Skip to content

Commit

Permalink
Added unit tests for CK Tile compute bound gemm pipeline (#1728)
Browse files Browse the repository at this point in the history
  • Loading branch information
@jakpiase
jakpiase authored Dec 17, 2024
1 parent d46196f commit 627a27b
Show file tree
Hide file tree
Showing 5 changed files with 90 additions and 62 deletions.
2 changes: 1 addition & 1 deletion test/ck_tile/gemm/CMakeLists.txt
View file
Original file line number Diff line number Diff line change
@@ -1,4 +1,4 @@
# Currently ck_tile is only built on gfx9
if(GPU_TARGETS MATCHES "gfx9")
add_gtest_executable(test_ck_tile_gemm_mem_pipeline test_gemm_mem_pipeline.cpp)
add_gtest_executable(test_ck_tile_gemm_pipeline test_gemm_pipeline.cpp)
endif()
36 changes: 0 additions & 36 deletions test/ck_tile/gemm/test_gemm_mem_pipeline.cpp
View file

This file was deleted.

42 changes: 42 additions & 0 deletions test/ck_tile/gemm/test_gemm_pipeline.cpp
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,42 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.

#include <tuple>

#include "gtest/gtest.h"

#include "ck_tile/host.hpp"
#include "test_gemm_pipeline_util.hpp"

using F16 = ck_tile::half_t;
using F32 = float;
using Row = ck_tile::tensor_layout::gemm::RowMajor;
using Col = ck_tile::tensor_layout::gemm::ColumnMajor;
using Intrawave = ck_tile::integral_constant<ck_tile::GemmPipelineScheduler,
ck_tile::GemmPipelineScheduler::Intrawave>;
using Interwave = ck_tile::integral_constant<ck_tile::GemmPipelineScheduler,
ck_tile::GemmPipelineScheduler::Interwave>;
using Mem = ck_tile::integral_constant<GemmPipelineType, GemmPipelineType::Mem>;
using Comp = ck_tile::integral_constant<GemmPipelineType, GemmPipelineType::Comp>;

// clang-format off
using KernelTypes = ::testing::Types<
// ALayout, BLayout, CLayout, ADataType, BDataType, AccDataType, CDataType, GemmPipelineScheduler, PipelineType
std::tuple< Row, Row, Row, F16, F16, F32, F16, Intrawave, Mem>,
std::tuple< Row, Row, Row, F16, F16, F32, F16, Intrawave, Comp>,
std::tuple< Row, Row, Row, F16, F16, F32, F16, Interwave, Mem>,
std::tuple< Row, Col, Row, F16, F16, F32, F16, Intrawave, Mem>,
std::tuple< Row, Col, Row, F16, F16, F32, F16, Intrawave, Comp>,
std::tuple< Row, Col, Row, F16, F16, F32, F16, Interwave, Mem>,
std::tuple< Col, Row, Row, F16, F16, F32, F16, Intrawave, Mem>,
std::tuple< Col, Row, Row, F16, F16, F32, F16, Intrawave, Comp>,
std::tuple< Col, Row, Row, F16, F16, F32, F16, Interwave, Mem>,
std::tuple< Col, Col, Row, F16, F16, F32, F16, Intrawave, Mem>,
std::tuple< Col, Col, Row, F16, F16, F32, F16, Intrawave, Comp>,
std::tuple< Col, Col, Row, F16, F16, F32, F16, Interwave, Mem>
>;
// clang-format on

TYPED_TEST_SUITE(TestCkTileGemmPipeline, KernelTypes);

#include "test_gemm_pipeline_ut_cases.inc"
10 changes: 5 additions & 5 deletions .../gemm/test_gemm_mem_pipeline_ut_cases.inc → ...tile/gemm/test_gemm_pipeline_ut_cases.inc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -3,7 +3,7 @@

#pragma once

TYPED_TEST(TestCkTileGemmMemPipeline, SmallM)
TYPED_TEST(TestCkTileGemmPipeline, SmallM)
{
std::vector<int> Ms{1, 2, 3, 4, 5, 6};
constexpr int N = 1024;
Expand All @@ -13,7 +13,7 @@ TYPED_TEST(TestCkTileGemmMemPipeline, SmallM)
this->Run(M, N, K);
}

TYPED_TEST(TestCkTileGemmMemPipeline, MidLargeM)
TYPED_TEST(TestCkTileGemmPipeline, MidLargeM)
{
std::vector<int> Ms{127, 255, 312, 799, 1573};
constexpr int N = 1024;
Expand All @@ -23,7 +23,7 @@ TYPED_TEST(TestCkTileGemmMemPipeline, MidLargeM)
this->Run(M, N, K);
}

TYPED_TEST(TestCkTileGemmMemPipeline, PaddK)
TYPED_TEST(TestCkTileGemmPipeline, PaddK)
{
std::vector<int> Ms{127};
constexpr int N = 1024;
Expand All @@ -33,7 +33,7 @@ TYPED_TEST(TestCkTileGemmMemPipeline, PaddK)
this->Run(M, N, K);
}

