ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-18 20:09:25 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge flatmm Operator with universal gemm (#2434)

Browse Source 
* Initial commit

* Adding new tile partitioner to flatmm

* intermediate changes

* debugging kernels

* Updating flatmm example to universal gemm example

* updated flatmm kernel to run via gemmKernel

* update universal gemm to incorporate flatmm

* debug

* Fix flatmm call

* Fixing other kernels and tests for API changes

* clang formatted

* fixing gemm tests

* added test for flatmm and simplify kernel arguments

* adding flatmm test

* fix test for flatmm

* simplify gemm kernel with flatmm

* remove flatmm related files

* addressing review comments and code clean up

* resolving empty file

* resolving empty file

* clang formatted

* addressing review comments

* enable persistent kernel for flatmm

* reverted the removed files for flatmm

* reverted the removed files for flatmm

* changed flatmm to weightPReshuffle; removed the _1 added in teh faltmm example

* some more renames

* clang formatted

[ROCm/composable_kernel commit: d239b91fd5]
This commit is contained in:
Khushbu Agarwal
2025-07-11 08:27:55 -07:00
committed by GitHub
parent fb42be79dc
commit e34599e8a9
34 changed files with 2736 additions and 338 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
test/ck_tile/CMakeLists.txt
View File

@@ -1,5 +1,6 @@
add_subdirectory(image_to_column)
add_subdirectory(gemm)
add_subdirectory(gemm_weight_preshuffle)
add_subdirectory(batched_gemm)
add_subdirectory(grouped_gemm)
add_subdirectory(gemm_multi_d)

7
test/ck_tile/gemm/test_gemm_pipeline_kernel_types.hpp
View File

@@ -22,9 +22,10 @@ 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 CompV3 = ck_tile::integral_constant<GemmPipelineType, GemmPipelineType::CompV3>;
using CompV4 = ck_tile::integral_constant<GemmPipelineType, GemmPipelineType::CompV4>;
using Mem = ck_tile::integral_constant<GemmPipelineType, GemmPipelineType::Mem>;
using CompV3 = ck_tile::integral_constant<GemmPipelineType, GemmPipelineType::CompV3>;
using CompV4 = ck_tile::integral_constant<GemmPipelineType, GemmPipelineType::CompV4>;
using Persistent = std::true_type;
using NonPersistent = std::false_type;

