Skip to content

GitLab

Commit a781d078 authored by Qianfeng Zhang's avatar Qianfeng Zhang
Browse files

Merge branch 'develop' into bnorm_bwd_pr

parents fd76c787 4c4c7328
Hide whitespace changes
Inline Side-by-side
Showing with 535 additions and 124 deletions
profiler/src/profile_softmax.cpp
View file @ a781d078
......@@ -8,14 +8,10 @@
#include "profiler/include/profile_softmax_impl.hpp"
using ck::index_t;
using ck::profiler::NormDataType;
using ck::profiler::NormType;
using ck::profiler::SoftmaxDataType;
struct ArgParser
{
std::unordered_map<std::string, NormType> norm_dict = {{"batchnorm", NormType::BATCHNORM},
{"softmax", NormType::SOFTMAX}};
std::unordered_map<std::string, std::vector<int>> long_opts = {
{"length", {}}, {"stride", {}}, {"reduce", {}}, {"alpha", {}}, {"beta", {}}};
......@@ -50,7 +46,7 @@ struct ArgParser
void print_help()
{
std::cout << "arg1: tensor operation (batchnorm/softmax)\n"
std::cout << "arg1: tensor operation (softmax)\n"
<< "arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n"
<< "arg3: verification (0: no; 1: yes)\n"
<< "arg4: initialization (0: no init; 1: integer value; 2: decimal value)\n"
......@@ -64,7 +60,7 @@ void print_help()
<< std::endl;
}
int profile_normalization(int argc, char* argv[])
int profile_softmax(int argc, char* argv[])
{
if(argc <= 2)
{
......@@ -75,12 +71,11 @@ int profile_normalization(int argc, char* argv[])
ArgParser arg_parser;
// short unnamed options
const NormType norm_type = arg_parser.norm_dict[argv[1]];
const NormDataType data_type = static_cast<NormDataType>(std::stoi(argv[2]));
const bool do_verification = std::stoi(argv[3]);
const int init_method = std::stoi(argv[4]);
const bool do_log = std::stoi(argv[5]);
const bool time_kernel = std::stoi(argv[6]);
const SoftmaxDataType data_type = static_cast<SoftmaxDataType>(std::stoi(argv[2]));
const bool do_verification = std::stoi(argv[3]);
const int init_method = std::stoi(argv[4]);
const bool do_log = std::stoi(argv[5]);
const bool time_kernel = std::stoi(argv[6]);
// parse the long options
arg_parser(argc, argv);
......@@ -91,9 +86,10 @@ int profile_normalization(int argc, char* argv[])
arg_parser.long_opts["alpha"].empty() ? 1 : arg_parser.long_opts["alpha"][0];
const index_t beta = arg_parser.long_opts["beta"].empty() ? 0 : arg_parser.long_opts["beta"][0];
// Rank 3
if(length.size() == 3)
{
if(data_type == NormDataType::F16_F16)
if(data_type == SoftmaxDataType::F16_F16)
{
ck::profiler::profile_softmax_impl<ck::half_t, float, ck::half_t, 3>(do_verification,
init_method,
......@@ -103,10 +99,9 @@ int profile_normalization(int argc, char* argv[])
stride,
reduce,
float(alpha),
float(beta),
norm_type);
float(beta));
}
else if(data_type == NormDataType::F32_F32)
else if(data_type == SoftmaxDataType::F32_F32)
{
ck::profiler::profile_softmax_impl<float, float, float, 3>(do_verification,
init_method,
......@@ -116,17 +111,17 @@ int profile_normalization(int argc, char* argv[])
stride,
reduce,
float(alpha),
float(beta),
norm_type);
float(beta));
}
else
{
throw std::runtime_error("not implemented yet");
}
}
// Rank 4
else if(length.size() == 4)
{
if(data_type == NormDataType::F16_F16)
if(data_type == SoftmaxDataType::F16_F16)
{
ck::profiler::profile_softmax_impl<ck::half_t, float, ck::half_t, 4>(do_verification,
init_method,
......@@ -136,10 +131,9 @@ int profile_normalization(int argc, char* argv[])
stride,
reduce,
float(alpha),
float(beta),
norm_type);
float(beta));
}
else if(data_type == NormDataType::F32_F32)
else if(data_type == SoftmaxDataType::F32_F32)
{
ck::profiler::profile_softmax_impl<float, float, float, 4>(do_verification,
init_method,
......@@ -149,8 +143,7 @@ int profile_normalization(int argc, char* argv[])
stride,
reduce,
float(alpha),
float(beta),
norm_type);
float(beta));
}
else
{
......
