ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-21 05:19:20 +00:00
Code Issues Packages Projects Releases Wiki Activity

Add DeviceBatchedGemmMultipleD_Dl (#732)

Browse Source 
* Add DeviceBatchedGemmMultipleD_Dl

* Fix batched_gemm tests

* Fix comments

* test_batched_gemm_multi_d fixes

* Fix args for isSupported batchedGemmMultipleDDl

* Disable tests for gfx90a

[ROCm/composable_kernel commit: fc9f97568f]
This commit is contained in:
Bartłomiej Kocot
2023-06-12 15:37:15 +02:00
committed by GitHub
parent 830b346bbb
commit a404cc8faf
31 changed files with 3347 additions and 123 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
profiler/src/CMakeLists.txt
View File

@@ -34,6 +34,7 @@ set(PROFILER_SOURCES
profile_grouped_gemm_fastgelu.cpp
profile_contraction_bilinear.cpp
profile_contraction_scale.cpp
profile_batched_gemm_multi_d.cpp
)
set(PROFILER_EXECUTABLE ckProfiler)
@@ -77,5 +78,5 @@ target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fastgel
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_bilinear_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_scale_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_pool_fwd_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_multi_d_instance)
rocm_install(TARGETS ${PROFILER_EXECUTABLE} COMPONENT profiler)

103
profiler/src/profile_batched_gemm.cpp
View File

@@ -10,6 +10,8 @@
#include "profiler/profile_batched_gemm_impl.hpp"
#include "profiler_operation_registry.hpp"
#include "ck/library/tensor_operation_instance/gpu/batched_gemm.hpp"
enum struct GemmMatrixLayout
{
MK_KN_MN, // 0
@@ -78,55 +80,72 @@ int profile_batched_gemm(int argc, char* argv[])
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
auto profile = [&](auto a_type,
auto b_type,
auto c_type,
auto a_layout,
auto b_layout,
auto c_layout) {
using ADataType = decltype(a_type);
using BDataType = decltype(b_type);
using CDataType = decltype(c_type);
auto profile =
[&](auto a_type, auto b_type, auto c_type, auto a_layout, auto b_layout, auto c_layout) {
using ADataType = decltype(a_type);
using BDataType = decltype(b_type);
using CDataType = decltype(c_type);
using ALayout = decltype(a_layout);
using BLayout = decltype(b_layout);
using CLayout = decltype(c_layout);
using ALayout = decltype(a_layout);
using BLayout = decltype(b_layout);
using CLayout = decltype(c_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 DefaultStrideC = ck::is_same_v<CLayout, Row> ? N : M;
const int DefaultStrideA = ck::is_same_v<ALayout, Row> ? K : M;
const int DefaultStrideB = ck::is_same_v<BLayout, Row> ? N : K;
const int DefaultStrideC = ck::is_same_v<CLayout, Row> ? N : M;
const int StrideA_ = (StrideA < 0) ? DefaultStrideA : StrideA;
const int StrideB_ = (StrideB < 0) ? DefaultStrideB : StrideB;
const int StrideC_ = (StrideC < 0) ? DefaultStrideC : StrideC;
const int StrideA_ = (StrideA < 0) ? DefaultStrideA : StrideA;
const int StrideB_ = (StrideB < 0) ? DefaultStrideB : StrideB;
const int StrideC_ = (StrideC < 0) ? DefaultStrideC : StrideC;
const int DefaultBatchStrideA = (ck::is_same_v<ALayout, Row> ? M : K) * StrideA_;
const int DefaultBatchStrideB = (ck::is_same_v<BLayout, Row> ? K : N) * StrideB_;
const int DefaultBatchStrideC = (ck::is_same_v<CLayout, Row> ? M : N) * StrideC_;
const int DefaultBatchStrideA = (ck::is_same_v<ALayout, Row> ? M : K) * StrideA_;
const int DefaultBatchStrideB = (ck::is_same_v<BLayout, Row> ? K : N) * StrideB_;
const int DefaultBatchStrideC = (ck::is_same_v<CLayout, Row> ? M : N) * StrideC_;
const int BatchStrideA_ = (BatchStrideA < 0) ? DefaultBatchStrideA : BatchStrideA;
const int BatchStrideB_ = (BatchStrideB < 0) ? DefaultBatchStrideB : BatchStrideB;
const int BatchStrideC_ = (BatchStrideC < 0) ? DefaultBatchStrideC : BatchStrideC;
const int BatchStrideA_ = (BatchStrideA < 0) ? DefaultBatchStrideA : BatchStrideA;
const int BatchStrideB_ = (BatchStrideB < 0) ? DefaultBatchStrideB : BatchStrideB;
const int BatchStrideC_ = (BatchStrideC < 0) ? DefaultBatchStrideC : BatchStrideC;
bool pass = ck::profiler::
profile_batched_gemm_impl<ADataType, BDataType, CDataType, ALayout, BLayout, CLayout>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
BatchStrideA_,
BatchStrideB_,
BatchStrideC_,
StrideA_,
StrideB_,
StrideC_,
BatchCount);
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
return pass ? 0 : 1;
};
using DeviceOp = ck::tensor_operation::device::DeviceBatchedGemm<ALayout,
BLayout,
CLayout,
ADataType,
BDataType,
CDataType,
AElementOp,
BElementOp,
CElementOp>;
bool pass = ck::profiler::profile_batched_gemm_impl<ADataType,
BDataType,
CDataType,
ALayout,
BLayout,
CLayout,
AElementOp,
BElementOp,
CElementOp,
DeviceOp>(do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
BatchStrideA_,
BatchStrideB_,
BatchStrideC_,
StrideA_,
StrideB_,
StrideC_,
BatchCount);
return pass ? 0 : 1;
};
if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_KN_MN)
{