25
test/ck_tile/gemm/test_gemm_pipeline_util.hpp
View File

@@ -90,7 +90,7 @@ class TestCkTileGemmPipeline : public ::testing::Test
ck_tile::tuple_element_or_default_t<Tuple, 9, std::false_type>::value;
// TODO: expose tile size through test t-param ?
template <bool PadM, bool PadN, bool PadK>
template <bool PadM, bool PadN, bool PadK, bool Preshuffle>
void invoke_gemm(const ck_tile::GemmHostArgs</*NumDTensor = 0*/>& args,
const ck_tile::stream_config& s)
{
@@ -107,9 +107,10 @@ class TestCkTileGemmPipeline : public ::testing::Test
constexpr ck_tile::index_t N_Warp_Tile = 32;
constexpr ck_tile::index_t K_Warp_Tile = 16;
constexpr bool kPadM = PadM;
constexpr bool kPadN = PadN;
constexpr bool kPadK = PadK;
constexpr bool kPadM = PadM;
constexpr bool kPadN = PadN;
constexpr bool kPadK = PadK;
constexpr bool preshuffle = Preshuffle;
constexpr bool DoubleSmemBuffer = (PipelineType == GemmPipelineType::CompV4) ? true : false;
@@ -131,7 +132,9 @@ class TestCkTileGemmPipeline : public ::testing::Test
using Traits = ck_tile::TileGemmTraits<kPadM, kPadN, kPadK, ALayout, BLayout, CLayout>;
static constexpr bool StructuredSparsity = false;
using GemmUniversalTraits = ck_tile::TileGemmUniversalTraits<kPadM,
static constexpr bool NumWaveGroup = 1;
using GemmUniversalTraits = ck_tile::TileGemmUniversalTraits<kPadM,
kPadN,
kPadK,
DoubleSmemBuffer,
@@ -140,7 +143,9 @@ class TestCkTileGemmPipeline : public ::testing::Test
CLayout,
TransposeC,
StructuredSparsity,
Persistent>;
Persistent,
NumWaveGroup,
preshuffle>;
using GemmPipelineProblem =
ck_tile::GemmPipelineProblem<ADataType, BDataType, AccDataType, GemmShape, Traits>;
@@ -261,7 +266,7 @@ class TestCkTileGemmPipeline : public ::testing::Test
}
}
template <bool PadM = true, bool PadN = true, bool PadK = true>
template <bool PadM = true, bool PadN = true, bool PadK = true, bool Preshuffle = false>
void Run(const int M,
const int N,
const int K,
@@ -271,11 +276,11 @@ class TestCkTileGemmPipeline : public ::testing::Test
{
for(auto kb : k_batches_)
{
RunSingle<PadM, PadN, PadK>(M, N, K, StrideA, StrideB, StrideC, kb);
RunSingle<PadM, PadN, PadK, Preshuffle>(M, N, K, StrideA, StrideB, StrideC, kb);
}
}
template <bool PadM, bool PadN, bool PadK>
template <bool PadM, bool PadN, bool PadK, bool Preshuffle>
void RunSingle(const int M,
const int N,
const int K,
@@ -352,7 +357,7 @@ class TestCkTileGemmPipeline : public ::testing::Test
args.stride_B = stride_B;
args.stride_E = stride_C;
invoke_gemm<PadM, PadN, PadK>(args, ck_tile::stream_config{nullptr, false});
invoke_gemm<PadM, PadN, PadK, Preshuffle>(args, ck_tile::stream_config{nullptr, false});
c_m_n_dev_buf.FromDevice(c_m_n_dev_result.data());
bool pass = true;

22
test/ck_tile/gemm_weight_preshuffle/CMakeLists.txt Normal file
View File

@@ -0,0 +1,22 @@
# Currently ck_tile_gemm is only built on gfx94/gfx95
set(EXAMPLE_GEMM_COMPILE_OPTIONS)
if(CK_USE_OCP_FP8)
list(APPEND EXAMPLE_GEMM_COMPILE_OPTIONS -DCK_TILE_USE_OCP_FP8)
endif()
set(EXAMPLE_GEMM_COMPILE_COMPUTE_V4_OPTIONS)
if(CK_USE_OCP_FP8)
list(APPEND EXAMPLE_GEMM_COMPILE_COMPUTE_V4_OPTIONS -DCK_TILE_USE_OCP_FP8)
endif()
list(APPEND EXAMPLE_GEMM_COMPILE_COMPUTE_V4_OPTIONS
-mllvm
-enable-noalias-to-md-conversion=0
)
if(GPU_TARGETS MATCHES "gfx94" OR GPU_TARGETS MATCHES "gfx95")
add_gtest_executable(test_ck_tile_gemm_pipeline_wp test_gemm_pipeline_wp.cpp)
target_compile_options(test_ck_tile_gemm_pipeline_wp PRIVATE ${EXAMPLE_GEMM_COMPILE_OPTIONS})
else()
message(DEBUG "Skipping ck_tile_gemm tests for current target")
endif()

32
test/ck_tile/gemm_weight_preshuffle/test_gemm_pipeline_kernel_types.hpp Normal file
View File

@@ -0,0 +1,32 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
#include <tuple>
#include <type_traits>
#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 F8 = ck_tile::fp8_t;
using Row = ck_tile::tensor_layout::gemm::RowMajor;
using Col = ck_tile::tensor_layout::gemm::ColumnMajor;
using Default = ck_tile::integral_constant<ck_tile::GemmPipelineScheduler,
ck_tile::GemmPipelineScheduler::Default>;
using WeightPreshuffle =
ck_tile::integral_constant<GemmPipelineType, GemmPipelineType::WeightPreshuffle>;
// clang-format off
using KernelTypesWeightPreshuffle = ::testing::Types<
std::tuple< Row, Col, Row, F16, F16, F32, F16, Default, WeightPreshuffle>,
std::tuple< Row, Col, Row, F8, F8, F32, F16, Default, WeightPreshuffle>
>;
// clang-format on