profiler/src/profiler.cpp
View file @ a781d078
......@@ -18,9 +18,9 @@ int profile_conv_fwd(int, char*[]);
int profile_conv_fwd_bias_relu(int, char*[]);
int profile_conv_fwd_bias_relu_add(int, char*[]);
int profile_conv_bwd_data(int, char*[]);
int profile_conv_bwd_weight(int, char*[]);
int profile_grouped_conv_fwd(int, char*[]);
int profile_normalization(int, char*[]);
int profile_grouped_conv_bwd_weight(int, char*[]);
int profile_softmax(int, char*[]);
int profile_layernorm(int, char*[]);
int profile_groupnorm(int, char*[]);
int profile_reduce(int, char*[]);
......@@ -43,8 +43,9 @@ static void print_helper_message()
" conv_fwd_bias_relu: ForwardConvolution+Bias+ReLU\n"
" conv_fwd_bias_relu_add: ForwardConvolution+Bias+ReLU+Add\n"
" conv_bwd_data: Convolution Backward Data\n"
" conv_bwd_weight: Convolution Backward Weight\n"
" grouped_conv_fwd: Grouped Convolution Forward\n"
" grouped_conv_bwd_weight: Grouped Convolution Backward Weight\n"
" softmax: Softmax\n"
" reduce: Reduce\n");
// clang-format on
}
......@@ -117,21 +118,21 @@ int main(int argc, char* argv[])
{
return profile_conv_bwd_data(argc, argv);
}
else if(strcmp(argv[1], "conv_bwd_weight") == 0)
{
return profile_conv_bwd_weight(argc, argv);
}
else if(strcmp(argv[1], "grouped_conv_fwd") == 0)
{
return profile_grouped_conv_fwd(argc, argv);
}
else if(strcmp(argv[1], "conv_bwd_weight") == 0)
{
return profile_grouped_conv_bwd_weight(argc, argv);
}
else if(strcmp(argv[1], "reduce") == 0)
{
return profile_reduce(argc, argv);
}
else if(strcmp(argv[1], "batchnorm") == 0 || strcmp(argv[1], "softmax") == 0)
else if(strcmp(argv[1], "softmax") == 0)
{
return profile_normalization(argc, argv);
return profile_softmax(argc, argv);
}
else if(strcmp(argv[1], "layernorm") == 0)
{
......
script/cmake-ck-dev.sh
View file @ a781d078
......@@ -11,7 +11,7 @@ cmake
-D CMAKE_CXX_FLAGS="-O3 -ftemplate-backtrace-limit=0 -gline-tables-only -save-temps=$PWD" \
-D CMAKE_BUILD_TYPE=Release \
-D BUILD_DEV=ON \
-D GPU_TARGETS=gfx908;gfx90a \
-D GPU_TARGETS="gfx908;gfx90a" \
-D CMAKE_VERBOSE_MAKEFILE:BOOL=ON \
-D USE_BITINT_EXTENSION_INT4=OFF \
${MY_PROJECT_SOURCE}
......
script/cmake-ck-release.sh
View file @ a781d078
......@@ -11,7 +11,7 @@ cmake
-D CMAKE_CXX_FLAGS="-O3" \
-D CMAKE_BUILD_TYPE=Release \
-D BUILD_DEV=OFF \
-D GPU_TARGETS=gfx908;gfx90a \
-D GPU_TARGETS="gfx908;gfx90a" \
-D CMAKE_VERBOSE_MAKEFILE:BOOL=ON \
-D USE_BITINT_EXTENSION_INT4=OFF \
${MY_PROJECT_SOURCE}
......
test/CMakeLists.txt
View file @ a781d078
......@@ -26,7 +26,7 @@ function(add_gtest_executable TEST_NAME)
# suppress gtest warnings
target_compile_options(${TEST_NAME} PRIVATE -Wno-global-constructors -Wno-undef)
target_link_libraries(${TEST_NAME} PRIVATE gtest_main)
gtest_discover_tests(${TEST_NAME})
add_test(NAME ${TEST_NAME} COMMAND $<TARGET_FILE:${TEST_NAME}> )
rocm_install(TARGETS ${TEST_NAME} COMPONENT tests)
endfunction(add_gtest_executable TEST_NAME)
......@@ -45,10 +45,11 @@ add_subdirectory(batched_gemm_softmax_gemm_permute)
add_subdirectory(grouped_gemm)
add_subdirectory(reduce)
add_subdirectory(convnd_fwd)
add_subdirectory(convnd_bwd_weight)
add_subdirectory(convnd_bwd_data)
add_subdirectory(grouped_convnd_fwd)
add_subdirectory(grouped_convnd_bwd_weight)
add_subdirectory(block_to_ctile_map)
add_subdirectory(softmax)
add_subdirectory(normalization)
add_subdirectory(data_type)
add_subdirectory(elementwise_normalization)
test/batched_gemm_softmax_gemm_permute/CMakeLists.txt
View file @ a781d078
add_custom_target(test_batched_gemm_softmax_gemm_permute)
add_gtest_executable(test_batched_gemm_softmax_gemm_permute_fp16 test_batched_gemm_softmax_gemm_permute_fp16.cpp)
add_gtest_executable(test_batched_gemm_softmax_gemm_permute_bf16 test_batched_gemm_softmax_gemm_permute_bf16.cpp)
target_link_libraries(test_batched_gemm_softmax_gemm_permute_fp16 PRIVATE utility device_batched_gemm_softmax_gemm_permute_instance)
add_dependencies(test_batched_gemm_softmax_gemm_permute test_batched_gemm_softmax_gemm_permute_fp16)
\ No newline at end of file
target_link_libraries(test_batched_gemm_softmax_gemm_permute_bf16 PRIVATE utility device_batched_gemm_softmax_gemm_permute_instance)
add_dependencies(test_batched_gemm_softmax_gemm_permute test_batched_gemm_softmax_gemm_permute_fp16)
add_dependencies(test_batched_gemm_softmax_gemm_permute test_batched_gemm_softmax_gemm_permute_bf16)
\ No newline at end of file
test/batched_gemm_softmax_gemm_permute/test_batched_gemm_softmax_gemm_permute_bf16.cpp 0 → 100644
View file @ a781d078
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "gtest/gtest.h"
#include "test_batched_gemm_softmax_gemm_permute_util.hpp"
template <typename Tuple>
class TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16
: public TestBatchedGemmMaskingScaleSoftmaxGemmPermute<Tuple>
{
};
