Skip to content

GitLab

Commit aea62819 authored by Chaitanya Inumella's avatar Chaitanya Inumella
Browse files

Rebase branch 'develop' of... 

Rebase branch 'develop' of https://github.com/ROCmSoftwarePlatform/composable_kernel into contraction_hipTENSOR
parents 75af5450 75ab874e
Hide whitespace changes
Inline Side-by-side
Showing with 577 additions and 1171 deletions
test/convnd_fwd/CMakeLists.txt
View file @ aea62819
add_custom_target(test_convnd_fwd)
add_gtest_executable(test_conv1d_fwd conv1d_fwd.cpp)
target_link_libraries(test_conv1d_fwd PRIVATE host_tensor device_conv1d_fwd_instance conv_util)
add_dependencies(test_convnd_fwd test_conv1d_fwd)
add_gtest_executable(test_conv2d_fwd conv2d_fwd.cpp)
target_link_libraries(test_conv2d_fwd PRIVATE host_tensor device_conv2d_fwd_instance device_convnd_2d_fwd_instance conv_util)
add_dependencies(test_convnd_fwd test_conv2d_fwd)
add_gtest_executable(test_conv3d_fwd conv3d_fwd.cpp)
target_link_libraries(test_conv3d_fwd PRIVATE host_tensor device_conv3d_fwd_instance conv_util)
add_dependencies(test_convnd_fwd test_conv3d_fwd)
add_gtest_executable(test_convnd_fwd convnd_fwd.cpp)
target_link_libraries(test_convnd_fwd PRIVATE utility device_conv2d_fwd_instance)
test/convnd_fwd/conv1d_fwd.cpp deleted 100644 → 0
View file @ 75af5450
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <tuple>
#include <vector>
#include <gtest/gtest.h>
#include "ck/utility/data_type.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/conv_util.hpp"
#include "test/convnd_fwd/conv_util.hpp"
namespace {
class Conv1dFwdNWCInstances : public ::testing::Test
{
public:
template <typename T>
bool test_conv1d_nwc_instances(const std::vector<test::conv::DeviceConvFwdNoOpPtr>& conv_ptrs,
const ck::utils::conv::ConvParams& params)
{
using namespace std::placeholders;
using namespace ck::utils;
namespace ctl = ck::tensor_layout::convolution;
conv::ConvFwdOpInstance<T,
T,
T,
ctl::NWC,
ctl::KXC,
ctl::NWK,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
FillUniformDistributionIntegerValue<T>,
FillUniformDistributionIntegerValue<T>>
conv_instance(params,
true,
FillUniformDistributionIntegerValue<T>{},
FillUniformDistributionIntegerValue<T>{});
auto reference_conv_fwd_fun =
std::bind(conv::run_reference_convolution_forward<1, T, T, T>, params, _1, _2, _3);
OpInstanceRunEngine<T, T, T> run_engine(conv_instance, reference_conv_fwd_fun);
run_engine.SetAtol(atol_);
run_engine.SetRtol(rtol_);
return run_engine.Test(conv_ptrs);
}
template <typename T>
bool test_default()
{
return test_conv1d_nwc_instances<T>(
ck::utils::conv::ConvolutionFwdInstances<T, T, T>::template Get<1>(), params_default_);
}
template <typename T>
bool test_filter1x1_stride1_pad0()
{
return test_conv1d_nwc_instances<T>(
ck::utils::conv::ConvolutionFwdInstances<T, T, T>::template Get<1>(),
params_filter1x1_stride1_pad0_);
}
template <typename T>
bool test_filter1x1_pad0()
{
return test_conv1d_nwc_instances<T>(
ck::utils::conv::ConvolutionFwdInstances<T, T, T>::template Get<1>(),
params_filter1x1_pad0_);
}
static inline ck::utils::conv::ConvParams params_default_{
1, 4, 256, 64, {3}, {71}, {2}, {2}, {2}, {2}};
static inline ck::utils::conv::ConvParams params_filter1x1_stride1_pad0_{
1, 4, 256, 64, {1}, {28}, {1}, {1}, {0}, {0}};
static inline ck::utils::conv::ConvParams params_filter1x1_pad0_{
1, 4, 256, 64, {1}, {28}, {2}, {1}, {0}, {0}};
private:
double atol_{1e-5};
double rtol_{1e-4};
};
} // anonymous namespace
TEST(Conv1DFwdNWC, IntegerValues)
{
using namespace std::placeholders;
using namespace ck::utils;
namespace ctl = ck::tensor_layout::convolution;
using T = float;
ck::utils::conv::ConvParams params{1, 4, 256, 64, {3}, {36}, {1}, {2}, {2}, {2}};
std::vector<test::conv::DeviceConvFwdNoOpPtr> conv_ptrs;
test::conv::get_test_convolution_fwd_instance<1, T, T, T, T>(conv_ptrs);
conv::ConvFwdOpInstance<T,
T,
T,
ctl::NWC,
ctl::KXC,
ctl::NWK,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
FillUniformDistributionIntegerValue<T>,
FillUniformDistributionIntegerValue<T>>
conv_instance(params,
true,
FillUniformDistributionIntegerValue<T>{},
FillUniformDistributionIntegerValue<T>{});
auto reference_conv_fwd_fun =
std::bind(conv::run_reference_convolution_forward<1, T, T, T>, params, _1, _2, _3);
OpInstanceRunEngine<T, T, T> run_engine(conv_instance, reference_conv_fwd_fun);
run_engine.SetAtol(1e-5);
run_engine.SetRtol(1e-4);
EXPECT_TRUE(run_engine.Test(conv_ptrs));
}
TEST(Conv1DFwdNWC, FloatingPointValues)
{
using namespace std::placeholders;
