ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-20 21:09:08 +00:00
Code Issues Packages Projects Releases Wiki Activity

Add unit tests for grouped gemm two stage (#1256)

Browse Source 
* add unit tests for grouped gemm two stage

* add reviewers suggestions

---------

Co-authored-by: Adam Osewski <19374865+aosewski@users.noreply.github.com>

[ROCm/composable_kernel commit: 3e3471d5d2]
This commit is contained in:
jakpiase
2024-05-15 10:03:39 +02:00
committed by GitHub
parent bfabd9aa9a
commit c63db2b2ab
5 changed files with 189 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
test/grouped_gemm/CMakeLists.txt
View File

@@ -6,6 +6,12 @@ if(result EQUAL 0)
add_dependencies(test_grouped_gemm test_grouped_gemm_splitk)
endif()
add_gtest_executable(test_grouped_gemm_two_stage_splitk test_grouped_gemm_two_stage_multiple_d_splitk_xdl.cpp)
if(result EQUAL 0)
target_link_libraries(test_grouped_gemm_two_stage_splitk PRIVATE utility device_grouped_gemm_instance)
add_dependencies(test_grouped_gemm test_grouped_gemm_two_stage_splitk)
endif()
add_gtest_executable(test_grouped_gemm_interface test_grouped_gemm_interface_xdl.cpp)
if(result EQUAL 0)
target_link_libraries(test_grouped_gemm_interface PRIVATE utility device_grouped_gemm_instance)

62
test/grouped_gemm/test_grouped_gemm_two_stage_multiple_d_splitk_xdl.cpp Normal file
View File

@@ -0,0 +1,62 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include <tuple>
#include <vector>
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/utility/data_type.hpp"
#include "gtest/gtest.h"
#include "test_grouped_gemm_util.hpp"
using F16 = ck::half_t;
using BF16 = ck::bhalf_t;
using I8 = int8_t;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using RRR_F16_F16_F16 = ck::test::TestGroupedGemmTwoStage<std::tuple<Row, Row, Row, F16, F16, F16>>;
using RCR_F16_F16_F16 = ck::test::TestGroupedGemmTwoStage<std::tuple<Row, Col, Row, F16, F16, F16>>;
using RRR_F16_F16_F16_LargeK =
ck::test::TestGroupedGemmTwoStage<std::tuple<Row, Row, Row, F16, F16, F16>>;
using RCR_F16_F16_F16_LargeK =
ck::test::TestGroupedGemmTwoStage<std::tuple<Row, Col, Row, F16, F16, F16>>;
using RRR_BF16_BF16_BF16 =
ck::test::TestGroupedGemmTwoStage<std::tuple<Row, Row, Row, BF16, BF16, BF16>>;
using RCR_BF16_BF16_BF16 =
ck::test::TestGroupedGemmTwoStage<std::tuple<Row, Col, Row, BF16, BF16, BF16>>;
using RRR_BF16_I8_BF16 =
ck::test::TestGroupedGemmTwoStage<std::tuple<Row, Row, Row, BF16, I8, BF16>>;
using RCR_BF16_I8_BF16 =
ck::test::TestGroupedGemmTwoStage<std::tuple<Row, Col, Row, BF16, I8, BF16>>;
const std::vector<int> KBATCH{1, 2, 3, 5, 8};
INSTANTIATE_TEST_SUITE_P(TestGroupedGemmTwoStage_splitk_MK_KN,
RRR_F16_F16_F16,
testing::ValuesIn(KBATCH));
INSTANTIATE_TEST_SUITE_P(TestGroupedGemmTwoStage_splitk_MK_NK,
RCR_F16_F16_F16,
testing::ValuesIn(KBATCH));
INSTANTIATE_TEST_SUITE_P(TestGroupedGemmTwoStage_splitk_MK_KN_BF16,
RRR_BF16_BF16_BF16,
testing::ValuesIn(KBATCH));
INSTANTIATE_TEST_SUITE_P(TestGroupedGemmTwoStage_splitk_MK_NK_BF16,
RCR_BF16_BF16_BF16,
testing::ValuesIn(KBATCH));
INSTANTIATE_TEST_SUITE_P(TestGroupedGemmTwoStage_splitk_MK_KN_BF16_INT8,
RRR_BF16_I8_BF16,
testing::ValuesIn(KBATCH));
INSTANTIATE_TEST_SUITE_P(TestGroupedGemmTwoStage_splitk_MK_NK_BF16_INT8,
RCR_BF16_I8_BF16,
testing::ValuesIn(KBATCH));
INSTANTIATE_TEST_SUITE_P(TestGroupedGemmTwoStage_splitk_LargeK_MK_KN,
RRR_F16_F16_F16_LargeK,
testing::Values(32, 64));
INSTANTIATE_TEST_SUITE_P(TestGroupedGemmTwoStage_splitk_LargeK_MK_NK,
RCR_F16_F16_F16_LargeK,
testing::Values(32, 64));
#include "test_grouped_gemm_ut_cases.inc"
#include "test_grouped_gemm_two_stage_ut_cases.inc"