using I1_t = ck::Number<1>;
using I2_t = ck::Number<2>;
using MaskDisabled_t =
ck::integral_constant<MaskingSpecialization, MaskingSpecialization::MaskDisabled>;
using MaskOutUpperTriangle_t =
ck::integral_constant<MaskingSpecialization, MaskingSpecialization::MaskOutUpperTriangle>;
// clang-format off
using KernelTypes = ::testing::Types<
std::tuple<I2_t, I1_t, I1_t, I1_t, I1_t, BF16, BF16, BF16, BF16, ck::Tuple<>, ck::Tuple<>, MaskDisabled_t>,
std::tuple<I2_t, I1_t, I1_t, I1_t, I1_t, BF16, BF16, BF16, BF16, ck::Tuple<>, ck::Tuple<>, MaskOutUpperTriangle_t>
>;
// clang-format on
TYPED_TEST_SUITE(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, KernelTypes);
TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, Test_BF16) { this->Run(); }
TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, Test_FPBF_PadM)
{
this->lengths_ = std::vector<std::vector<int>>{
{136, 128, 32, 128, 2, 3},
};
this->Run();
}
TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, Test_BF16_PadN)
{
this->lengths_ = std::vector<std::vector<int>>{
{128, 136, 32, 128, 3, 2},
};
this->Run();
}
TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, Test_BF16_PadK)
{
this->lengths_ = std::vector<std::vector<int>>{
{128, 128, 40, 128, 2, 4},
{128, 128, 136, 128, 4, 2},
};
this->Run();
}
TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, Test_BF16_PadO)
{
this->lengths_ = std::vector<std::vector<int>>{
{128, 128, 32, 136, 1, 3},
};
this->Run();
}
TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, Test_BF16_OddM)
{
this->lengths_ = std::vector<std::vector<int>>{
{129, 128, 32, 128, 2, 3},
};
this->Run();
}
TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, Test_BF16_OddN)
{
this->lengths_ = std::vector<std::vector<int>>{
{128, 129, 32, 128, 4, 3},
};
this->Run();
}
TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, Test_BF16_OddK)
{
this->lengths_ = std::vector<std::vector<int>>{
{128, 128, 33, 128, 2, 3},
{128, 128, 129, 128, 2, 3},
};
this->Run();
}
// If kernel B1Layout is RowMajor, expect not to support odd O size
TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, Test_BF16_OddO)
{
this->lengths_ = std::vector<std::vector<int>>{
{128, 128, 32, 129, 2, 3},
};
this->Run();
}
TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, DISABLED_Bench_BF16_IrregularK)
{
this->lengths_ = std::vector<std::vector<int>>{{256, 256, 160, 160, 1, 16},
{256, 64, 160, 64, 1, 16},
{1024, 1024, 80, 80, 1, 16},
{1024, 64, 80, 64, 1, 16},
{4096, 4096, 40, 40, 1, 16},
{4096, 64, 40, 64, 1, 16}};
this->bench_ = true;
this->verify_ = false;
this->Run();
}
TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, DISABLED_Bench_BF16)
{
this->lengths_ = std::vector<std::vector<int>>{
{256, 256, 64, 64, 48, 16},
{256, 256, 128, 128, 48, 16},
{512, 512, 64, 64, 48, 16},
{512, 512, 128, 128, 48, 16},
{1024, 1024, 64, 64, 48, 16},
{1024, 1024, 128, 128, 48, 16},
{2048, 2048, 64, 64, 48, 16},
{2048, 2048, 128, 128, 48, 16},
{4096, 4096, 64, 64, 48, 16},
{4096, 4096, 128, 128, 48, 16},
};
this->bench_ = true;
this->verify_ = false;
this->Run();
}
using ck::tensor_operation::device::GemmSpecialization;
TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteInterface, GemmSpecializationSizeMatch)
{
int P = 120; // requires padding
int Q = 128; // do not require padding
// IsSupported(M, N, K, O)
// clang-format off
EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::Default>{}.IsSupported(Q, Q, Q, Q));
EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MPadding>{}.IsSupported(P, Q, Q, Q));
EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::NPadding>{}.IsSupported(Q, P, Q, Q));
EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::KPadding>{}.IsSupported(Q, Q, P, Q));
EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MNPadding>{}.IsSupported(P, P, Q, Q));
EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MKPadding>{}.IsSupported(P, Q, P, Q));
EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::NKPadding>{}.IsSupported(Q, P, P, Q));
EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MNKPadding>{}.IsSupported(P, P, P, Q));
EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::OPadding>{}.IsSupported(Q, Q, Q, P));
EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MOPadding>{}.IsSupported(P, Q, Q, P));
EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::NOPadding>{}.IsSupported(Q, P, Q, P));
EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::KOPadding>{}.IsSupported(Q, Q, P, P));
EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MNOPadding>{}.IsSupported(P, P, Q, P));
EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MKOPadding>{}.IsSupported(P, Q, P, P));
EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::NKOPadding>{}.IsSupported(Q, P, P, P));
EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(P, P, P, P));