using namespace ck::utils;
namespace ctl = ck::tensor_layout::convolution;
using T = ck::half_t;
ck::utils::conv::ConvParams params{1, 4, 256, 64, {3}, {36}, {1}, {2}, {2}, {2}};
std::vector<test::conv::DeviceConvFwdNoOpPtr> conv_ptrs;
test::conv::get_test_convolution_fwd_instance<1, T, T, T, float>(conv_ptrs);
conv::ConvFwdOpInstance<T,
T,
T,
ctl::NWC,
ctl::KXC,
ctl::NWK,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
FillUniformDistribution<T>,
FillUniformDistribution<T>>
conv_instance(params, true, FillUniformDistribution<T>{}, FillUniformDistribution<T>{});
auto reference_conv_fwd_fun =
std::bind(conv::run_reference_convolution_forward<1, T, T, T>, params, _1, _2, _3);
OpInstanceRunEngine<T, T, T> run_engine(conv_instance, reference_conv_fwd_fun);
run_engine.SetAtol(0.1);
run_engine.SetRtol(1e-2);
EXPECT_TRUE(run_engine.Test(conv_ptrs));
}
TEST_F(Conv1dFwdNWCInstances, BF16_default) { EXPECT_TRUE(this->test_default<ck::bhalf_t>()); }
TEST_F(Conv1dFwdNWCInstances, BF16_filter1x1_stride1_pad0)
{
EXPECT_TRUE(this->test_filter1x1_stride1_pad0<ck::bhalf_t>());
}
TEST_F(Conv1dFwdNWCInstances, BF16_filter1x1_pad0)
{
EXPECT_TRUE(this->test_filter1x1_pad0<ck::bhalf_t>());
}
TEST_F(Conv1dFwdNWCInstances, F16_default) { EXPECT_TRUE(this->test_default<ck::half_t>()); }
TEST_F(Conv1dFwdNWCInstances, F16_filter1x1_stride1_pad0)
{
EXPECT_TRUE(this->test_filter1x1_stride1_pad0<ck::half_t>());
}
TEST_F(Conv1dFwdNWCInstances, F16_filter1x1_pad0)
{
EXPECT_TRUE(this->test_filter1x1_pad0<ck::half_t>());
}
TEST_F(Conv1dFwdNWCInstances, F32_default) { EXPECT_TRUE(this->test_default<float>()); }
TEST_F(Conv1dFwdNWCInstances, F32_filter1x1_stride1_pad0)
{
EXPECT_TRUE(this->test_filter1x1_stride1_pad0<float>());
}
TEST_F(Conv1dFwdNWCInstances, F32_filter1x1_pad0)
{
EXPECT_TRUE(this->test_filter1x1_pad0<float>());
}
TEST_F(Conv1dFwdNWCInstances, I8_default) { EXPECT_TRUE(this->test_default<int8_t>()); }
TEST_F(Conv1dFwdNWCInstances, I8_filter1x1_stride1_pad0)
{
EXPECT_TRUE(this->test_filter1x1_stride1_pad0<int8_t>());
}
TEST_F(Conv1dFwdNWCInstances, I8_filter1x1_pad0)
{
EXPECT_TRUE(this->test_filter1x1_pad0<int8_t>());
}
test/convnd_fwd/conv2d_fwd.cpp deleted 100644 → 0
View file @ 75af5450
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <tuple>
#include <vector>
#include <gtest/gtest.h>
#include "ck/utility/data_type.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/conv_util.hpp"
#include "test/convnd_fwd/conv_util.hpp"
namespace {
class Conv2dFwdNHWCInstances : public ::testing::Test
{
public:
template <typename T>
bool test_conv2d_nhwc_instances(const std::vector<test::conv::DeviceConvFwdNoOpPtr>& conv_ptrs,
const ck::utils::conv::ConvParams& params)
{
using namespace std::placeholders;
using namespace ck::utils;
conv::ConvFwdOpInstance<T,
T,
T,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
FillUniformDistributionIntegerValue<T>,
FillUniformDistributionIntegerValue<T>>
conv_instance(params,
true,
FillUniformDistributionIntegerValue<T>{},
FillUniformDistributionIntegerValue<T>{});
auto reference_conv_fwd_fun =
std::bind(conv::run_reference_convolution_forward<2, T, T, T>, params, _1, _2, _3);
OpInstanceRunEngine<T, T, T> run_engine(conv_instance, reference_conv_fwd_fun);
run_engine.SetAtol(atol_);
run_engine.SetRtol(rtol_);
return run_engine.Test(conv_ptrs);
}
template <typename T>
bool test_default(bool use_convnd = false)
{
if(use_convnd)
{
return test_conv2d_nhwc_instances<T>(
test::conv::ConvolutionNDFwdInstances<T, T, T>::Get(2), params_default_);
}
else
{
return test_conv2d_nhwc_instances<T>(
ck::utils::conv::ConvolutionFwdInstances<T, T, T>::template Get<2>(),
params_default_);
}
}
template <typename T>
bool test_filter1x1_stride1_pad0(bool use_convnd = false)
{
if(use_convnd)
{
return test_conv2d_nhwc_instances<T>(
test::conv::ConvolutionNDFwdInstances<T, T, T>::Get(2),
params_filter1x1_stride1_pad0_);
}
else
{
return test_conv2d_nhwc_instances<T>(
ck::utils::conv::ConvolutionFwdInstances<T, T, T>::template Get<2>(),
params_filter1x1_stride1_pad0_);
}
}
template <typename T>
bool test_filter1x1_pad0(bool use_convnd = false)
{
if(use_convnd)
{
return test_conv2d_nhwc_instances<T>(
test::conv::ConvolutionNDFwdInstances<T, T, T>::Get(2), params_filter1x1_pad0_);
}
else
{
return test_conv2d_nhwc_instances<T>(
ck::utils::conv::ConvolutionFwdInstances<T, T, T>::template Get<2>(),
params_filter1x1_pad0_);
}
}
template <typename T>
bool test_oddC()
{
return test_conv2d_nhwc_instances<T>(
ck::utils::conv::ConvolutionFwdInstances<T, T, T>::template Get<2>(), params_oddC_);