21
test/ck_tile/gemm_weight_preshuffle/test_gemm_pipeline_ut_cases.inc Executable file
View File

@@ -0,0 +1,21 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#ifndef TEST_GEMM_PIPELINE_UT_CASES_INC
#define TEST_GEMM_PIPELINE_UT_CASES_INC
TYPED_TEST(TEST_SUITE_NAME, GemmPreshuffle)
{
constexpr int M = 2048;
constexpr int N = 4096;
constexpr int K = 5120;
constexpr bool PadM = false;
constexpr bool PadN = false;
constexpr bool PadK = false;
constexpr bool Preshuffle = true;
this->template Run<PadM, PadN, PadK, Preshuffle>(M, N, K);
}
#endif

384
test/ck_tile/gemm_weight_preshuffle/test_gemm_pipeline_util.hpp Normal file
View File

@@ -0,0 +1,384 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <sstream>
#include <gtest/gtest.h>
#include "ck_tile/core.hpp"
#include "ck_tile/host.hpp"
#include "ck_tile/host/kernel_launch.hpp"
#include "ck_tile/ops/epilogue.hpp"
#include "ck_tile/ops/gemm.hpp"
template <typename ADataType, typename BDataType, typename AccDataType, typename CDataType>
auto calculate_rtol_atol(const ck_tile::index_t K,
const ck_tile::index_t kbatch,
const float max_accumulated_value)
{
using ComputeType =
std::conditional_t<sizeof(ADataType) < sizeof(BDataType), ADataType, BDataType>;
// Calculate thresholds
const auto rtol = ck_tile::get_relative_threshold<ComputeType, CDataType, AccDataType>(
ck_tile::integer_divide_ceil(K, kbatch));
const auto atol = ck_tile::get_absolute_threshold<ComputeType, CDataType, AccDataType>(
max_accumulated_value / kbatch, ck_tile::integer_divide_ceil(K, kbatch));
// Calculate error due to split_k accumulation
const auto rtol_split_k =
ck_tile::get_relative_threshold<CDataType, CDataType, CDataType>(kbatch);
const auto atol_split_k = ck_tile::get_absolute_threshold<CDataType, CDataType, CDataType>(
max_accumulated_value, kbatch);
// Use higher threshold
return ck_tile::make_tuple(std::max(rtol, rtol_split_k), std::max(atol, atol_split_k));
}
enum struct GemmPipelineType
{
WeightPreshuffle
};
template <GemmPipelineType PT, typename Problem>
struct GemmPipelineTypeSelector;
template <typename Problem>
struct GemmPipelineTypeSelector<GemmPipelineType::WeightPreshuffle, Problem>
{
using base_pipeline = ck_tile::BaseWeightPreshufflePipelineAGmemBGmemCRegV1<Problem>;
using pipeline = ck_tile::WeightPreshufflePipelineAGmemBGmemCRegV1<Problem>;
static constexpr auto GetName() { return "GemmPipelineAgBgCrWeightPreshuffle"; }
};
template <typename Datatype>
struct config
{
static constexpr ck_tile::index_t M_Tile = 128;
static constexpr ck_tile::index_t N_Tile = 128;
static constexpr ck_tile::index_t K_Tile = 128 / sizeof(Datatype);
static constexpr ck_tile::index_t M_Warp = 1;
static constexpr ck_tile::index_t N_Warp = 4;
static constexpr ck_tile::index_t K_Warp = 1;
static constexpr ck_tile::index_t M_Warp_Tile = 32;
static constexpr ck_tile::index_t N_Warp_Tile = 32;
static constexpr ck_tile::index_t K_Warp_Tile = sizeof(Datatype) == 2 ? 16 : 32;
};
template <typename Tuple>
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;
static constexpr auto PipelineType = std::tuple_element_t<8, Tuple>::value;