// clang-format on
}
TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteInterface, GemmSpecializationSizeMismatch)
{
// IsSupported(M, N, K, O)
// clang-format off
EXPECT_FALSE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::Default>{}.IsSupported(128, 128, 120, 128));
EXPECT_FALSE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MNKPadding>{}.IsSupported(128, 128, 128, 120));
// Kernel can't support odd K size because SrcVectorDim == KDim and must satisfy SizeKRaw % ABSrcScalarPerVector == 0
EXPECT_FALSE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 129, 128));
EXPECT_FALSE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 130, 128));
// Kernel can't support odd O size because SrcVectorDim == ODim and must satisfy SizeORaw % B1SrcScalarPerVector == 0
EXPECT_FALSE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 128, 129));
// clang-format on
}
TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, AdhocTest)
{
this->lengths_ = std::vector<std::vector<int>>{
{49, 49, 64, 64, 4, 6},
{64, 49, 64, 64, 4, 6},
{1020, 1020, 64, 128, 4, 6},
{576, 576, 64, 64, 4, 6},
};
this->Run();
}
test/batched_gemm_softmax_gemm_permute/test_batched_gemm_softmax_gemm_permute_util.hpp
View file @ a781d078
......@@ -16,7 +16,8 @@ using ck::tensor_operation::device::TensorSpecialization;
template <ck::index_t N>
using I = ck::Number<N>;
using F16 = ck::half_t;
using F16 = ck::half_t;
using BF16 = ck::bhalf_t;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
......@@ -63,7 +64,7 @@ struct TestBatchedGemmMaskingScaleSoftmaxGemmPermute : public ::testing::Test
ck::Tuple<>,
ck::Tuple<>,
MaskingType::value>(
verify_, 1, false, bench_, M, N, K, O, G0, G1);
verify_, 2, false, bench_, M, N, K, O, G0, G1);
EXPECT_TRUE(pass);
}
......@@ -224,3 +225,144 @@ struct DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128
return gemm.IsSupportedArgument(argument);
}
};
template <GemmSpecialization GemmSpec>
struct DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128
{
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using Scale = ck::tensor_operation::element_wise::Scale;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using ADataType = BF16;
using B0DataType = BF16;
using B1DataType = BF16;
using AccDataType = float;
using CShuffleDataType = BF16;
using CDataType = BF16;
using AElementOp = PassThrough;
using B0ElementOp = PassThrough;
using Acc0ElementOp = Scale;
using B1ElementOp = PassThrough;
using CElementOp = PassThrough;
// static constexpr auto GemmSpec = std::tuple_element_t<0, Tuple>::value;
using DeviceGemmGemmInstance =
ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<
2,
1,
1,
1,
1,
ADataType,
B0DataType,
B1DataType,
CDataType,
ck::Tuple<>,
ck::Tuple<>,
AccDataType,
CShuffleDataType,
AElementOp,
B0ElementOp,
Acc0ElementOp,
B1ElementOp,
CElementOp,
GemmSpec,
TensorSpecialization::Default, // ATensorSpec
TensorSpecialization::Default, // B0TensorSpec
TensorSpecialization::Default, // B1TensorSpec
TensorSpecialization::Default, // CTensorSpec
1,
256,
128, // MPerBlock
128, // NPerBlock
32, // KPerBlock
128, // Gemm1NPerBlock
32, // Gemm1KPerBlock
8, // AK1
8, // BK1
2, // B1K1
32, // MPerXDL
32, // NPerXDL
1, // MXdlPerWave
4, // NXdlPerWave
4, // Gemm1NXdlPerWave
S<4, 64, 1>, // ABlockTransfer
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
true,
S<4, 64, 1>, // BBlockTransfer
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
true,
S<8, 32, 1>, // B1BlockTransfer
S<0, 2, 1>,
S<0, 2, 1>,
1,
4,
2,
false,
1, // CShuffleMXdlPerWavePerShuffle
2, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
8, // CShuffleBlockTransferScalarPerVector_NPerBlock
MaskingSpecialization::MaskOutUpperTriangle>; // MaskOutUpperTriangle
bool IsSupported(int M, int N, int K, int O)
{
const int G0 = 1, G1 = 1;
// A layout [G0, M, G1, K]
std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, G1, M, K};
std::vector<ck::index_t> a_gs_ms_ks_strides{M * G1 * K, K, G1 * K, 1};
// B0 layout [G0, N, G1, K]
std::vector<ck::index_t> b0_gs_ns_ks_lengths{G0, G1, N, K};
std::vector<ck::index_t> b0_gs_ns_ks_strides{N * G1 * K, K, G1 * K, 1};
// B1 layout [G0, N, G1, O]
std::vector<ck::index_t> b1_gs_os_ns_lengths{G0, G1, O, N};
std::vector<ck::index_t> b1_gs_os_ns_strides{N * G1 * O, O, 1, G1 * O};
// C layout [G0, M, G1, O]
std::vector<ck::index_t> c_gs_ms_os_lengths{G0, G1, M, O};
std::vector<ck::index_t> c_gs_ms_os_strides{M * G1 * O, O, G1 * O, 1};
auto gemm = DeviceGemmGemmInstance{};
auto invoker = gemm.MakeInvoker();
auto argument = gemm.MakeArgument(static_cast<ADataType*>(nullptr),
static_cast<B0DataType*>(nullptr),
static_cast<B1DataType*>(nullptr),
static_cast<CDataType*>(nullptr),
{}, // p_acc0_biases
{}, // p_acc1_biases
a_gs_ms_ks_lengths,
a_gs_ms_ks_strides,
b0_gs_ns_ks_lengths,
