mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-20 04:49:54 +00:00
GEMM with Multiple Source, GEMM+Bias+Add+FastGeLU example and ckProfiler (#241)
* 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 comment
* use type_convert
* clean
* clean element wise op
[ROCm/composable_kernel commit: 56adf7e9cc]
This commit is contained in:
Chao Liu
152
profiler/src/profile_gemm_add_add_fastgelu.cpp
Normal file
152
profiler/src/profile_gemm_add_add_fastgelu.cpp
Normal file
@@ -0,0 +1,152 @@
|
||||
#include <iostream>
|
||||
#include <numeric>
|
||||
#include <initializer_list>
|
||||
#include <cstdlib>
|
||||
#include <stdlib.h>
|
||||
|
||||
#include "profile_gemm_add_add_fastgelu_impl.hpp"
|
||||
|
||||
int profile_gemm_add_add_fastgelu(int argc, char* argv[])
|
||||
{
|
||||
enum struct MatrixLayout
|
||||
{
|
||||
MK_KN_MN_MN_MN, // 0
|
||||
MK_NK_MN_MN_MN, // 1
|
||||
KM_KN_MN_MN_MN, // 2
|
||||
KM_NK_MN_MN_MN, // 3
|
||||
MK_KN_NM_MN_MN, // 4
|
||||
MK_NK_NM_MN_MN, // 5
|
||||
KM_KN_NM_MN_MN, // 6
|
||||
KM_NK_NM_MN_MN, // 7
|
||||
};
|
||||
|
||||
enum struct MatrixDataType
|
||||
{
|
||||
F32_F32_F32_F32_F32, // 0
|
||||
F16_F16_F16_F16_F16, // 1
|
||||
BF16_BF16_BF16_BF16_BF16, // 2
|
||||
INT8_INT8_INT8_INT8_INT8, // 3
|
||||
};
|
||||
|
||||
if(argc != 16)
|
||||
{
|
||||
// clang-format off
|
||||
printf("arg1: tensor operation (gemm_add_add_fastgelu: GEMM+Add+Add+GeLU)\n");
|
||||
printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n");
|
||||
printf("arg3: matrix layout (0: E[m, n] = FastGeLU(A[m, k] * B[k, n] + D0[m, n] + D1[m, n]);\n");
|
||||
printf(" 1: E[m, n] = FastGeLU(A[m, k] * B[n, k] + D0[m, n] + D1[m, n]);\n");
|
||||
printf(" 2: E[m, n] = FastGeLU(A[k, m] * B[k, n] + D0[m, n] + D1[m, n]);\n");
|
||||
printf(" 3: E[m, n] = FastGeLU(A[k, m] * B[n, k] + D0[m, n] + D1[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=no, 1=yes)\n");
|
||||
printf("arg8 to 13: M, N, K, StrideA, StrideB, StrideD0, StrideD1, StrideE\n");
|
||||
// clang-format on
|
||||
exit(1);
|
||||
}
|
||||
|
||||
const auto data_type = static_cast<MatrixDataType>(std::stoi(argv[2]));
|
||||
const auto layout = static_cast<MatrixLayout>(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 StrideD0 = std::stoi(argv[13]);
|
||||
const int StrideD1 = std::stoi(argv[14]);
|
||||
const int StrideE = std::stoi(argv[15]);
|
||||
|
||||
using F16 = ck::half_t;
|
||||
using F32 = float;
|
||||
|
||||
using Row = ck::tensor_layout::gemm::RowMajor;
|
||||
using Col = ck::tensor_layout::gemm::ColumnMajor;
|
||||
|
||||
auto profile = [&](auto a_type,
|
||||
auto b_type,
|
||||
auto acc_type,
|
||||
auto d0_type,
|
||||
auto d1_type,
|
||||
auto e_type,
|
||||