61
test/grouped_gemm/test_grouped_gemm_two_stage_ut_cases.inc Normal file
View File

@@ -0,0 +1,61 @@
#pragma once
TEST_P(RRR_BF16_BF16_BF16, MNKPadded)
{
const std::vector<int> Ms{127, 150, 188, 210};
constexpr int N = 136;
constexpr int K = 280;
const std::vector<int> Ns(Ms.size(), N);
const std::vector<int> Ks(Ms.size(), K);
const std::vector<int> StrideAs(Ms.size(), K);
const std::vector<int> StrideBs(Ms.size(), N);
const std::vector<int> StrideCs(Ms.size(), N);
this->Run(Ms, Ns, Ks, StrideAs, StrideBs, StrideCs, this->GetParam());
}
TEST_P(RCR_BF16_BF16_BF16, MNKPadded)
{
const std::vector<int> Ms{127, 150, 188, 210};
constexpr int N = 136;
constexpr int K = 280;
const std::vector<int> Ns(Ms.size(), N);
const std::vector<int> Ks(Ms.size(), K);
const std::vector<int> StrideAs(Ms.size(), K);
const std::vector<int> StrideBs(Ms.size(), K);
const std::vector<int> StrideCs(Ms.size(), N);
this->Run(Ms, Ns, Ks, StrideAs, StrideBs, StrideCs, this->GetParam());
}
TEST_P(RRR_BF16_I8_BF16, MNKPadded)
{
const std::vector<int> Ms{127, 150, 188, 210};
constexpr int N = 136;
constexpr int K = 280;
const std::vector<int> Ns(Ms.size(), N);
const std::vector<int> Ks(Ms.size(), K);
const std::vector<int> StrideAs(Ms.size(), K);
const std::vector<int> StrideBs(Ms.size(), N);
const std::vector<int> StrideCs(Ms.size(), N);
this->Run(Ms, Ns, Ks, StrideAs, StrideBs, StrideCs, this->GetParam());
}
TEST_P(RCR_BF16_I8_BF16, MNKPadded)
{
const std::vector<int> Ms{127, 150, 188, 210};
constexpr int N = 136;
constexpr int K = 280;
const std::vector<int> Ns(Ms.size(), N);
const std::vector<int> Ks(Ms.size(), K);
const std::vector<int> StrideAs(Ms.size(), K);
const std::vector<int> StrideBs(Ms.size(), K);
const std::vector<int> StrideCs(Ms.size(), N);
this->Run(Ms, Ns, Ks, StrideAs, StrideBs, StrideCs, this->GetParam());
}

55
test/grouped_gemm/test_grouped_gemm_util.hpp
View File

@@ -1,5 +1,5 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
@@ -22,6 +22,7 @@
#include "ck/utility/tuple.hpp"
#include "ck/utility/number.hpp"
#include "profiler/profile_grouped_gemm_impl.hpp"
#include "profiler/profile_grouped_gemm_two_stage_impl.hpp"
namespace ck {
namespace test {
@@ -90,6 +91,58 @@ class TestGroupedGemm : public testing::TestWithParam<int>
}
};
template <typename Tuple>
class TestGroupedGemmTwoStage : public testing::TestWithParam<int>
{
protected:
using ALayout = std::tuple_element_t<0, Tuple>;
using BLayout = std::tuple_element_t<1, Tuple>;
using ELayout = std::tuple_element_t<2, Tuple>;
using ADataType = std::tuple_element_t<3, Tuple>;
using BDataType = std::tuple_element_t<4, Tuple>;
using EDataType = std::tuple_element_t<5, Tuple>;
public:
static constexpr bool verify_ = true;
static constexpr int init_method_ = 1; // decimal value initialization
static constexpr bool log_ = false;
static constexpr bool bench_ = false; // measure kernel performance
void SetUp() override {}
void Run(const std::vector<int>& Ms,
const std::vector<int>& Ns,
const std::vector<int>& Ks,
const std::vector<int>& StrideAs,
const std::vector<int>& StrideBs,
const std::vector<int>& StrideCs,
int kbatch = 1,
int n_warmup = 1,
int n_iter = 10)
{
bool pass = ck::profiler::profile_grouped_gemm_two_stage_impl<ADataType,
BDataType,
EDataType,
float,
ALayout,
BLayout,
ELayout>(verify_,
init_method_,
log_,
bench_,
Ms,
Ns,
Ks,
StrideAs,
StrideBs,
StrideCs,
kbatch,
n_warmup,
n_iter);
EXPECT_TRUE(pass);
}
};
template <typename ALayout,
typename BLayout,
typename ELayout,