b0_gs_ns_ks_strides,
b1_gs_os_ns_lengths,
b1_gs_os_ns_strides,
c_gs_ms_os_lengths,
c_gs_ms_os_strides,
{}, // acc0_biases_gs_ms_ns_lengths
{}, // acc0_biases_gs_ms_ns_strides
{}, // acc1_biases_gs_ms_os_lengths
{}, // acc1_biases_gs_ms_os_strides
PassThrough{}, // a_element_op
PassThrough{}, // b0_element_op
Scale{1.f}, // acc0_element_op
PassThrough{}, // b1_element_op
PassThrough{}); // c_element_op
return gemm.IsSupportedArgument(argument);
}
};
test/convnd_bwd_weight/CMakeLists.txt deleted 100644 → 0
View file @ fd76c787
add_gtest_executable(test_convnd_bwd_weight convnd_bwd_weight.cpp)
target_link_libraries(test_convnd_bwd_weight PRIVATE utility device_conv1d_bwd_weight_instance device_conv2d_bwd_weight_instance device_conv3d_bwd_weight_instance)
test/elementwise_normalization/CMakeLists.txt 0 → 100644
View file @ a781d078
add_custom_target(test_elementwise_normalization)
add_gtest_executable(test_elementwise_layernorm_fp16 test_elementwise_layernorm_fp16.cpp)
target_link_libraries(test_elementwise_layernorm_fp16 PRIVATE utility device_elementwise_normalization_instance)
add_dependencies(test_elementwise_normalization test_elementwise_layernorm_fp16)
test/elementwise_normalization/test_elementwise_layernorm_fp16.cpp 0 → 100644
View file @ a781d078
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "gtest/gtest.h"
#include "profiler/include/profile_elementwise_layernorm_impl.hpp"
using F16 = ck::half_t;
using F32 = float;
using ck::index_t;
template <typename Tuple>
class TestElementwiseLayernorm : public ::testing::Test
{
protected:
using ADataType = std::tuple_element_t<0, Tuple>;
using BDataType = std::tuple_element_t<1, Tuple>;
using GammaDataType = std::tuple_element_t<2, Tuple>;
using BetaDataType = std::tuple_element_t<3, Tuple>;
using AccDataType = std::tuple_element_t<4, Tuple>;
using YDataType = std::tuple_element_t<5, Tuple>;
void Run()
{
// M, N
std::vector<std::vector<ck::index_t>> lengths = {
{1, 1}, {25, 16}, {39, 777}, {100, 200}, {1024, 1024}, {48 * 256, 2048}};
for(auto length : lengths)
{
bool success = ck::profiler::profile_elementwise_layernorm_impl<ADataType,
BDataType,
GammaDataType,
BetaDataType,
AccDataType,
YDataType>(
true, 2, false, false, length);
EXPECT_TRUE(success);
}
}
};
using KernelTypes = ::testing::Types<
// ADataType, BDataType, GammaDataType, BetaDataType, AccDataType, YDataType>
std::tuple<F16, F16, F16, F16, F32, F16>>;
TYPED_TEST_SUITE(TestElementwiseLayernorm, KernelTypes);
TYPED_TEST(TestElementwiseLayernorm, Test_FP16) { this->Run(); }
test/gemm/gemm_util.hpp
View file @ a781d078
......@@ -9,6 +9,7 @@
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
namespace ck {
......@@ -128,15 +129,15 @@ struct TestGemm
{
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
using namespace ck::literals;
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({stride, 1}));
return HostTensorDescriptor({row, col}, {stride, 1_uz});
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({1, stride}));
return HostTensorDescriptor({row, col}, {1_uz, stride});
}
};
......@@ -229,27 +230,27 @@ struct TestGemm
bool res = false;
if(std::is_same<CDataType, float>::value)
{
res = ck::utils::check_err(c_device.mData, c_host.mData);
res = ck::utils::check_err(c_device, c_host);
std::cout << (res ? "SUCCESS" : "FAILURE") << std::endl;
}
else if(std::is_same<CDataType, ck::half_t>::value)
{
res = ck::utils::check_err(c_device.mData, c_host.mData);
res = ck::utils::check_err(c_device, c_host);
std::cout << (res ? "SUCCESS" : "FAILURE") << std::endl;
}
else if(std::is_same<CDataType, ck::bhalf_t>::value)
{
res = ck::utils::check_err(c_device.mData, c_host.mData);
res = ck::utils::check_err(c_device, c_host);
std::cout << (res ? "SUCCESS" : "FAILURE") << std::endl;
}
else if(std::is_same<CDataType, int8_t>::value)
{
res = ck::utils::check_err(c_device.mData, c_host.mData);
res = ck::utils::check_err(c_device, c_host);
std::cout << (res ? "SUCCESS" : "FAILURE") << std::endl;
}
else if(std::is_same<CDataType, double>::value)
{
res = ck::utils::check_err(c_device.mData, c_host.mData);
res = ck::utils::check_err(c_device, c_host);
std::cout << (res ? "SUCCESS" : "FAILURE") << std::endl;
}
......
test/gemm_split_k/gemm_split_k.cpp
View file @ a781d078
......@@ -16,6 +16,7 @@
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "ck/library/utility/host_gemm.hpp"
......@@ -93,15 +94,15 @@ int test_gemm(const gemmArgs& args)
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, bool row_major) {
using namespace ck::literals;
if(row_major)
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({stride, 1}));
return HostTensorDescriptor({row, col}, {stride, 1_uz});
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({1, stride}));
return HostTensorDescriptor({row, col}, {1_uz, stride});
}
};
......