auto a_layout,
|
||||
auto b_layout,
|
||||
auto d0_layout,
|
||||
auto d1_layout,
|
||||
auto e_layout) {
|
||||
using ADataType = decltype(a_type);
|
||||
using BDataType = decltype(b_type);
|
||||
using AccDataType = decltype(acc_type);
|
||||
using D0DataType = decltype(d0_type);
|
||||
using D1DataType = decltype(d1_type);
|
||||
using EDataType = decltype(e_type);
|
||||
|
||||
using ALayout = decltype(a_layout);
|
||||
using BLayout = decltype(b_layout);
|
||||
using D0Layout = decltype(d0_layout);
|
||||
using D1Layout = decltype(d1_layout);
|
||||
using ELayout = decltype(e_layout);
|
||||
|
||||
const int DefaultStrideA = ck::is_same_v<ALayout, Row> ? K : M;
|
||||
const int DefaultStrideB = ck::is_same_v<BLayout, Row> ? N : K;
|
||||
const int DefaultStrideD0 = ck::is_same_v<D0Layout, Row> ? N : M;
|
||||
const int DefaultStrideD1 = ck::is_same_v<D1Layout, Row> ? N : M;
|
||||
const int DefaultStrideE = ck::is_same_v<ELayout, Row> ? N : M;
|
||||
|
||||
return ck::profiler::profile_gemm_add_add_fastgelu_impl<ADataType,
|
||||
BDataType,
|
||||
AccDataType,
|
||||
D0DataType,
|
||||
D1DataType,
|
||||
EDataType,
|
||||
ALayout,
|
||||
BLayout,
|
||||
D0Layout,
|
||||
D1Layout,
|
||||
ELayout>(
|
||||
do_verification,
|
||||
init_method,
|
||||
do_log,
|
||||
time_kernel,
|
||||
M,
|
||||
N,
|
||||
K,
|
||||
(StrideA < 0) ? DefaultStrideA : StrideA,
|
||||
(StrideB < 0) ? DefaultStrideB : StrideB,
|
||||
(StrideD0 < 0) ? DefaultStrideD0 : StrideD0,
|
||||
(StrideD1 < 0) ? DefaultStrideD1 : StrideD1,
|
||||
(StrideE < 0) ? DefaultStrideE : StrideE);
|
||||
};
|
||||
|
||||
if(data_type == MatrixDataType::F16_F16_F16_F16_F16 && layout == MatrixLayout::MK_KN_MN_MN_MN)
|
||||
{
|
||||
return profile(F16{}, F16{}, F32{}, F16{}, F16{}, F16{}, Row{}, Row{}, Row{}, Row{}, Row{});
|
||||
}
|
||||
else if(data_type == MatrixDataType::F16_F16_F16_F16_F16 &&
|
||||
layout == MatrixLayout::MK_NK_MN_MN_MN)
|
||||
{
|
||||
return profile(F16{}, F16{}, F32{}, F16{}, F16{}, F16{}, Row{}, Col{}, Row{}, Row{}, Row{});
|
||||
}
|
||||
else if(data_type == MatrixDataType::F16_F16_F16_F16_F16 &&
|
||||
layout == MatrixLayout::KM_KN_MN_MN_MN)
|
||||
{
|
||||
return profile(F16{}, F16{}, F32{}, F16{}, F16{}, F16{}, Col{}, Row{}, Row{}, Row{}, Row{});
|
||||
}
|
||||
else if(data_type == MatrixDataType::F16_F16_F16_F16_F16 &&
|
||||
layout == MatrixLayout::KM_NK_MN_MN_MN)
|
||||
{
|
||||
return profile(F16{}, F16{}, F32{}, F16{}, F16{}, F16{}, Col{}, Col{}, Row{}, Row{}, Row{});
|
||||
}
|
||||
else
|
||||
{
|
||||
std::cout << "this data_type & layout is not implemented" << std::endl;
|
||||
|
||||
return 0;
|
||||
}
|
||||
}
|
||||
@@ -22,9 +22,39 @@ int profile_convnd_bwd_data(int, char*[], int);