using DsLayout = ck_tile::tuple<>;
using DsDataType = ck_tile::tuple<>;
using GemmConfig = config<ADataType>;
static constexpr bool Persistent =
ck_tile::tuple_element_or_default_t<Tuple, 9, std::false_type>::value;
// TODO: expose tile size through test t-param ?
template <bool PadM, bool PadN, bool PadK, bool Preshuffle>
void invoke_gemm(const ck_tile::GemmHostArgs</*NumDTensor = 0*/>& args,
const ck_tile::stream_config& s)
{
// TODO: This should be parameterized in tests
// constexpr ck_tile::index_t M_Tile = 128;
// constexpr ck_tile::index_t N_Tile = 128;
// constexpr ck_tile::index_t K_Tile = 128;
// constexpr ck_tile::index_t M_Warp = 1;
// constexpr ck_tile::index_t N_Warp = 4;
// constexpr ck_tile::index_t K_Warp = 1;
// constexpr ck_tile::index_t M_Warp_Tile = 32;
// constexpr ck_tile::index_t N_Warp_Tile = 32;
// constexpr ck_tile::index_t K_Warp_Tile = sizeof(ADataType) == 2 ? 16 : 32;
constexpr bool kPadM = PadM;
constexpr bool kPadN = PadN;
constexpr bool kPadK = PadK;
constexpr bool preshuffle = Preshuffle;
constexpr bool DoubleSmemBuffer = false;
// TODO: For now - but this should also be a test parameter
constexpr bool TransposeC = false;
constexpr int kBlockPerCu = 2;
constexpr ck_tile::index_t TileParitionerGroupNum = 8;
constexpr ck_tile::index_t TileParitionerM01 = 4;
// ===============================================
using GemmShape = ck_tile::TileGemmShape<
ck_tile::sequence<GemmConfig::M_Tile, GemmConfig::N_Tile, GemmConfig::K_Tile>,
ck_tile::sequence<GemmConfig::M_Warp, GemmConfig::N_Warp, GemmConfig::K_Warp>,
ck_tile::sequence<GemmConfig::M_Warp_Tile,
GemmConfig::N_Warp_Tile,
GemmConfig::K_Warp_Tile>>;
using TilePartitioner = ck_tile::
GemmSpatiallyLocalTilePartitioner<GemmShape, TileParitionerGroupNum, TileParitionerM01>;
using Traits = ck_tile::TileGemmTraits<kPadM, kPadN, kPadK, ALayout, BLayout, CLayout>;
static constexpr bool StructuredSparsity = false;
static constexpr bool NumWaveGroup = 1;
using GemmUniversalTraits = ck_tile::TileGemmUniversalTraits<kPadM,
kPadN,
kPadK,
DoubleSmemBuffer,
ALayout,
BLayout,
CLayout,
TransposeC,
StructuredSparsity,
Persistent,
NumWaveGroup,
preshuffle>;
using GemmPipelineProblem =
ck_tile::GemmPipelineProblem<ADataType, BDataType, AccDataType, GemmShape, Traits>;
using BaseGemmPipeline =
typename GemmPipelineTypeSelector<PipelineType, GemmPipelineProblem>::base_pipeline;
const ck_tile::index_t k_grain = args.k_batch * GemmConfig::K_Tile;
const ck_tile::index_t K_split = (args.K + k_grain - 1) / k_grain * GemmConfig::K_Tile;
const ck_tile::index_t num_loop = TilePartitioner::GetLoopNum(K_split);
const bool has_hot_loop = BaseGemmPipeline::BlockHasHotloop(num_loop);
const ck_tile::TailNumber tail_num = BaseGemmPipeline::GetBlockLoopTailNum(num_loop);
const auto Run = [&](const auto has_hot_loop_,
const auto tail_number_,
const auto memory_operation_) {
constexpr bool has_hot_loop_v = has_hot_loop_.value;
constexpr auto tail_number_v = tail_number_.value;
constexpr auto memory_operation = memory_operation_.value;
using UniversalGemmProblem = ck_tile::UniversalGemmPipelineProblem<ADataType,
BDataType,
AccDataType,