test/grouped_convnd_bwd_weight/CMakeLists.txt 0 → 100644
View file @ a781d078
add_gtest_executable(test_grouped_convnd_bwd_weight grouped_convnd_bwd_weight.cpp)
target_link_libraries(test_grouped_convnd_bwd_weight PRIVATE utility device_grouped_conv1d_bwd_weight_instance device_grouped_conv2d_bwd_weight_instance device_grouped_conv3d_bwd_weight_instance)
test/convnd_bwd_weight/convnd_bwd_weight.cpp → test/grouped_convnd_bwd_weight/grouped_convnd_bwd_weight.cpp
View file @ a781d078
......@@ -4,14 +4,15 @@
#include <cstdlib>
#include <iostream>
#include <initializer_list>
#include <vector>
#include <tuple>
#include <vector>
#include <gtest/gtest.h>
#include "profiler/include/profile_conv_bwd_weight_impl.hpp"
#include "profiler/include/profile_grouped_conv_bwd_weight_impl.hpp"
template <typename Tuple>
class TestConvndBwdWeight : public ::testing::Test
class TestGroupedConvndBwdWeight : public ::testing::Test
{
protected:
using DataType = std::tuple_element_t<0, Tuple>;
......@@ -25,20 +26,20 @@ class TestConvndBwdWeight : public ::testing::Test
{
bool pass;
EXPECT_FALSE(conv_params.empty());
pass = ck::profiler::profile_conv_bwd_weight_impl<
pass = ck::profiler::profile_grouped_conv_bwd_weight_impl<
NDimSpatial,
ck::tuple_element_t<NDimSpatial - 1,
ck::Tuple<ck::tensor_layout::convolution::NWC,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::NDHWC>>,
ck::Tuple<ck::tensor_layout::convolution::GNWC,
ck::tensor_layout::convolution::GNHWC,
ck::tensor_layout::convolution::GNDHWC>>,
ck::tuple_element_t<NDimSpatial - 1,
ck::Tuple<ck::tensor_layout::convolution::KXC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::KZYXC>>,
ck::Tuple<ck::tensor_layout::convolution::GKXC,
ck::tensor_layout::convolution::GKYXC,
ck::tensor_layout::convolution::GKZYXC>>,
ck::tuple_element_t<NDimSpatial - 1,
ck::Tuple<ck::tensor_layout::convolution::NWK,
ck::tensor_layout::convolution::NHWK,
ck::tensor_layout::convolution::NDHWK>>,
ck::Tuple<ck::tensor_layout::convolution::GNWK,
ck::tensor_layout::convolution::GNHWK,
ck::tensor_layout::convolution::GNDHWK>>,
DataType,
DataType,
DataType>(true, // do_verification
......@@ -54,37 +55,37 @@ class TestConvndBwdWeight : public ::testing::Test
using KernelTypes =
::testing::Types<std::tuple<float>, std::tuple<ck::half_t>, std::tuple<ck::bhalf_t>>;
TYPED_TEST_SUITE(TestConvndBwdWeight, KernelTypes);
TYPED_TEST_SUITE(TestGroupedConvndBwdWeight, KernelTypes);
TYPED_TEST(TestConvndBwdWeight, Test1D)
TYPED_TEST(TestGroupedConvndBwdWeight, Test1D)
{
this->conv_params.clear();
this->conv_params.push_back({1, 1, 128, 128, 256, {1}, {14}, {2}, {1}, {0}, {0}});
this->conv_params.push_back({1, 1, 128, 128, 256, {3}, {28}, {1}, {1}, {1}, {1}});
this->conv_params.push_back({1, 1, 128, 128, 256, {1}, {3}, {1}, {1}, {0}, {0}});
this->conv_params.push_back({1, 4, 128, 128, 256, {1}, {14}, {2}, {1}, {0}, {0}});
this->conv_params.push_back({1, 4, 128, 128, 256, {3}, {28}, {1}, {1}, {1}, {1}});
this->conv_params.push_back({1, 4, 128, 128, 256, {1}, {3}, {1}, {1}, {0}, {0}});
this->template Run<1>();
}
TYPED_TEST(TestConvndBwdWeight, Test2D)
TYPED_TEST(TestGroupedConvndBwdWeight, Test2D)
{
this->conv_params.clear();
this->conv_params.push_back(
{2, 1, 128, 128, 256, {1, 1}, {7, 7}, {2, 2}, {1, 1}, {0, 0}, {0, 0}});
{2, 4, 128, 128, 256, {1, 1}, {7, 7}, {2, 2}, {1, 1}, {0, 0}, {0, 0}});
this->conv_params.push_back(
{2, 1, 32, 128, 256, {3, 3}, {14, 14}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
{2, 4, 32, 128, 256, {3, 3}, {14, 14}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
this->conv_params.push_back(
{2, 1, 128, 128, 256, {1, 1}, {3, 3}, {1, 1}, {1, 1}, {0, 0}, {0, 0}});
{2, 4, 128, 128, 256, {1, 1}, {3, 3}, {1, 1}, {1, 1}, {0, 0}, {0, 0}});
this->template Run<2>();
}
TYPED_TEST(TestConvndBwdWeight, Test3D)
TYPED_TEST(TestGroupedConvndBwdWeight, Test3D)
{
this->conv_params.clear();
this->conv_params.push_back(
{3, 1, 128, 128, 256, {1, 1, 1}, {7, 7, 7}, {2, 2, 2}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