190
profiler/src/profile_batched_gemm_multi_d.cpp Normal file
View File

@@ -0,0 +1,190 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdint>
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "profiler/profile_batched_gemm_impl.hpp"
#include "profiler_operation_registry.hpp"
#include "ck/library/tensor_operation_instance/gpu/batched_gemm_multi_d.hpp"
enum struct GemmMatrixLayout
{
MK_KN_MN, // 0
MK_NK_MN, // 1
KM_KN_MN, // 2
KM_NK_MN, // 3
};
enum struct GemmDataType
{
F16_F16_F16, // 0
INT8_INT8_INT8, // 1
};
#define OP_NAME "batched_gemm_multi_d"
#define OP_DESC "Batched GEMM multi D"
int profile_batched_gemm_multi_d(int argc, char* argv[])
{
if(argc != 18)
{
// clang-format off
printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
printf("arg2: data type (0: fp16; 1: int8)\n");
printf("arg3: matrix layout (0: A[g, m, k] * B[g, k, n] = C[g, m, n];\n");
printf(" 1: A[g, m, k] * B[g, n, k] = C[g, m, n];\n");
printf(" 2: A[g, k, m] * B[g, k, n] = C[g, m, n];\n");
printf(" 3: A[g, k, m] * B[g, n, k] = C[g, 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 17: M, N, K, StrideA, StrideB, StrideC, BatchStrideA, BatchStrideB, BatchStrideC, BatchCount\n");
// clang-format on
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]);
const int BatchStrideA = std::stoi(argv[14]);
const int BatchStrideB = std::stoi(argv[15]);
const int BatchStrideC = std::stoi(argv[16]);
const int BatchCount = std::stoi(argv[17]);
using F16 = ck::half_t;
using INT8 = int8_t;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
auto profile =
[&](auto a_type, auto b_type, auto c_type, auto a_layout, auto b_layout, auto c_layout) {
using ADataType = decltype(a_type);
using BDataType = decltype(b_type);
using CDataType = decltype(c_type);
using DsDataType = ck::Tuple<>;
using ALayout = decltype(a_layout);
using BLayout = decltype(b_layout);
using CLayout = decltype(c_layout);
using DsLayout = ck::Tuple<>;
const int DefaultStrideA = ck::is_same_v<ALayout, Row> ? K : M;
const int DefaultStrideB = ck::is_same_v<BLayout, Row> ? N : K;
const int DefaultStrideC = ck::is_same_v<CLayout, Row> ? N : M;
const int StrideA_ = (StrideA < 0) ? DefaultStrideA : StrideA;
const int StrideB_ = (StrideB < 0) ? DefaultStrideB : StrideB;
const int StrideC_ = (StrideC < 0) ? DefaultStrideC : StrideC;
const int DefaultBatchStrideA = (ck::is_same_v<ALayout, Row> ? M : K) * StrideA_;
const int DefaultBatchStrideB = (ck::is_same_v<BLayout, Row> ? K : N) * StrideB_;
const int DefaultBatchStrideC = (ck::is_same_v<CLayout, Row> ? M : N) * StrideC_;
const int BatchStrideA_ = (BatchStrideA < 0) ? DefaultBatchStrideA : BatchStrideA;
const int BatchStrideB_ = (BatchStrideB < 0) ? DefaultBatchStrideB : BatchStrideB;
const int BatchStrideC_ = (BatchStrideC < 0) ? DefaultBatchStrideC : BatchStrideC;
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
using DeviceOp = ck::tensor_operation::device::DeviceBatchedGemmMultiD<ALayout,
BLayout,
DsLayout,
CLayout,
ADataType,
BDataType,
DsDataType,
CDataType,
AElementOp,
BElementOp,
CElementOp>;
bool pass = ck::profiler::profile_batched_gemm_impl<ADataType,
BDataType,
CDataType,
ALayout,
BLayout,
CLayout,
AElementOp,
BElementOp,
CElementOp,
DeviceOp>(do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
BatchStrideA_,
BatchStrideB_,
BatchStrideC_,
StrideA_,
StrideB_,
StrideC_,
BatchCount);
return pass ? 0 : 1;
};
if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
{
return profile(F16{}, F16{}, F16{}, Row{}, Row{}, Row{});
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_NK_MN)
{
return profile(F16{}, F16{}, F16{}, Row{}, Col{}, Row{});
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_KN_MN)
{
return profile(F16{}, F16{}, F16{}, Col{}, Row{}, Row{});
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_NK_MN)
{
return profile(F16{}, F16{}, F16{}, Col{}, Col{}, Row{});
}
else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::MK_KN_MN)
{
return profile(INT8{}, INT8{}, INT8{}, Row{}, Row{}, Row{});
}
else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::MK_NK_MN)
{
return profile(INT8{}, INT8{}, INT8{}, Row{}, Col{}, Row{});
}
else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::KM_KN_MN)
{
return profile(INT8{}, INT8{}, INT8{}, Col{}, Row{}, Row{});
}
else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::KM_NK_MN)
{
return profile(INT8{}, INT8{}, INT8{}, Col{}, Col{}, Row{});
}
else
{
std::cout << "this data_type & layout is not implemented" << std::endl;
return 1;
}
}
REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_batched_gemm_multi_d);