GemmShape,
GemmUniversalTraits,
Scheduler,
has_hot_loop_v,
tail_number_v>;
using GemmPipeline =
typename GemmPipelineTypeSelector<PipelineType, UniversalGemmProblem>::pipeline;
using GemmEpilogue = ck_tile::CShuffleEpilogue<
ck_tile::CShuffleEpilogueProblem<ADataType,
BDataType,
DsDataType,
AccDataType,
CDataType,
DsLayout,
CLayout,
ck_tile::element_wise::PassThrough,
GemmPipeline::BlockSize,
TilePartitioner::MPerBlock,
TilePartitioner::NPerBlock,
GemmConfig::M_Warp,
GemmConfig::N_Warp,
GemmConfig::M_Warp_Tile,
GemmConfig::N_Warp_Tile,
GemmConfig::K_Warp_Tile,
UniversalGemmProblem::TransposeC,
memory_operation>>;
using Kernel = ck_tile::GemmKernel<TilePartitioner, GemmPipeline, GemmEpilogue>;
auto kargs = Kernel::MakeKernelArgs(args);
dim3 grids;
if constexpr(Persistent)
{
grids = Kernel::MaxOccupancyGridSize(s);
}
else
{
grids = Kernel::GridSize(args.M, args.N, args.k_batch);
}
constexpr dim3 blocks = Kernel::BlockSize();
if(!Kernel::IsSupportedArgument(kargs))
{
throw std::runtime_error("Wrong! Arguments not supported! Skipping gemm!\n");
}
if(s.log_level_ > 0)
{
std::cout << "Launching kernel with args:"
<< " grid: {" << grids.x << ", " << grids.y << ", " << grids.z << "}"
<< ", blocks: {" << blocks.x << ", " << blocks.y << ", " << blocks.z
<< "}" << std::endl;
}
ck_tile::launch_kernel(
s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
};
const auto RunSplitk = [&](const auto has_hot_loop_, const auto tail_number_) {
if(args.k_batch == 1)
{
Run(has_hot_loop_,
tail_number_,
ck_tile::integral_constant<ck_tile::memory_operation_enum,
ck_tile::memory_operation_enum::set>{});
}
else
{
Run(has_hot_loop_,
tail_number_,
ck_tile::integral_constant<ck_tile::memory_operation_enum,
ck_tile::memory_operation_enum::atomic_add>{});
}
};
BaseGemmPipeline::TailHandler(RunSplitk, has_hot_loop, tail_num);
}
public:
std::vector<int> k_batches_;
void SetUp() override
{
// Otherwise, use k_batch = 1 and 2
k_batches_ = {1};
}
template <bool PadM = true, bool PadN = true, bool PadK = true, bool Preshuffle = false>
void Run(const int M,
const int N,
const int K,
const int StrideA = 0,
const int StrideB = 0,
const int StrideC = 0)
{
for(auto kb : k_batches_)
{
RunSingle<PadM, PadN, PadK, Preshuffle>(M, N, K, StrideA, StrideB, StrideC, kb);
}
}
template <bool PadM, bool PadN, bool PadK, bool Preshuffle>
void RunSingle(const int M,
const int N,
const int K,
const int StrideA,
const int StrideB,
const int StrideC,
int kbatch = 1)
{
using namespace ck_tile::literals;
auto f_host_tensor_descriptor = [](std::size_t row,
std::size_t col,
std::size_t stride,
auto layout) {
if constexpr(std::is_same_v<decltype(layout), ck_tile::tensor_layout::gemm::RowMajor>)
{
return ck_tile::HostTensorDescriptor({row, col}, {stride, 1_uz});
}
else
{
return ck_tile::HostTensorDescriptor({row, col}, {1_uz, stride});
}
};
auto f_get_default_stride =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
if(stride == 0)
{
// give a chance if stride is zero, return a default packed stride
if constexpr(std::is_same_v<decltype(layout),
ck_tile::tensor_layout::gemm::RowMajor>)
{
return col;
}
else
{