{3, 4, 128, 128, 256, {1, 1, 1}, {7, 7, 7}, {2, 2, 2}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
this->conv_params.push_back(
{3, 1, 32, 128, 256, {3, 3, 3}, {14, 14, 3}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
{3, 4, 32, 128, 256, {3, 3, 3}, {14, 14, 3}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
this->conv_params.push_back(
{3, 1, 128, 128, 256, {1, 1, 1}, {3, 3, 3}, {1, 1, 1}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
{3, 4, 128, 128, 256, {1, 1, 1}, {3, 3, 3}, {1, 1, 1}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
this->template Run<3>();
}
test/normalization/CMakeLists.txt
View file @ a781d078
......@@ -3,10 +3,11 @@ add_custom_target(test_layernorm)
add_gtest_executable(test_layernorm2d_fp32 test_layernorm2d_fp32.cpp)
add_gtest_executable(test_layernorm2d_fp16 test_layernorm2d_fp16.cpp)
add_gtest_executable(test_groupnorm_fp16 test_groupnorm_fp16.cpp)
add_gtest_executable(test_groupnorm_fp32 test_groupnorm_fp32.cpp)
add_gtest_executable(test_groupnorm_fp32 test_groupnorm_fp32.cpp)
target_link_libraries(test_layernorm2d_fp32 PRIVATE utility)
target_link_libraries(test_layernorm2d_fp16 PRIVATE utility)
target_link_libraries(test_layernorm2d_fp32 PRIVATE utility device_normalization_instance)
target_link_libraries(test_layernorm2d_fp16 PRIVATE utility device_normalization_instance)
target_link_libraries(test_groupnorm_fp16 PRIVATE utility device_normalization_instance)
target_link_libraries(test_groupnorm_fp32 PRIVATE utility device_normalization_instance)
......@@ -14,4 +15,3 @@ add_dependencies(test_layernorm test_layernorm2d_fp32)
add_dependencies(test_layernorm test_layernorm2d_fp16)
add_dependencies(test_layernorm test_groupnorm_fp16)
add_dependencies(test_layernorm test_groupnorm_fp32)
test/normalization/test_groupnorm_fp16.cpp
View file @ a781d078
......@@ -20,7 +20,7 @@ class TestGroupnorm : public ::testing::Test
void Run()
{
// N, H, W, G, C
// [N, H, W, G, C], reduce H, W, C
std::vector<std::vector<ck::index_t>> lengths = {{1, 1, 1, 1, 1},
{1, 2, 3, 4, 5},
{256, 9, 9, 9, 9},
......
test/normalization/test_groupnorm_fp32.cpp
View file @ a781d078
......@@ -20,7 +20,7 @@ class TestGroupnorm : public ::testing::Test
void Run()
{
// N, H, W, G, C
// [N, H, W, G, C], reduce H, W, C
std::vector<std::vector<ck::index_t>> lengths = {{1, 1, 1, 1, 1},
{1, 2, 3, 4, 5},
{256, 9, 9, 9, 9},
......
test/normalization/test_layernorm2d_fp16.cpp
View file @ a781d078
......@@ -2,28 +2,44 @@
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "gtest/gtest.h"
#include "test_layernorm2d_util.hpp"
#include "profiler/include/profile_layernorm_impl.hpp"
template <ck::index_t N>
using I = ck::Number<N>;
using F16 = ck::half_t;
using F32 = float;
using ck::index_t;
template <typename Tuple>
class TestLayernorm2dFP16 : public ck::TestLayernorm2d<Tuple>
class TestLayernorm2d : public ::testing::Test
{
protected:
using XDataType = std::tuple_element_t<0, Tuple>;
using GammaDataType = std::tuple_element_t<1, Tuple>;
using BetaDataType = std::tuple_element_t<2, Tuple>;
using AccDataType = std::tuple_element_t<3, Tuple>;
using YDataType = std::tuple_element_t<4, Tuple>;
void Run()
{
// [N, D], reduce D
std::vector<std::vector<ck::index_t>> lengths = {
{4, 256}, {8, 511}, {9, 1032}, {4, 2048}, {1, 8192}, {4000, 2000}};
for(auto length : lengths)
{
bool success = ck::profiler::profile_layernorm_impl<XDataType,
GammaDataType,
BetaDataType,
AccDataType,
YDataType,
2>(true, 2, false, false, length);
EXPECT_TRUE(success);
}
}
};
// clang-format off
using KernelTypes = ::testing::Types<
// XDataType, GammaDataType, BetaDataType, AccDataType, YDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorDim , GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize>
std::tuple<ck::half_t, ck::half_t, ck::half_t, float, ck::half_t, I<2>, I<1>, I<256>, I<8>, I<32>, I<1>, I<8>, I<1>, I<8>, I<1>, I<8>, I<1>, I<8>, I<8>>,
std::tuple<ck::half_t, ck::half_t, ck::half_t, float, ck::half_t, I<2>, I<1>, I<256>, I<8>, I<32>, I<2>, I<8>, I<1>, I<8>, I<1>, I<8>, I<1>, I<8>, I<8>>,