|
||||
int profile_reduce(int, char*[]);
|
||||
int profile_conv_bwd_weight(int, char*[]);
|
||||
int profile_batched_gemm_reduce(int, char*[]);
|
||||
int profile_gemm_add_add_fastgelu(int, char*[]);
|
||||
|
||||
static void print_helper_message()
|
||||
{
|
||||
// clang-format off
|
||||
printf("arg1: tensor operation (gemm: GEMM\n"
|
||||
" gemm_bias_2d: GEMM+Bias(2D)\n"
|
||||
" gemm_bias_relu: GEMM+Bias+ReLU\n"
|
||||
" gemm_bias_relu_add: GEMM+Bias+ReLU+Add\n"
|
||||
" gemm_reduce: GEMM+Reduce\n"
|
||||
" grouped_gemm: Grouped GEMM\n"
|
||||
" conv_fwd: ForwardConvolution\n"
|
||||
" conv_fwd_bias_relu: ForwardConvolution+Bias+ReLU\n"
|
||||
" conv_fwd_bias_relu_add: ForwardConvolution+Bias+ReLU+Add\n"
|
||||
" conv_fwd_bias_relu_atomic_add: ForwardConvolution+Bias+ReLU+AtomicAdd\n"
|
||||
" conv1d_bwd_data: BackwardConvolution data 1 dim\n"
|
||||
" conv2d_bwd_data: BackwardConvolution data 2 dim\n"
|
||||
" conv3d_bwd_data: BackwardConvolution data 3 dim\n"
|
||||
" reduce: Reduce\n"
|
||||
" conv2d_bwd_weight: Backward Weight Convolution 2d\n"
|
||||
" gemm_add_add_fastgelu: GEMM+Add+Add+FastGeLU\n");
|
||||
// clang-format on
|
||||
}
|
||||
|
||||
int main(int argc, char* argv[])
|
||||
{
|
||||
if(argc == 1)
|
||||
{
|
||||
print_helper_message();
|
||||
|
||||
return 0;
|
||||
}
|
||||
|
||||
if(strcmp(argv[1], "gemm") == 0)
|
||||
{
|
||||
return profile_gemm(argc, argv);
|
||||
@@ -97,25 +127,14 @@ int main(int argc, char* argv[])
|
||||
{
|
||||
return profile_conv_bwd_weight(argc, argv);
|
||||
}
|
||||
else if(strcmp(argv[1], "gemm_add_add_fastgelu") == 0)
|
||||
{
|
||||
return profile_gemm_add_add_fastgelu(argc, argv);
|
||||
}
|
||||
else
|
||||
{
|
||||
// clang-format off
|
||||
printf("arg1: tensor operation (gemm: GEMM\n"
|
||||
" gemm_bias_2d: GEMM+Bias(2D)\n"
|
||||
" gemm_bias_relu: GEMM+Bias+ReLU\n"
|
||||
" gemm_bias_relu_add: GEMM+Bias+ReLU+Add\n"
|
||||
" gemm_reduce: GEMM+Reduce\n"
|
||||
" grouped_gemm: Grouped GEMM\n"
|
||||
" conv_fwd: ForwardConvolution\n"
|
||||
" conv_fwd_bias_relu: ForwardConvolution+Bias+ReLU\n"
|
||||
" conv_fwd_bias_relu_add: ForwardConvolution+Bias+ReLU+Add\n"
|
||||
" conv_fwd_bias_relu_atomic_add: ForwardConvolution+Bias+ReLU+AtomicAdd\n"
|
||||
" conv1d_bwd_data: BackwardConvolution data 1 dim\n"
|
||||
" conv2d_bwd_data: BackwardConvolution data 2 dim\n"
|
||||
" conv3d_bwd_data: BackwardConvolution data 3 dim\n"
|
||||
" reduce: Reduce\n"
|
||||
" conv2d_bwd_weight: Backward Weight Convolution 2d\n");
|
||||
// clang-format on
|
||||
print_helper_message();
|
||||
|
||||
return 0;
|
||||
}
|
||||
return 0;
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user