Skip to content

GitLab

Commit 2d92594a authored by Chao Liu's avatar Chao Liu
Browse files

Merge remote-tracking branch 'origin/develop' into clean_conv

parents ba6664b0 a11680cc
Hide whitespace changes
Inline Side-by-side
Showing with 397 additions and 26 deletions
script/run_full_performance_tests.sh 0 → 100755
View file @ 2d92594a
#!/bin/bash
#
# in order to run this script you'd first need to build the ckProfiler executable in ../build/bin/
# and make sure the following python packages are installed in your environment:
pip3 install --upgrade pip
pip3 install sqlalchemy pymysql pandas sshtunnel
# you would also need to set up some environment variables in order to
# post your new test results to the database and compare them to the baseline
# please contact Illia.Silin@amd.com for more details
#
# run the script as "./run_full_performance_tests.sh <tag for your test environment>
#get the test environment type:
export env_type=$1
echo 'Environment type ' $env_type
function print_log_header(){
rm -f $1;
git status | grep -e 'On branch' > $1;
echo -n 'Node name: ' >>$1; hostname >> $1;
#get GPU_arch and number of compute units from rocminfo
echo -n "GPU_arch: " >> $1; rocminfo | grep "Name:" | grep "gfx" >> $1;
rocminfo | grep "Compute Unit:" >> $1;
hipcc --version | grep -e 'HIP version' >> $1;
echo 'Environment type: ' $2 >>$1;
/opt/rocm/bin/amdclang++ --version | grep -e 'InstalledDir' >> $1;
}
#run gemm tests
export gemm_log="perf_gemm.log"
print_log_header $gemm_log $env_type
./profile_gemm.sh gemm 0 0 0 1 0 5 | tee -a $gemm_log
./profile_gemm.sh gemm 1 0 0 1 0 5 | tee -a $gemm_log
./profile_gemm.sh gemm 2 0 0 1 0 5 | tee -a $gemm_log
./profile_gemm.sh gemm 3 0 0 1 0 5 | tee -a $gemm_log
./profile_gemm.sh gemm 0 1 0 1 0 5 | tee -a $gemm_log
./profile_gemm.sh gemm 1 1 0 1 0 5 | tee -a $gemm_log
./profile_gemm.sh gemm 2 1 0 1 0 5 | tee -a $gemm_log
./profile_gemm.sh gemm 3 1 0 1 0 5 | tee -a $gemm_log
./profile_gemm.sh gemm 0 2 0 1 0 5 | tee -a $gemm_log
./profile_gemm.sh gemm 1 2 0 1 0 5 | tee -a $gemm_log
./profile_gemm.sh gemm 2 2 0 1 0 5 | tee -a $gemm_log
./profile_gemm.sh gemm 3 2 0 1 0 5 | tee -a $gemm_log
./profile_gemm.sh gemm 0 3 0 1 0 5 | tee -a $gemm_log
./profile_gemm.sh gemm 1 3 0 1 0 5 | tee -a $gemm_log
./profile_gemm.sh gemm 2 3 0 1 0 5 | tee -a $gemm_log
./profile_gemm.sh gemm 3 3 0 1 0 5 | tee -a $gemm_log
python3 process_perf_data.py $gemm_log
#run resnet50 tests
export resnet256_log="perf_resnet50_N256.log"
print_log_header $resnet256_log $env_type
./profile_resnet50.sh conv_fwd_bias_relu 1 1 1 1 0 2 0 1 256 | tee -a $resnet256_log
python3 process_perf_data.py $resnet256_log
export resnet4_log="perf_resnet50_N4.log"
print_log_header $resnet4_log $env_type
./profile_resnet50.sh conv_fwd_bias_relu 1 1 1 1 0 2 0 1 4 | tee -a $resnet4_log
python3 process_perf_data.py $resnet4_log
#run batched_gemm tests
export batched_gemm_log="perf_batched_gemm.log"
print_log_header $batched_gemm_log $env_type