std::tuple<ck::half_t, ck::half_t, ck::half_t, float, ck::half_t, I<2>, I<1>, I<256>, I<4>, I<64>, I<1>, I<8>, I<1>, I<8>, I<1>, I<8>, I<1>, I<8>, I<8>>,
std::tuple<ck::half_t, ck::half_t, ck::half_t, float, ck::half_t, I<2>, I<1>, I<256>, I<4>, I<64>, I<2>, I<8>, I<1>, I<8>, I<1>, I<8>, I<1>, I<8>, I<8>>,
std::tuple<ck::half_t, ck::half_t, ck::half_t, float, ck::half_t, I<2>, I<1>, I<256>, I<2>, I<128>, I<1>, I<8>, I<1>, I<8>, I<1>, I<8>, I<1>, I<8>, I<8>>,
std::tuple<ck::half_t, ck::half_t, ck::half_t, float, ck::half_t, I<2>, I<1>, I<256>, I<2>, I<128>, I<2>, I<8>, I<1>, I<8>, I<1>, I<8>, I<1>, I<8>, I<8>>,
std::tuple<ck::half_t, ck::half_t, ck::half_t, float, ck::half_t, I<2>, I<1>, I<256>, I<1>, I<256>, I<1>, I<8>, I<1>, I<8>, I<1>, I<8>, I<1>, I<8>, I<8>>,
std::tuple<ck::half_t, ck::half_t, ck::half_t, float, ck::half_t, I<2>, I<1>, I<256>, I<1>, I<256>, I<2>, I<8>, I<1>, I<8>, I<1>, I<8>, I<1>, I<8>, I<8>>
>;
// clang-format on
TYPED_TEST_SUITE(TestLayernorm2dFP16, KernelTypes);
TYPED_TEST(TestLayernorm2dFP16, Test_FP16) { this->Run(); }
// XDataType, GammaDataType, BetaDataType, AccDataType, YDataType>
std::tuple<F16, F16, F16, F32, F16>>;
TYPED_TEST_SUITE(TestLayernorm2d, KernelTypes);
TYPED_TEST(TestLayernorm2d, Test_FP16) { this->Run(); }
test/normalization/test_layernorm2d_fp32.cpp
View file @ a781d078
......@@ -2,28 +2,44 @@
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "gtest/gtest.h"
#include "test_layernorm2d_util.hpp"
#include "profiler/include/profile_layernorm_impl.hpp"
template <ck::index_t N>
using I = ck::Number<N>;
using F16 = ck::half_t;
using F32 = float;
using ck::index_t;
template <typename Tuple>
class TestLayernorm2dFP32 : public ck::TestLayernorm2d<Tuple>
class TestLayernorm2d : public ::testing::Test
{
protected:
using XDataType = std::tuple_element_t<0, Tuple>;
using GammaDataType = std::tuple_element_t<1, Tuple>;
using BetaDataType = std::tuple_element_t<2, Tuple>;
using AccDataType = std::tuple_element_t<3, Tuple>;
using YDataType = std::tuple_element_t<4, Tuple>;
void Run()
{
// [N, D], reduce D
std::vector<std::vector<ck::index_t>> lengths = {
{4, 256}, {8, 511}, {9, 1032}, {4, 2048}, {1, 8192}, {4000, 2000}};
for(auto length : lengths)
{
bool success = ck::profiler::profile_layernorm_impl<XDataType,
GammaDataType,
BetaDataType,
AccDataType,
YDataType,
2>(true, 2, false, false, length);
EXPECT_TRUE(success);
}
}
};
// clang-format off
using KernelTypes = ::testing::Types<
// XDataType, GammaDataType, BetaDataType, AccDataType, YDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize>
std::tuple<float, float, float, float, float, I<2>, I<1>, I<256>, I<8>, I<32>, I<1>, I<8>, I<1>, I<4>, I<1>, I<4>, I<1>, I<4>, I<4>>,
std::tuple<float, float, float, float, float, I<2>, I<1>, I<256>, I<8>, I<32>, I<2>, I<8>, I<1>, I<4>, I<1>, I<4>, I<1>, I<4>, I<4>>,
std::tuple<float, float, float, float, float, I<2>, I<1>, I<256>, I<4>, I<64>, I<1>, I<8>, I<1>, I<4>, I<1>, I<4>, I<1>, I<4>, I<4>>,
std::tuple<float, float, float, float, float, I<2>, I<1>, I<256>, I<4>, I<64>, I<2>, I<8>, I<1>, I<4>, I<1>, I<4>, I<1>, I<4>, I<4>>,
std::tuple<float, float, float, float, float, I<2>, I<1>, I<256>, I<2>, I<128>, I<1>, I<8>, I<1>, I<4>, I<1>, I<4>, I<1>, I<4>, I<4>>,
std::tuple<float, float, float, float, float, I<2>, I<1>, I<256>, I<2>, I<128>, I<2>, I<8>, I<1>, I<4>, I<1>, I<4>, I<1>, I<4>, I<4>>,
std::tuple<float, float, float, float, float, I<2>, I<1>, I<256>, I<1>, I<256>, I<1>, I<8>, I<1>, I<4>, I<1>, I<4>, I<1>, I<4>, I<4>>,
std::tuple<float, float, float, float, float, I<2>, I<1>, I<256>, I<1>, I<256>, I<2>, I<8>, I<1>, I<4>, I<1>, I<4>, I<1>, I<4>, I<4>>
>;
// clang-format on
TYPED_TEST_SUITE(TestLayernorm2dFP32, KernelTypes);
TYPED_TEST(TestLayernorm2dFP32, Test_FP32) { this->Run(); }
// XDataType, GammaDataType, BetaDataType, AccDataType, YDataType>
std::tuple<F32, F32, F32, F32, F32>>;
TYPED_TEST_SUITE(TestLayernorm2d, KernelTypes);
TYPED_TEST(TestLayernorm2d, Test_FP32) { this->Run(); }
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment