK_MN)
|
|
{
|
|
ck::profiler::profile_gemm_impl<ck::bhalf_t,
|
|
ck::bhalf_t,
|
|
ck::bhalf_t,
|
|
float,
|
|
ck::tensor_layout::gemm::ColumnMajor,
|
|
ck::tensor_layout::gemm::ColumnMajor,
|
|
ck::tensor_layout::gemm::RowMajor>(
|
|
do_verification,
|
|
init_method,
|
|
do_log,
|
|
time_kernel,
|
|
M,
|
|
N,
|
|
K,
|
|
(StrideA < 0) ? M : StrideA,
|
|
(StrideB < 0) ? K : StrideB,
|
|
(StrideC < 0) ? N : StrideC,
|
|
KBatch);
|
|
}
|
|
else
|
|
{
|
|
throw std::runtime_error("wrong! this GEMM data_type & layout is not implemented");
|
|
}
|
|
|
|
return 0;
|
|
}
|