./profile_batched_gemm.sh batched_gemm 0 0 0 2 0 5 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 0 1 0 2 0 5 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 0 2 0 2 0 5 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 0 3 0 2 0 5 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 1 0 0 2 0 5 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 1 1 0 2 0 5 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 1 2 0 2 0 5 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 1 3 0 2 0 5 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 2 0 0 2 0 5 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 2 1 0 2 0 5 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 2 2 0 2 0 5 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 2 3 0 2 0 5 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 3 0 0 2 0 5 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 3 1 0 2 0 5 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 3 2 0 2 0 5 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 3 3 0 2 0 5 | tee -a $batched_gemm_log
python3 process_perf_data.py $batched_gemm_log
#run grouped_gemm tests
export grouped_gemm_log="perf_grouped_gemm.log"
print_log_header $grouped_gemm_log $env_type
./profile_grouped_gemm.sh grouped_gemm 1 0 0 2 0 5 | tee -a $grouped_gemm_log
./profile_grouped_gemm.sh grouped_gemm 1 1 0 2 0 5 | tee -a $grouped_gemm_log
./profile_grouped_gemm.sh grouped_gemm 1 2 0 2 0 5 | tee -a $grouped_gemm_log
./profile_grouped_gemm.sh grouped_gemm 1 3 0 2 0 5 | tee -a $grouped_gemm_log
python3 process_perf_data.py $grouped_gemm_log
#run fwd_conv tests
export fwd_conv_log="perf_fwd_conv.log"
print_log_header $fwd_conv_log $env_type
./profile_conv.sh conv_fwd 0 1 0 2 0 5 2 256 | tee -a $fwd_conv_log
./profile_conv.sh conv_fwd 1 1 0 2 0 5 2 256 | tee -a $fwd_conv_log
./profile_conv.sh conv_fwd 2 1 0 2 0 5 2 256 | tee -a $fwd_conv_log
./profile_conv.sh conv_fwd 3 1 0 2 0 5 2 256 | tee -a $fwd_conv_log
python3 process_perf_data.py $fwd_conv_log
#run bwd_conv tests
export bwd_conv_log="perf_bwd_conv.log"
print_log_header $bwd_conv_log $env_type
./profile_conv.sh conv2d_bwd_data 0 1 1 1 0 2 0 5 128 | tee -a $bwd_conv_log
./profile_conv.sh conv2d_bwd_data 1 1 1 1 0 2 0 5 128 | tee -a $bwd_conv_log
./profile_conv.sh conv2d_bwd_data 2 1 1 1 0 2 0 5 128 | tee -a $bwd_conv_log
./profile_conv.sh conv2d_bwd_data 3 1 1 1 0 2 0 5 128 | tee -a $bwd_conv_log
python3 process_perf_data.py $bwd_conv_log
#run fusion tests
export fusion_log="perf_fusion.log"
print_log_header $fusion_log $env_type
./profile_gemm_bias_relu_add.sh gemm_bias_relu_add 1 0 0 2 0 5 | tee -a $fusion_log
./profile_gemm_bias_relu_add.sh gemm_bias_relu_add 1 1 0 2 0 5 | tee -a $fusion_log
./profile_gemm_bias_relu_add.sh gemm_bias_relu_add 1 2 0 2 0 5 | tee -a $fusion_log
./profile_gemm_bias_relu_add.sh gemm_bias_relu_add 1 3 0 2 0 5 | tee -a $fusion_log
python3 process_perf_data.py $fusion_log
#run reduction tests
export reduction_log="perf_reduction.log"
print_log_header $reduction_log $env_type
./profile_reduce_with_index.sh 0 2 10 --half | tee -a $reduction_log
./profile_reduce_no_index.sh 0 2 10 --half | tee -a $reduction_log
python3 process_perf_data.py $reduction_log
\ No newline at end of file
script/run_performance_tests.sh
View file @ 2d92594a
......@@ -10,17 +10,27 @@ pip3 install sqlalchemy pymysql pandas sshtunnel
# post your new test results to the database and compare them to the baseline
# please contact Illia.Silin@amd.com for more details
#
# run the script as "./run_performance_tests.sh <tag for your test environment>
#get the test environment type:
export env_type=$1
echo 'Environment type ' $env_type
function print_log_header(){
rm -f $1;
git status | grep -e 'On branch' > $1;
echo -n 'Node name: ' >>$1; hostname >> $1;
#get GPU_arch and number of compute units from rocminfo
echo -n "GPU_arch: " >> $1; rocminfo | grep "Name:" | grep "gfx" >> $1;
rocminfo | grep "Compute Unit:" >> $1;
hipcc --version | grep -e 'HIP version' >> $1;
echo 'Environment type: ' $2 >>$1;
/opt/rocm/bin/amdclang++ --version | grep -e 'InstalledDir' >> $1;
}
#run gemm tests
export gemm_log="perf_gemm.log"
rm -f $gemm_log
git status | grep -e 'On branch' > ${gemm_log}
echo -n 'Node name: ' >>${gemm_log}; hostname >> ${gemm_log}
#get GPU_arch and number of compute units from rocminfo
echo -n "GPU_arch: " >> ${gemm_log}; rocminfo | grep "Name:" | grep "gfx" >> ${gemm_log}
rocminfo | grep "Compute Unit:" >> ${gemm_log}
hipcc --version | grep -e 'HIP version' >> ${gemm_log}
/opt/rocm/bin/amdclang++ --version | grep -e 'InstalledDir' >> ${gemm_log}
./profile_gemm.sh gemm 0 0 0 1 0 5 | tee -a ${gemm_log}
print_log_header $gemm_log $env_type
./profile_gemm.sh gemm 0 0 0 1 0 5 | tee -a $gemm_log
./profile_gemm.sh gemm 1 0 0 1 0 5 | tee -a $gemm_log
./profile_gemm.sh gemm 2 0 0 1 0 5 | tee -a $gemm_log
./profile_gemm.sh gemm 3 0 0 1 0 5 | tee -a $gemm_log
......@@ -36,22 +46,14 @@ hipcc --version | grep -e 'HIP version' >> ${gemm_log}
./profile_gemm.sh gemm 1 3 0 1 0 5 | tee -a $gemm_log
./profile_gemm.sh gemm 2 3 0 1 0 5 | tee -a $gemm_log
./profile_gemm.sh gemm 3 3 0 1 0 5 | tee -a $gemm_log
python3 parse_perf_data.py ${gemm_log}
python3 process_perf_data.py $gemm_log
#run resnet50 test
export resnet_log="perf_resnet50.log"
rm -f $resnet_log
git status | grep -e 'On branch' > ${resnet_log}
echo -n 'Node name: '>>${resnet_log}; hostname >>${resnet_log}
#get GPU_arch and number of compute units from rocminfo
echo -n "GPU_arch: " >> ${resnet_log}; rocminfo | grep "Name:" | grep "gfx" >> ${resnet_log}
rocminfo | grep "Compute Unit:" >> ${resnet_log}
hipcc --version | grep -e 'HIP version' >> ${resnet_log}
/opt/rocm/bin/amdclang++ --version | grep -e 'InstalledDir' >> ${resnet_log}
#first run tests with N=256
./profile_conv.sh conv_fwd_bias_relu 1 1 1 1 0 2 0 1 256 | tee -a ${resnet_log}
#then run with N=4
./profile_conv.sh conv_fwd_bias_relu 1 1 1 1 0 2 0 1 4 | tee -a ${resnet_log}
#the script will put the results from N=256 and N=4 runs into separate tables
python3 parse_perf_data.py ${resnet_log}
export resnet256_log="perf_resnet50_N256.log"
print_log_header $resnet256_log $env_type
./profile_resnet50.sh conv_fwd_bias_relu 1 1 1 1 0 2 0 1 256 | tee -a $resnet256_log
python3 process_perf_data.py $resnet256_log
export resnet4_log="perf_resnet50_N4.log"
print_log_header $resnet4_log $env_type
./profile_resnet50.sh conv_fwd_bias_relu 1 1 1 1 0 2 0 1 4 | tee -a $resnet4_log
python3 process_perf_data.py $resnet4_log
test/CMakeLists.txt
View file @ 2d92594a
......@@ -47,3 +47,4 @@ add_subdirectory(convnd_bwd_weight)
add_subdirectory(convnd_bwd_data)
add_subdirectory(block_to_ctile_map)
add_subdirectory(softmax)
add_subdirectory(layernorm)
test/layernorm/CMakeLists.txt 0 → 100644
View file @ 2d92594a
add_custom_target(test_layernorm)
add_gtest_executable(test_layernorm_fp32 test_layernorm_fp32.cpp)
add_gtest_executable(test_layernorm_fp16 test_layernorm_fp16.cpp)
target_link_libraries(test_layernorm_fp32 PRIVATE host_tensor)
target_link_libraries(test_layernorm_fp16 PRIVATE host_tensor)
add_dependencies(test_layernorm test_layernorm_fp32)
add_dependencies(test_layernorm test_layernorm_fp16)
\ No newline at end of file
test/layernorm/test_layernorm_fp16.cpp 0 → 100644
View file @ 2d92594a
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "gtest/gtest.h"
#include "test_layernorm_util.hpp"
template <ck::index_t N>
using I = ck::Number<N>;
template <typename Tuple>
class TestLayernormFP16 : public ck::TestLayernorm<Tuple>
{
};
// clang-format off
using KernelTypes = ::testing::Types<
// XDataType, GammaDataType, BetaDataType, AccDataType, YDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, , GammaSrcVectorSize, 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<8>, 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<8>, 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<8>, 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<8>, 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<8>, 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<8>, 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<8>, 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<8>, I<8>, I<8>>
>;
// clang-format on
TYPED_TEST_SUITE(TestLayernormFP16, KernelTypes);
TYPED_TEST(TestLayernormFP16, Test_FP16) { this->Run(); }
test/layernorm/test_layernorm_fp32.cpp 0 → 100644
View file @ 2d92594a
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "gtest/gtest.h"
#include "test_layernorm_util.hpp"
template <ck::index_t N>
using I = ck::Number<N>;
template <typename Tuple>
class TestLayernormFP32 : public ck::TestLayernorm<Tuple>
{
};
// clang-format off
using KernelTypes = ::testing::Types<
// XDataType, GammaDataType, BetaDataType, AccDataType, YDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, , GammaSrcVectorSize, 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<4>, 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<4>, 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<4>, 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<4>, 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<4>, 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<4>, 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<4>, 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<4>, I<4>, I<4>>
>;
// clang-format on
TYPED_TEST_SUITE(TestLayernormFP32, KernelTypes);
TYPED_TEST(TestLayernormFP32, Test_FP32) { this->Run(); }
test/layernorm/test_layernorm_util.hpp 0 → 100644
View file @ 2d92594a
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <vector>
#include <iostream>
#include <gtest/gtest.h>
#include "ck/ck.hpp"
#include "ck/utility/number.hpp"
#include "ck/tensor_operation/gpu/device/device_layernorm.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/host_tensor/host_tensor.hpp"
#include "ck/library/host_tensor/device_memory.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_layernorm.hpp"
namespace ck {
template <typename Range>
std::string serialize_range(const Range& range)
{
std::stringstream ss;
for(auto& r : range)
{
ss << r << ", ";
}
std::string str = ss.str();
return std::string(str.begin(), str.end() - 2);
}
template <typename Tuple>
class TestLayernorm : 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>;
static constexpr index_t Rank = std::tuple_element_t<5, Tuple>{}.value;
static constexpr index_t NumReduceDim = std::tuple_element_t<6, Tuple>{}.value;
static constexpr index_t BlockSize = std::tuple_element_t<7, Tuple>{}.value;
static constexpr index_t MThreadClusterSize = std::tuple_element_t<8, Tuple>{}.value;
static constexpr index_t KThreadClusterSize = std::tuple_element_t<9, Tuple>{}.value;
static constexpr index_t MThreadSliceSize = std::tuple_element_t<10, Tuple>{}.value;
static constexpr index_t KThreadSliceSize = std::tuple_element_t<11, Tuple>{}.value;
static constexpr index_t XYSrcVectorDim = std::tuple_element_t<12, Tuple>{}.value;
static constexpr index_t XSrcVectorSize = std::tuple_element_t<13, Tuple>{}.value;
static constexpr index_t GammaSrcVectorSize = std::tuple_element_t<14, Tuple>{}.value;
static constexpr index_t BetaSrcVectorSize = std::tuple_element_t<15, Tuple>{}.value;
static constexpr index_t YDstVectorSize = std::tuple_element_t<16, Tuple>{}.value;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ReferenceInstance = tensor_operation::host::ReferenceLayernorm<XDataType,
GammaDataType,
BetaDataType,
YDataType,
AccDataType,
PassThrough,
Rank,
NumReduceDim>;
using DeviceInstance = tensor_operation::device::DeviceLayernorm<XDataType,
GammaDataType,
BetaDataType,
AccDataType,
YDataType,
PassThrough,
Rank,
NumReduceDim,
BlockSize,
MThreadClusterSize,
KThreadClusterSize,
MThreadSliceSize,
KThreadSliceSize,
XYSrcVectorDim,
XSrcVectorSize,
GammaSrcVectorSize,
BetaSrcVectorSize,
YDstVectorSize>;
TestLayernorm() : ref_instance_invoker_(ReferenceInstance{}.MakeInvoker()) {}
void RunSingle(std::vector<index_t> lengths, std::vector<index_t> reduceDims)
{
std::vector<index_t> reduceLength(reduceDims.size());
for(int i = 0; i < NumReduceDim; ++i)
{
reduceLength[i] = lengths[reduceDims[i]];
}
Tensor<XDataType> x(lengths);
Tensor<GammaDataType> gamma(reduceLength);
Tensor<BetaDataType> beta(reduceLength);
Tensor<YDataType> y(lengths);
Tensor<YDataType> y_ref(lengths);
x.GenerateTensorValue(GeneratorTensor_3<XDataType>{0, 1.0});
gamma.GenerateTensorValue(GeneratorTensor_3<GammaDataType>{0.0, 1.0});
beta.GenerateTensorValue(GeneratorTensor_3<BetaDataType>{0.0, 1.0});
DeviceMem x_dev(sizeof(XDataType) * x.mDesc.GetElementSpace());
DeviceMem gamma_dev(sizeof(GammaDataType) * gamma.mDesc.GetElementSpace());
DeviceMem beta_dev(sizeof(BetaDataType) * beta.mDesc.GetElementSpace());
DeviceMem y_dev(sizeof(YDataType) * y.mDesc.GetElementSpace());
x_dev.ToDevice(x.mData.data());
gamma_dev.ToDevice(gamma.mData.data());
beta_dev.ToDevice(beta.mData.data());
auto device_instance = DeviceInstance{};
auto argument_ptr = device_instance.MakeArgumentPointer(
lengths,
std::vector<ck::index_t>{x.mDesc.GetStrides().begin(), x.mDesc.GetStrides().end()},
std::vector<ck::index_t>{gamma.mDesc.GetStrides().begin(),
gamma.mDesc.GetStrides().end()},
std::vector<ck::index_t>{beta.mDesc.GetStrides().begin(),
beta.mDesc.GetStrides().end()},
reduceDims,
1e-4,
x_dev.GetDeviceBuffer(),
gamma_dev.GetDeviceBuffer(),
beta_dev.GetDeviceBuffer(),
y_dev.GetDeviceBuffer(),
PassThrough{});
if(!device_instance.IsSupportedArgument(argument_ptr.get()))
{
return;
}
auto invoker_ptr = device_instance.MakeInvokerPointer();
invoker_ptr->Run(argument_ptr.get());
ref_instance_invoker_.Run(
{x, gamma, beta, y_ref, PassThrough{}, lengths, reduceDims, 1e-4});
y_dev.FromDevice(y.mData.data());
bool pass;
if(std::is_same<XDataType, int8_t>::value)
{
EXPECT_TRUE(pass = ck::utils::check_err(
y.mData, y_ref.mData, "Error: Incorrect results!", 0, 1));
}
else
{
EXPECT_TRUE(pass = ck::utils::check_err(
y.mData, y_ref.mData, "Error: Incorrect results d1", 1e-3, 1e-3));
}
if(!pass)
{
FAIL() << "Failure in input lengths = [" << serialize_range(lengths) << "], "
<< "reduce dim = [" << serialize_range(reduceDims) << "].";
}
}
void Run()
{
for(auto length : this->lengths_)
{
this->RunSingle(length, reduceDims_[0]);
}
}
std::vector<std::vector<index_t>> lengths_ = {
{4, 256}, {8, 511}, {9, 1032}, {4, 2048}, {1, 8192}, {4000, 2000}};
std::vector<std::vector<index_t>> reduceDims_ = {{1}};
typename ReferenceInstance::Invoker ref_instance_invoker_;
};
} // namespace ck
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