ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-24 14:54:47 +00:00
Code Issues Packages Projects Releases Wiki Activity

Batched gemm and reduction (#156)

Browse Source 
* adding batched_gemm_and_reduction

* batched_gemm_reduce works with bactch_count=1

* fix a bug in grid_size; batched_gemm_reduce works for batch_count > 1

* adding profiler for batched_gemm_fp16

* fixed a bug in declaration of d1 and d0; both example and profiler work

* clang-format

* cleanup

* batched_gemm_reduce: add test

* minor change

* fixed some typo in function names

[ROCm/composable_kernel commit: 34c661e71c]
This commit is contained in:
Jianfeng Yan
2022-03-30 11:21:18 -05:00
committed by GitHub
parent 6d537a8c3e
commit cb97ce68d8
27 changed files with 2145 additions and 62 deletions
Download Patch File Download Diff File Expand all files Collapse all files

154
profiler/src/profile_batched_gemm_reduce.cpp Normal file
View File

@@ -0,0 +1,154 @@
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <stdlib.h>
#include <half.hpp>
#include "profile_batched_gemm_reduce_impl.hpp"
int profile_batched_gemm_reduce(int argc, char* argv[])
{
enum struct GemmMatrixLayout_t
{
MK_KN_MN, // 0
MK_NK_MN, // 1
KM_KN_MN, // 2
KM_NK_MN, // 3
};
enum struct GemmReduceDataType_t
{
F32_F32_F32_F32_F32, // 0
F16_F16_F16_F32_F32, // 1
};
if(!(argc == 15 || argc == 16))
{
printf("arg1: tensor operation (batched_gemm: BatchedGEMM+Reduce)\n");
printf("arg2: data type (0: fp32; 1: fp16)\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("arg8: print tensor value (0: no; 1: yes)\n");
printf("arg7: run kernel # of times (>1)\n");
printf("arg8 to 14: M, N, K, StrideA, StrideB, StrideC, BatchCount\n");
printf("arg15: split k into mulitiple batch\n");
exit(1);
}
const auto data_type = static_cast<GemmReduceDataType_t>(std::stoi(argv[2]));
const auto layout = static_cast<GemmMatrixLayout_t>(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 int nrepeat = 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]);
const int BatchCount = std::stoi(argv[14]);
if(data_type == GemmReduceDataType_t::F16_F16_F16_F32_F32 &&
layout == GemmMatrixLayout_t::MK_KN_MN)
{
ck::profiler::profile_batched_gemm_reduce_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,
nrepeat,
M,
N,
K,
(StrideA < 0) ? K : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
BatchCount);
}
else if(data_type == GemmReduceDataType_t::F16_F16_F16_F32_F32 &&
layout == GemmMatrixLayout_t::MK_NK_MN)
{
ck::profiler::profile_batched_gemm_reduce_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,
nrepeat,
M,
N,
K,
(StrideA < 0) ? K : StrideA,
(StrideB < 0) ? K : StrideB,
(StrideC < 0) ? N : StrideC,
BatchCount);
}
else if(data_type == GemmReduceDataType_t::F16_F16_F16_F32_F32 &&
layout == GemmMatrixLayout_t::KM_KN_MN)
{
ck::profiler::profile_batched_gemm_reduce_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,
nrepeat,
M,
N,
K,
(StrideA < 0) ? M : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
BatchCount);
}
else if(data_type == GemmReduceDataType_t::F16_F16_F16_F32_F32 &&
layout == GemmMatrixLayout_t::KM_NK_MN)
{
ck::profiler::profile_batched_gemm_reduce_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,
nrepeat,
M,
N,
K,
(StrideA < 0) ? M : StrideA,
(StrideB < 0) ? K : StrideB,
(StrideC < 0) ? N : StrideC,
BatchCount);
}
else
{
throw std::runtime_error("wrong! this data_type & layout is not implemented");
}
return 1;
}

5
profiler/src/profiler.cpp
View File

@@ -17,6 +17,7 @@ int profile_conv_fwd_bias_relu_add(int, char*[]);
int profile_conv_fwd_bias_relu_atomic_add(int, char*[]);
int profile_convnd_bwd_data(int, char*[], int);
int profile_reduce(int, char*[]);
int profile_batched_gemm_reduce(int, char*[]);
int main(int argc, char* argv[])
{
@@ -44,6 +45,10 @@ int main(int argc, char* argv[])
{
return profile_batched_gemm(argc, argv);
}
else if(strcmp(argv[1], "batched_gemm_reduce") == 0)
{
return profile_batched_gemm_reduce(argc, argv);
}
else if(strcmp(argv[1], "grouped_gemm") == 0)
{
profile_grouped_gemm(argc, argv);