return row;
}
}
else
return stride;
};
std::size_t stride_A = f_get_default_stride(M, K, StrideA, ALayout{});
std::size_t stride_B = f_get_default_stride(K, N, StrideB, BLayout{});
std::size_t stride_C = f_get_default_stride(M, N, StrideC, CLayout{});
ck_tile::HostTensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, stride_A, ALayout{}));
ck_tile::HostTensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, stride_B, BLayout{}));
ck_tile::HostTensor<CDataType> c_m_n_dev_result(
f_host_tensor_descriptor(M, N, stride_C, CLayout{}));
ck_tile::FillUniformDistributionIntegerValue<ADataType>{-5, 5}(a_m_k);
ck_tile::FillUniformDistributionIntegerValue<BDataType>{-5, 5}(b_k_n);
ck_tile::DeviceMem a_m_k_dev_buf(a_m_k.get_element_space_size_in_bytes());
ck_tile::DeviceMem b_k_n_dev_buf(b_k_n.get_element_space_size_in_bytes());
ck_tile::DeviceMem c_m_n_dev_buf(c_m_n_dev_result.get_element_space_size_in_bytes());
constexpr int divisor = GemmConfig::N_Warp_Tile == 32 ? 2 : 4;
ck_tile::HostTensor<BDataType> t_view({N / GemmConfig::N_Warp_Tile,
GemmConfig::N_Warp_Tile,
K / GemmConfig::K_Warp_Tile,
divisor,
GemmConfig::K_Warp_Tile / divisor});
std::copy(b_k_n.begin(), b_k_n.end(), t_view.begin());
ck_tile::HostTensor<BDataType> b_shuffle_host =
ck_tile::reference_permute(t_view, {0, 2, 3, 1, 4});
a_m_k_dev_buf.ToDevice(a_m_k.data());
b_k_n_dev_buf.ToDevice(b_shuffle_host.data());
c_m_n_dev_buf.SetZero();
c_m_n_dev_result.SetZero();
ck_tile::GemmHostArgs</*NumDTensor = 0*/> args;
args.a_ptr = a_m_k_dev_buf.GetDeviceBuffer();
args.b_ptr = b_k_n_dev_buf.GetDeviceBuffer();
args.e_ptr = c_m_n_dev_buf.GetDeviceBuffer();
args.k_batch = kbatch;
args.M = M;
args.N = N;
args.K = K;
args.stride_A = stride_A;
args.stride_B = stride_B;
args.stride_E = stride_C;
invoke_gemm<PadM, PadN, PadK, Preshuffle>(args, ck_tile::stream_config{nullptr, false});
c_m_n_dev_buf.FromDevice(c_m_n_dev_result.data());
bool pass = true;
ck_tile::HostTensor<CDataType> c_m_n_host_ref(
f_host_tensor_descriptor(M, N, stride_C, CLayout{}));
c_m_n_host_ref.SetZero();
ck_tile::reference_gemm<ADataType, BDataType, AccDataType, CDataType>(
a_m_k, b_k_n, c_m_n_host_ref);
const float max_accumulated_value =
*std::max_element(c_m_n_host_ref.mData.begin(), c_m_n_host_ref.mData.end());
const auto rtol_atol = calculate_rtol_atol<ADataType, BDataType, AccDataType, CDataType>(
K, kbatch, max_accumulated_value);
pass = ck_tile::check_err(c_m_n_dev_result,
c_m_n_host_ref,
"Error: Incorrect results!",
rtol_atol.at(ck_tile::number<0>{}),
rtol_atol.at(ck_tile::number<1>{}));
EXPECT_TRUE(pass);
}
};

16
test/ck_tile/gemm_weight_preshuffle/test_gemm_pipeline_wp.cpp Normal file
View File

@@ -0,0 +1,16 @@
#include "test_gemm_pipeline_kernel_types.hpp"
#include "test_gemm_pipeline_util.hpp"
#include "gtest/gtest.h"
template <typename T>
class TestCkTileGemmPipelineWeightPreshuffle : public TestCkTileGemmPipeline<T>
{
};
#define TEST_SUITE_NAME TestCkTileGemmPipelineWeightPreshuffle
TYPED_TEST_SUITE(TestCkTileGemmPipelineWeightPreshuffle, KernelTypesWeightPreshuffle);
#include "test_gemm_pipeline_ut_cases.inc"
#undef TEST_SUITE_NAME