TYPED_TEST(TestCkTileGemmMemPipeline, Regular)
TYPED_TEST(TestCkTileGemmPipeline, Regular)
{
std::vector<int> Ms{512};
constexpr int N = 1024;
Expand All @@ -43,7 +43,7 @@ TYPED_TEST(TestCkTileGemmMemPipeline, Regular)
this->Run(M, N, K);
}

TYPED_TEST(TestCkTileGemmMemPipeline, NotSupportedArgument)
TYPED_TEST(TestCkTileGemmPipeline, NotSupportedArgument)
{
constexpr int M = 512;
constexpr int N = 1025;
Expand Down
62 changes: 42 additions & 20 deletions ...tile/gemm/test_gemm_mem_pipeline_util.hpp → .../ck_tile/gemm/test_gemm_pipeline_util.hpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -11,18 +11,24 @@
#include "ck_tile/ops/epilogue.hpp"
#include "ck_tile/ops/gemm.hpp"

enum struct GemmPipelineType
{
Mem,
Comp
};
template <typename Tuple>
class TestCkTileGemmMemPipeline : public ::testing::Test
class TestCkTileGemmPipeline : public ::testing::Test
{
protected:
using ALayout = std::tuple_element_t<0, Tuple>;
using BLayout = std::tuple_element_t<1, Tuple>;
using CLayout = std::tuple_element_t<2, Tuple>;
using ADataType = std::tuple_element_t<3, Tuple>;
using BDataType = std::tuple_element_t<4, Tuple>;
using AccDataType = std::tuple_element_t<5, Tuple>;
using CDataType = std::tuple_element_t<6, Tuple>;
static constexpr auto Scheduler = std::tuple_element_t<7, Tuple>::value;
using ALayout = std::tuple_element_t<0, Tuple>;
using BLayout = std::tuple_element_t<1, Tuple>;
using CLayout = std::tuple_element_t<2, Tuple>;
using ADataType = std::tuple_element_t<3, Tuple>;
using BDataType = std::tuple_element_t<4, Tuple>;
using AccDataType = std::tuple_element_t<5, Tuple>;
using CDataType = std::tuple_element_t<6, Tuple>;
static constexpr auto Scheduler = std::tuple_element_t<7, Tuple>::value;
static constexpr auto PipelineType = std::tuple_element_t<8, Tuple>::value;
// TODO: expose tile size through test t-param ?

struct gemm_args
Expand Down Expand Up @@ -74,8 +80,13 @@ class TestCkTileGemmMemPipeline : public ::testing::Test

using Traits = ck_tile::TileGemmTraits<kPadM, kPadN, kPadK, ALayout, BLayout, CLayout>;

using BaseGemmPipeline = ck_tile::BaseGemmPipelineAgBgCrMem<
ck_tile::GemmPipelineProblem<ADataType, BDataType, AccDataType, GemmShape, Traits>>;
using BaseGemmPipeline = std::conditional_t<
PipelineType == GemmPipelineType::Mem,
ck_tile::BaseGemmPipelineAgBgCrMem<
ck_tile::GemmPipelineProblem<ADataType, BDataType, AccDataType, GemmShape, Traits>>,
ck_tile::BaseGemmPipelineAgBgCrCompV3<
ck_tile::
GemmPipelineProblem<ADataType, BDataType, AccDataType, GemmShape, Traits>>>;

const ck_tile::index_t num_loop = TilePartitioner::GetLoopNum(args.K);
const bool has_hot_loop = BaseGemmPipeline::BlockHasHotloop(num_loop);
Expand All @@ -85,15 +96,26 @@ class TestCkTileGemmMemPipeline : public ::testing::Test
constexpr bool has_hot_loop_v = has_hot_loop_.value;
constexpr auto tail_number_v = tail_number_.value;

using GemmPipeline = ck_tile::GemmPipelineAgBgCrMem<
ck_tile::UniversalGemmPipelineProblem<ADataType,
BDataType,
AccDataType,
GemmShape,
Traits,
Scheduler,
has_hot_loop_v,
tail_number_v>>;
using GemmPipeline =
std::conditional_t<PipelineType == GemmPipelineType::Mem,
ck_tile::GemmPipelineAgBgCrMem<
ck_tile::UniversalGemmPipelineProblem<ADataType,
BDataType,
AccDataType,
GemmShape,
Traits,
Scheduler,
has_hot_loop_v,
tail_number_v>>,
ck_tile::GemmPipelineAgBgCrCompV3<
ck_tile::UniversalGemmPipelineProblem<ADataType,
BDataType,
AccDataType,
GemmShape,
Traits,
Scheduler,
has_hot_loop_v,
tail_number_v>>>;
using Kernel = ck_tile::GemmKernel<TilePartitioner, GemmPipeline, GemmEpilogue>;
auto kargs = Kernel::MakeKargs(args.p_a,
args.p_b,
Expand Down

0 comments on commit 627a27b

Please sign in to comment.