}
static inline ck::utils::conv::ConvParams params_default_{
2, 4, 256, 64, {3, 3}, {36, 36}, {2, 2}, {2, 2}, {2, 2}, {2, 2}};
static inline ck::utils::conv::ConvParams params_filter1x1_stride1_pad0_{
2, 4, 256, 64, {1, 1}, {28, 28}, {1, 1}, {1, 1}, {0, 0}, {0, 0}};
static inline ck::utils::conv::ConvParams params_filter1x1_pad0_{
2, 4, 256, 64, {1, 1}, {28, 28}, {2, 2}, {1, 1}, {0, 0}, {0, 0}};
static inline ck::utils::conv::ConvParams params_oddC_{
2, 4, 256, 3, {3, 3}, {28, 28}, {1, 1}, {1, 1}, {0, 0}, {0, 0}};
private:
double atol_{1e-5};
double rtol_{1e-4};
};
} // anonymous namespace
TEST(Conv2DFwdNHWC, IntegerValues)
{
using namespace std::placeholders;
using namespace ck::utils;
using T = float;
ck::utils::conv::ConvParams params{
2, 4, 256, 64, {3, 3}, {36, 36}, {1, 1}, {2, 2}, {2, 2}, {2, 2}};
std::vector<test::conv::DeviceConvFwdNoOpPtr> conv_ptrs;
test::conv::get_test_convolution_fwd_instance<2, T, T, T, T>(conv_ptrs);
conv::ConvFwdOpInstance<T,
T,
T,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
FillUniformDistributionIntegerValue<T>,
FillUniformDistributionIntegerValue<T>>
conv_instance(params,
true,
FillUniformDistributionIntegerValue<T>{},
FillUniformDistributionIntegerValue<T>{});
auto reference_conv_fwd_fun =
std::bind(conv::run_reference_convolution_forward<2, T, T, T>, params, _1, _2, _3);
OpInstanceRunEngine<T, T, T> run_engine(conv_instance, reference_conv_fwd_fun);
run_engine.SetAtol(1e-5);
run_engine.SetRtol(1e-4);
EXPECT_TRUE(run_engine.Test(conv_ptrs));
}
TEST(Conv2DFwdNHWC, FloatingPointValues)
{
using namespace std::placeholders;
using namespace ck::utils;
using T = ck::half_t;
ck::utils::conv::ConvParams params{
2, 4, 256, 64, {3, 3}, {36, 36}, {2, 2}, {2, 2}, {2, 2}, {2, 2}};
std::vector<test::conv::DeviceConvFwdNoOpPtr> conv_ptrs;
test::conv::get_test_convolution_fwd_instance<2, T, T, T, float>(conv_ptrs);
conv::ConvFwdOpInstance<T,
T,
T,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
FillUniformDistribution<T>,
FillUniformDistribution<T>>
conv_instance(params, true, FillUniformDistribution<T>{}, FillUniformDistribution<T>{});
auto reference_conv_fwd_fun =
std::bind(conv::run_reference_convolution_forward<2, T, T, T>, params, _1, _2, _3);
OpInstanceRunEngine<T, T, T> run_engine(conv_instance, reference_conv_fwd_fun);
run_engine.SetAtol(2e-4);
run_engine.SetRtol(1e-3);
EXPECT_TRUE(run_engine.Test(conv_ptrs));
}
TEST_F(Conv2dFwdNHWCInstances, BF16_default) { EXPECT_TRUE(this->test_default<ck::bhalf_t>()); }
TEST_F(Conv2dFwdNHWCInstances, BF16_filter1x1_stride1_pad0)
{
EXPECT_TRUE(this->test_filter1x1_stride1_pad0<ck::bhalf_t>());
}
TEST_F(Conv2dFwdNHWCInstances, BF16_filter1x1_pad0)
{
EXPECT_TRUE(this->test_filter1x1_pad0<ck::bhalf_t>());
}
TEST_F(Conv2dFwdNHWCInstances, F16_default) { EXPECT_TRUE(this->test_default<ck::half_t>()); }
TEST_F(Conv2dFwdNHWCInstances, F16_filter1x1_stride1_pad0)
{
EXPECT_TRUE(this->test_filter1x1_stride1_pad0<ck::half_t>());
}
TEST_F(Conv2dFwdNHWCInstances, F16_filter1x1_pad0)
{
EXPECT_TRUE(this->test_filter1x1_pad0<ck::half_t>());
}
TEST_F(Conv2dFwdNHWCInstances, F16_oddC) { EXPECT_TRUE(this->test_oddC<ck::half_t>()); }
TEST_F(Conv2dFwdNHWCInstances, F32_default) { EXPECT_TRUE(this->test_default<float>()); }
TEST_F(Conv2dFwdNHWCInstances, F32_filter1x1_stride1_pad0)
{
EXPECT_TRUE(this->test_filter1x1_stride1_pad0<float>());
}
TEST_F(Conv2dFwdNHWCInstances, F32_filter1x1_pad0)
{
EXPECT_TRUE(this->test_filter1x1_pad0<float>());
}
TEST_F(Conv2dFwdNHWCInstances, I8_default) { EXPECT_TRUE(this->test_default<int8_t>()); }
TEST_F(Conv2dFwdNHWCInstances, I8_filter1x1_stride1_pad0)
{
EXPECT_TRUE(this->test_filter1x1_stride1_pad0<int8_t>());
}
TEST_F(Conv2dFwdNHWCInstances, I8_filter1x1_pad0)
{
EXPECT_TRUE(this->test_filter1x1_pad0<int8_t>());
}
TEST_F(Conv2dFwdNHWCInstances, ND_BF16_default)
{
EXPECT_TRUE(this->test_default<ck::bhalf_t>(true));
}
TEST_F(Conv2dFwdNHWCInstances, ND_BF16_filter1x1_stride1_pad0)
{
EXPECT_TRUE(this->test_filter1x1_stride1_pad0<ck::bhalf_t>(true));
}
TEST_F(Conv2dFwdNHWCInstances, ND_BF16_filter1x1_pad0)
{
EXPECT_TRUE(this->test_filter1x1_pad0<ck::bhalf_t>(true));
}
TEST_F(Conv2dFwdNHWCInstances, ND_F16_default)
{
EXPECT_TRUE(this->test_default<ck::half_t>(true));
}
TEST_F(Conv2dFwdNHWCInstances, ND_F16_filter1x1_stride1_pad0)
{
EXPECT_TRUE(this->test_filter1x1_stride1_pad0<ck::half_t>(true));
}
TEST_F(Conv2dFwdNHWCInstances, ND_F16_filter1x1_pad0)
{
EXPECT_TRUE(this->test_filter1x1_pad0<ck::half_t>(true));
}
TEST_F(Conv2dFwdNHWCInstances, ND_F32_default) { EXPECT_TRUE(this->test_default<float>(true)); }
TEST_F(Conv2dFwdNHWCInstances, ND_F32_filter1x1_stride1_pad0)
{
EXPECT_TRUE(this->test_filter1x1_stride1_pad0<float>(true));
}
TEST_F(Conv2dFwdNHWCInstances, ND_F32_filter1x1_pad0)
{
EXPECT_TRUE(this->test_filter1x1_pad0<float>(true));
}
TEST_F(Conv2dFwdNHWCInstances, ND_I8_default) { EXPECT_TRUE(this->test_default<int8_t>(true)); }
TEST_F(Conv2dFwdNHWCInstances, ND_I8_filter1x1_stride1_pad0)
{
EXPECT_TRUE(this->test_filter1x1_stride1_pad0<int8_t>(true));
}
TEST_F(Conv2dFwdNHWCInstances, ND_I8_filter1x1_pad0)
{
EXPECT_TRUE(this->test_filter1x1_pad0<int8_t>(true));
}
test/convnd_fwd/conv3d_fwd.cpp deleted 100644 → 0
View file @ 75af5450
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <stdexcept>
#include <tuple>
#include <vector>
#include <gtest/gtest.h>
#include "ck/utility/data_type.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/conv_util.hpp"
#include "test/convnd_fwd/conv_util.hpp"
namespace {
class Conv3dFwdNDHWCInstances : public ::testing::Test
{
public:
template <typename T>
bool test_conv3d_nwc_instances(const std::vector<test::conv::DeviceConvFwdNoOpPtr>& conv_ptrs,
const ck::utils::conv::ConvParams& params)
{
using namespace std::placeholders;
using namespace ck::utils;
namespace ctl = ck::tensor_layout::convolution;
conv::ConvFwdOpInstance<T,
T,
T,
ctl::NDHWC,
ctl::KZYXC,
ctl::NDHWK,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
FillUniformDistributionIntegerValue<T>,
FillUniformDistributionIntegerValue<T>>
conv_instance(params,
true,
FillUniformDistributionIntegerValue<T>{},
FillUniformDistributionIntegerValue<T>{});
auto reference_conv_fwd_fun =
std::bind(conv::run_reference_convolution_forward<3, T, T, T>, params, _1, _2, _3);
OpInstanceRunEngine<T, T, T> run_engine(conv_instance, reference_conv_fwd_fun);
run_engine.SetAtol(atol_);
run_engine.SetRtol(rtol_);
return run_engine.Test(conv_ptrs);
}
template <typename T>
bool test_default()
{
return test_conv3d_nwc_instances<T>(
ck::utils::conv::ConvolutionFwdInstances<T, T, T>::template Get<3>(), params_default_);
}
template <typename T>
bool test_filter1x1_stride1_pad0()
{
return test_conv3d_nwc_instances<T>(
ck::utils::conv::ConvolutionFwdInstances<T, T, T>::template Get<3>(),
params_filter1x1_stride1_pad0_);
}
template <typename T>
bool test_filter1x1_pad0()
{
return test_conv3d_nwc_instances<T>(
ck::utils::conv::ConvolutionFwdInstances<T, T, T>::template Get<3>(),
params_filter1x1_pad0_);
}
static inline ck::utils::conv::ConvParams params_default_{
3, 4, 256, 64, {3, 3, 3}, {28, 28, 28}, {2, 2, 2}, {2, 2, 2}, {2, 2, 2}, {2, 2, 2}};
static inline ck::utils::conv::ConvParams params_filter1x1_stride1_pad0_{
3, 4, 256, 64, {1, 1, 1}, {28, 28, 28}, {1, 1, 1}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}};
static inline ck::utils::conv::ConvParams params_filter1x1_pad0_{
3, 4, 256, 64, {1, 1, 1}, {28, 28, 28}, {2, 2, 2}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}};
private:
double atol_{1e-5};
double rtol_{1e-4};
};
} // anonymous namespace
TEST(Conv3DFwdNDHWC, IntegerValues)
{
using namespace std::placeholders;
using namespace ck::utils;
namespace ctl = ck::tensor_layout::convolution;
using T = float;
ck::utils::conv::ConvParams params{
3, 4, 256, 64, {3, 3, 3}, {18, 18, 18}, {1, 1, 1}, {2, 2, 2}, {2, 2, 2}, {2, 2, 2}};
std::vector<test::conv::DeviceConvFwdNoOpPtr> conv_ptrs;
test::conv::get_test_convolution_fwd_instance<3, T, T, T, T>(conv_ptrs);
conv::ConvFwdOpInstance<T,
T,
T,
ctl::NDHWC,
ctl::KZYXC,
ctl::NDHWK,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
FillUniformDistributionIntegerValue<T>,
FillUniformDistributionIntegerValue<T>>
conv_instance(params,
true,
FillUniformDistributionIntegerValue<T>{},
FillUniformDistributionIntegerValue<T>{});
auto reference_conv_fwd_fun =
std::bind(conv::run_reference_convolution_forward<3, T, T, T>, params, _1, _2, _3);
OpInstanceRunEngine<T, T, T> run_engine(conv_instance, reference_conv_fwd_fun);
run_engine.SetAtol(1e-5);
run_engine.SetRtol(1e-3);
EXPECT_TRUE(run_engine.Test(conv_ptrs));
}
TEST(Conv3DFwdNDHWC, FloatingPointValues)
{
using namespace std::placeholders;
using namespace ck::utils;
namespace ctl = ck::tensor_layout::convolution;
using T = ck::half_t;
ck::utils::conv::ConvParams params{
3, 4, 256, 64, {3, 3, 3}, {18, 18, 18}, {1, 1, 1}, {2, 2, 2}, {2, 2, 2}, {2, 2, 2}};
std::vector<test::conv::DeviceConvFwdNoOpPtr> conv_ptrs;
test::conv::get_test_convolution_fwd_instance<3, T, T, T, float>(conv_ptrs);
conv::ConvFwdOpInstance<T,
T,
T,
ctl::NDHWC,
ctl::KZYXC,
ctl::NDHWK,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
FillUniformDistribution<T>,
FillUniformDistribution<T>>
conv_instance(params, true, FillUniformDistribution<T>{}, FillUniformDistribution<T>{});
auto reference_conv_fwd_fun =
std::bind(conv::run_reference_convolution_forward<3, T, T, T>, params, _1, _2, _3);
OpInstanceRunEngine<T, T, T> run_engine(conv_instance, reference_conv_fwd_fun);
run_engine.SetAtol(1e-3);
run_engine.SetRtol(1e-3);
EXPECT_TRUE(run_engine.Test(conv_ptrs));
}
TEST(Conv3DFwdNDHWC, InputOver2GB)
{
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using namespace ck::utils;
using T = float;
// >2GB Input
conv::ConvParams params;
params.num_dim_spatial_ = 3;
params.N_ = 2;
params.K_ = 16;
params.C_ = 32;
params.filter_spatial_lengths_ = std::vector<ck::index_t>{3, 3, 3};
params.input_spatial_lengths_ = std::vector<ck::index_t>{32, 1000, 1000};
params.conv_filter_strides_ = std::vector<ck::index_t>{1, 1, 1};
params.conv_filter_dilations_ = std::vector<ck::index_t>{1, 1, 1};
params.input_left_pads_ = std::vector<ck::index_t>{1, 1, 1};
params.input_right_pads_ = std::vector<ck::index_t>{1, 1, 1};
std::vector<test::conv::DeviceConvFwdNoOpPtr> conv_ptrs;
test::conv::get_test_convolution_fwd_instance<3, T, T, T, T>(conv_ptrs);
auto arg = conv_ptrs.back()->MakeArgumentPointer(nullptr,
nullptr,
nullptr,
params.N_,
params.K_,
params.C_,
params.input_spatial_lengths_,
params.filter_spatial_lengths_,
params.GetOutputSpatialLengths(),
params.conv_filter_strides_,
params.conv_filter_dilations_,
params.input_left_pads_,
params.input_right_pads_,
PassThrough{},
PassThrough{},
PassThrough{});
EXPECT_FALSE(conv_ptrs.back()->IsSupportedArgument(arg.get()));
}
TEST(Conv3DFwdNDHWC, FiltersOver2GB)
{
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using namespace ck::utils;
using T = float;
// >2GB Filters
conv::ConvParams params;
params.num_dim_spatial_ = 3;
params.N_ = 2;
params.K_ = 16;
params.C_ = 32;
params.filter_spatial_lengths_ = std::vector<ck::index_t>{4, 1000, 1000};
params.input_spatial_lengths_ = std::vector<ck::index_t>{16, 16, 16};
params.conv_filter_strides_ = std::vector<ck::index_t>{1, 1, 1};
params.conv_filter_dilations_ = std::vector<ck::index_t>{1, 1, 1};
params.input_left_pads_ = std::vector<ck::index_t>{1, 1, 1};
params.input_right_pads_ = std::vector<ck::index_t>{1, 1, 1};
std::vector<test::conv::DeviceConvFwdNoOpPtr> conv_ptrs;
test::conv::get_test_convolution_fwd_instance<3, T, T, T, T>(conv_ptrs);
auto arg = conv_ptrs.back()->MakeArgumentPointer(nullptr,
nullptr,
nullptr,
params.N_,
params.K_,
params.C_,
params.input_spatial_lengths_,
params.filter_spatial_lengths_,
params.GetOutputSpatialLengths(),
params.conv_filter_strides_,
params.conv_filter_dilations_,
params.input_left_pads_,
params.input_right_pads_,
PassThrough{},
PassThrough{},
PassThrough{});
EXPECT_FALSE(conv_ptrs.back()->IsSupportedArgument(arg.get()));
}
TEST(Conv3DFwdNDHWC, OutputOver2GB)
{
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using namespace ck::utils;
using T = float;
// >2GB Output
conv::ConvParams params;
params.num_dim_spatial_ = 3;
params.N_ = 2;
params.K_ = 16;
params.C_ = 2;
params.filter_spatial_lengths_ = std::vector<ck::index_t>{1, 1, 1};
params.input_spatial_lengths_ = std::vector<ck::index_t>{1000, 1000, 30};
params.conv_filter_strides_ = std::vector<ck::index_t>{1, 1, 1};
params.conv_filter_dilations_ = std::vector<ck::index_t>{1, 1, 1};
params.input_left_pads_ = std::vector<ck::index_t>{2, 2, 2};
params.input_right_pads_ = std::vector<ck::index_t>{2, 2, 2};
std::vector<test::conv::DeviceConvFwdNoOpPtr> conv_ptrs;
test::conv::get_test_convolution_fwd_instance<3, T, T, T, T>(conv_ptrs);
auto arg = conv_ptrs.back()->MakeArgumentPointer(nullptr,
nullptr,
nullptr,
params.N_,
params.K_,
params.C_,
params.input_spatial_lengths_,
params.filter_spatial_lengths_,
params.GetOutputSpatialLengths(),
params.conv_filter_strides_,
params.conv_filter_dilations_,
params.input_left_pads_,
params.input_right_pads_,
PassThrough{},
PassThrough{},
PassThrough{});
EXPECT_FALSE(conv_ptrs.back()->IsSupportedArgument(arg.get()));
}
TEST_F(Conv3dFwdNDHWCInstances, BF16_default) { EXPECT_TRUE(this->test_default<ck::bhalf_t>()); }
TEST_F(Conv3dFwdNDHWCInstances, BF16_filter1x1_stride1_pad0)
{
EXPECT_TRUE(this->test_filter1x1_stride1_pad0<ck::bhalf_t>());
}
TEST_F(Conv3dFwdNDHWCInstances, BF16_filter1x1_pad0)
{
EXPECT_TRUE(this->test_filter1x1_pad0<ck::bhalf_t>());
}
TEST_F(Conv3dFwdNDHWCInstances, F16_default) { EXPECT_TRUE(this->test_default<ck::half_t>()); }
TEST_F(Conv3dFwdNDHWCInstances, F16_filter1x1_stride1_pad0)
{
EXPECT_TRUE(this->test_filter1x1_stride1_pad0<ck::half_t>());
}
TEST_F(Conv3dFwdNDHWCInstances, F16_filter1x1_pad0)
{
EXPECT_TRUE(this->test_filter1x1_pad0<ck::half_t>());
}
TEST_F(Conv3dFwdNDHWCInstances, F32_default) { EXPECT_TRUE(this->test_default<float>()); }
TEST_F(Conv3dFwdNDHWCInstances, F32_filter1x1_stride1_pad0)
{
EXPECT_TRUE(this->test_filter1x1_stride1_pad0<float>());
}
TEST_F(Conv3dFwdNDHWCInstances, F32_filter1x1_pad0)
{
EXPECT_TRUE(this->test_filter1x1_pad0<float>());
}
TEST_F(Conv3dFwdNDHWCInstances, I8_default) { EXPECT_TRUE(this->test_default<int8_t>()); }
TEST_F(Conv3dFwdNDHWCInstances, I8_filter1x1_stride1_pad0)
{
EXPECT_TRUE(this->test_filter1x1_stride1_pad0<int8_t>());
}
TEST_F(Conv3dFwdNDHWCInstances, I8_filter1x1_pad0)
{
EXPECT_TRUE(this->test_filter1x1_pad0<int8_t>());
}
test/convnd_fwd/conv_util.hpp deleted 100644 → 0
View file @ 75af5450
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <tuple>
#include "ck/ck.hpp"
#include "ck/utility/sequence.hpp"
#include "ck/utility/data_type.hpp"
#include "ck/tensor_operation/gpu/device/device_convnd_fwd_xdl_nhwc_kyxc_nhwk.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/host_tensor/host_tensor.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
using DeviceConvFwdNoOpPtr = DeviceConvFwdPtr<element_wise::PassThrough,
element_wise::PassThrough,
element_wise::PassThrough>;
namespace instance {
void add_device_convnd_2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instances(std::vector<DeviceConvFwdNoOpPtr>&);
void add_device_convnd_2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instances(std::vector<DeviceConvFwdNoOpPtr>&);
void add_device_convnd_2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instances(std::vector<DeviceConvFwdNoOpPtr>&);
void add_device_convnd_2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instances(std::vector<DeviceConvFwdNoOpPtr>&);
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
namespace test {
namespace conv {
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using InElementOp = ck::tensor_operation::element_wise::PassThrough;
using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
using DeviceConvFwdNoOpPtr =
ck::tensor_operation::device::DeviceConvFwdPtr<InElementOp, WeiElementOp, OutElementOp>;
static constexpr auto ConvFwdDefault =
ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
template <ck::index_t SpatialDims,
typename InDataType,
typename WeiDataType,
typename OutDataType,
typename AccDataType>
using DeviceConvNDFwdInstance = ck::tensor_operation::device::
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<
// clang-format off
InDataType, //
WeiDataType, //
OutDataType, //
AccDataType, // Accumulator data type.
InElementOp, // Input Elementwise Operation
WeiElementOp, // Weights Elementwise Operation
OutElementOp, // Output Elementwise Operation
ConvFwdDefault, // ConvForwardSpecialization
SpatialDims, // SptialDims
256, // BlockSize
128, // MPerBlock
256, // NPerBlock
4, // K0PerBlock
8, // K1
32, // MPerXdl
32, // NPerXdl
2, // MXdlPerWave
4, // NXdlPerWave
S<4, 64, 1>, // ABlockTransferThreadClusterLengths_K0_M_K1
S<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // ABlockTransferSrcAccessOrder
2, // ABlockTransferSrcVectorDim
8, // ABlockTransferSrcScalarPerVector
8, // ABlockTransferDstScalarPerVector_K1
true, // ABlockLdsAddExtraM
S<4, 64, 1>, // BBlockTransferThreadClusterLengths_K0_N_K1
S<1, 0, 2>, // BBlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // BBlockTransferSrcAccessOrder
2, // BBlockTransferSrcVectorDim
8, // BBlockTransferSrcScalarPerVector
8, // BBlockTransferDstScalarPerVector_K1
true, // BBlockLdsAddExtraN
7, // CThreadTransferSrcDstVectorDim
1>; // CThreadTransferDstScalarPerVector
// clang-format on
template <ck::index_t NDim,
typename InDataType,
typename WeiDataType,
typename OutDataType,
typename AccDataType>
void get_test_convolution_fwd_instance(std::vector<DeviceConvFwdNoOpPtr>& instances)
{
using ConvInstanceT =
DeviceConvNDFwdInstance<NDim, InDataType, WeiDataType, OutDataType, AccDataType>;
instances.emplace_back(std::make_unique<ConvInstanceT>());
}
// TODO (aosewski)
// Temporary solution to get all DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
// instances. When switched over to DeviceConvNDFwdXdl for 2D remove ConvolutionNDFwdInstances
// structures.
template <typename InDataType, typename WeiDataType, typename OutDataType>
struct ConvolutionNDFwdInstances;
template <>
struct ConvolutionNDFwdInstances<float, float, float>
{
static std::vector<DeviceConvFwdNoOpPtr> Get(std::size_t num_dim_spatial)
{
std::vector<DeviceConvFwdNoOpPtr> conv_ptrs;
if(num_dim_spatial == 2)
{
ck::tensor_operation::device::instance::
add_device_convnd_2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instances(conv_ptrs);
}
return conv_ptrs;
}
};
template <>
struct ConvolutionNDFwdInstances<ck::half_t, ck::half_t, ck::half_t>
{
static std::vector<DeviceConvFwdNoOpPtr> Get(std::size_t num_dim_spatial)
{
std::vector<DeviceConvFwdNoOpPtr> conv_ptrs;
if(num_dim_spatial == 2)
{
ck::tensor_operation::device::instance::
add_device_convnd_2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instances(conv_ptrs);
}
return conv_ptrs;
}
};
template <>
struct ConvolutionNDFwdInstances<ck::bhalf_t, ck::bhalf_t, ck::bhalf_t>
{
static std::vector<DeviceConvFwdNoOpPtr> Get(std::size_t num_dim_spatial)
{
std::vector<DeviceConvFwdNoOpPtr> conv_ptrs;
if(num_dim_spatial == 2)
{
ck::tensor_operation::device::instance::
add_device_convnd_2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instances(conv_ptrs);
}
return conv_ptrs;
}
};
template <>
struct ConvolutionNDFwdInstances<int8_t, int8_t, int8_t>
{
static std::vector<DeviceConvFwdNoOpPtr> Get(std::size_t num_dim_spatial)
{
std::vector<DeviceConvFwdNoOpPtr> conv_ptrs;
if(num_dim_spatial == 2)
{
ck::tensor_operation::device::instance::
add_device_convnd_2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instances(conv_ptrs);
}
return conv_ptrs;
}
};
} // namespace conv
} // namespace test
test/convnd_fwd/convnd_fwd.cpp 0 → 100644
View file @ aea62819
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include <iostream>
#include <initializer_list>
#include <vector>
#include <gtest/gtest.h>
#include "profiler/include/profile_conv_fwd_impl.hpp"
class TestConvndFwd : public ::testing::Test
{
protected:
std::vector<ck::utils::conv::ConvParam> conv_params;
};
// 1d
TEST_F(TestConvndFwd, Conv1dFwd)
{
conv_params.clear();
conv_params.push_back({1, 1, 128, 128, 256, {1}, {14}, {2}, {1}, {0}, {0}});
conv_params.push_back({1, 1, 128, 128, 256, {3}, {28}, {1}, {1}, {1}, {1}});
conv_params.push_back({1, 1, 128, 128, 256, {1}, {3}, {1}, {1}, {0}, {0}});
for(auto& param : conv_params)
{
bool pass;
// fp32
pass = ck::profiler::profile_conv_fwd_impl<1,
ck::tensor_layout::convolution::NWC,
ck::tensor_layout::convolution::KXC,
ck::tensor_layout::convolution::NWK,
float,
float,
float>(true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param);
EXPECT_TRUE(pass);
// fp16
pass = ck::profiler::profile_conv_fwd_impl<1,
ck::tensor_layout::convolution::NWC,
ck::tensor_layout::convolution::KXC,
ck::tensor_layout::convolution::NWK,
ck::half_t,
ck::half_t,
ck::half_t>(true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param);
EXPECT_TRUE(pass);
// bf16
pass = ck::profiler::profile_conv_fwd_impl<1,
ck::tensor_layout::convolution::NWC,
ck::tensor_layout::convolution::KXC,
ck::tensor_layout::convolution::NWK,
ck::bhalf_t,
ck::bhalf_t,
ck::bhalf_t>(true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param);
EXPECT_TRUE(pass);
// int8
pass = ck::profiler::profile_conv_fwd_impl<1,
ck::tensor_layout::convolution::NWC,
ck::tensor_layout::convolution::KXC,
ck::tensor_layout::convolution::NWK,
int8_t,
int8_t,
int8_t>(true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param);
EXPECT_TRUE(pass);
}
}
// 2d
TEST_F(TestConvndFwd, Conv2dFwd)
{
conv_params.clear();
conv_params.push_back({2, 1, 128, 128, 256, {1, 1}, {7, 7}, {2, 2}, {1, 1}, {0, 0}, {0, 0}});
conv_params.push_back({2, 1, 128, 128, 256, {3, 3}, {14, 14}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
conv_params.push_back({2, 1, 128, 128, 256, {1, 1}, {3, 3}, {1, 1}, {1, 1}, {0, 0}, {0, 0}});
for(auto& param : conv_params)
{
bool pass;
// fp32
pass = ck::profiler::profile_conv_fwd_impl<2,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK,
float,
float,
float>(true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param);
EXPECT_TRUE(pass);
// fp16
pass = ck::profiler::profile_conv_fwd_impl<2,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK,
ck::half_t,
ck::half_t,
ck::half_t>(true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param);
EXPECT_TRUE(pass);
// bf16
pass = ck::profiler::profile_conv_fwd_impl<2,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK,
ck::bhalf_t,
ck::bhalf_t,
ck::bhalf_t>(true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param);
EXPECT_TRUE(pass);
// int8
pass = ck::profiler::profile_conv_fwd_impl<2,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK,
int8_t,
int8_t,
int8_t>(true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param);
EXPECT_TRUE(pass);
}
}
// 3d
TEST_F(TestConvndFwd, Conv3dFwd)
{
conv_params.clear();
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}});
conv_params.push_back(
{3, 1, 128, 128, 256, {3, 3, 3}, {14, 14, 3}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
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}});
for(auto& param : conv_params)
{
bool pass;
// fp32
pass = ck::profiler::profile_conv_fwd_impl<3,
ck::tensor_layout::convolution::NDHWC,
ck::tensor_layout::convolution::KZYXC,
ck::tensor_layout::convolution::NDHWK,
float,
float,
float>(true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param);
EXPECT_TRUE(pass);
// fp16
pass = ck::profiler::profile_conv_fwd_impl<3,
ck::tensor_layout::convolution::NDHWC,
ck::tensor_layout::convolution::KZYXC,
ck::tensor_layout::convolution::NDHWK,
ck::half_t,
ck::half_t,
ck::half_t>(true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param);
EXPECT_TRUE(pass);
// bf16
pass = ck::profiler::profile_conv_fwd_impl<3,
ck::tensor_layout::convolution::NDHWC,
ck::tensor_layout::convolution::KZYXC,
ck::tensor_layout::convolution::NDHWK,
ck::bhalf_t,
ck::bhalf_t,
ck::bhalf_t>(true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param);
EXPECT_TRUE(pass);
// int8
pass = ck::profiler::profile_conv_fwd_impl<3,
ck::tensor_layout::convolution::NDHWC,
ck::tensor_layout::convolution::KZYXC,
ck::tensor_layout::convolution::NDHWK,
int8_t,
int8_t,
int8_t>(true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param);
EXPECT_TRUE(pass);
}
}
test/gemm/CMakeLists.txt
View file @ aea62819
add_test_executable(test_gemm_fp32 gemm_fp32.cpp)
target_link_libraries(test_gemm_fp32 PRIVATE host_tensor)
target_link_libraries(test_gemm_fp32 PRIVATE utility)
target_link_libraries(test_gemm_fp32 PRIVATE device_gemm_instance)
add_test_executable(test_gemm_fp16 gemm_fp16.cpp)
target_link_libraries(test_gemm_fp16 PRIVATE host_tensor)
target_link_libraries(test_gemm_fp16 PRIVATE utility)
target_link_libraries(test_gemm_fp16 PRIVATE device_gemm_instance)
add_test_executable(test_gemm_bf16 gemm_bf16.cpp)
target_link_libraries(test_gemm_bf16 PRIVATE host_tensor)
target_link_libraries(test_gemm_bf16 PRIVATE utility)
target_link_libraries(test_gemm_bf16 PRIVATE device_gemm_instance)
add_test_executable(test_gemm_int8 gemm_int8.cpp)
target_link_libraries(test_gemm_int8 PRIVATE host_tensor)
target_link_libraries(test_gemm_int8 PRIVATE utility)
target_link_libraries(test_gemm_int8 PRIVATE device_gemm_instance)
test/gemm/gemm_bf16.cpp
View file @ aea62819
......@@ -17,9 +17,9 @@
#include "ck/library/tensor_operation_instance/gpu/gemm.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/host_tensor/device_memory.hpp"
#include "ck/library/host_tensor/host_tensor.hpp"
#include "ck/library/host_tensor/host_tensor_generator.hpp"
#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/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "test/gemm/gemm_util.hpp"
......
test/gemm/gemm_fp16.cpp
View file @ aea62819
......@@ -17,9 +17,9 @@
#include "ck/library/tensor_operation_instance/gpu/gemm.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/host_tensor/device_memory.hpp"
#include "ck/library/host_tensor/host_tensor.hpp"
#include "ck/library/host_tensor/host_tensor_generator.hpp"
#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/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "test/gemm/gemm_util.hpp"
......
test/gemm/gemm_fp32.cpp
View file @ aea62819
......@@ -17,9 +17,9 @@
#include "ck/library/tensor_operation_instance/gpu/gemm.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/host_tensor/device_memory.hpp"
#include "ck/library/host_tensor/host_tensor.hpp"
#include "ck/library/host_tensor/host_tensor_generator.hpp"
#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/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "test/gemm/gemm_util.hpp"
......
test/gemm/gemm_fp64.cpp
View file @ aea62819
......@@ -17,9 +17,9 @@
#include "ck/library/tensor_operation_instance/gpu/gemm.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/host_tensor/device_memory.hpp"
#include "ck/library/host_tensor/host_tensor.hpp"
#include "ck/library/host_tensor/host_tensor_generator.hpp"
#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/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "test/gemm/gemm_util.hpp"
......
test/gemm/gemm_int8.cpp
View file @ aea62819
......@@ -17,9 +17,9 @@
#include "ck/library/tensor_operation_instance/gpu/gemm.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/host_tensor/device_memory.hpp"
#include "ck/library/host_tensor/host_tensor.hpp"
#include "ck/library/host_tensor/host_tensor_generator.hpp"
#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/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "test/gemm/gemm_util.hpp"
......
test/gemm/gemm_util.hpp
View file @ aea62819
......@@ -6,9 +6,9 @@
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/host_tensor/device_memory.hpp"
#include "ck/library/host_tensor/host_tensor.hpp"
#include "ck/library/host_tensor/host_tensor_generator.hpp"
#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/reference_tensor_operation/cpu/reference_gemm.hpp"
namespace ck {
......@@ -71,9 +71,9 @@ bool RunDeviceGEMM(DeviceGemmPtr_& gemmPtr,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op)
{
DeviceMem a_m_k_device_buf(sizeof(ADataType) * A.mDesc.GetElementSpace());
DeviceMem b_k_n_device_buf(sizeof(BDataType) * B.mDesc.GetElementSpace());
DeviceMem c_m_n_device_buf(sizeof(CDataType) * C.mDesc.GetElementSpace());
DeviceMem a_m_k_device_buf(sizeof(ADataType) * A.mDesc.GetElementSpaceSize());
DeviceMem b_k_n_device_buf(sizeof(BDataType) * B.mDesc.GetElementSpaceSize());
DeviceMem c_m_n_device_buf(sizeof(CDataType) * C.mDesc.GetElementSpaceSize());
auto invoker_ptr = gemmPtr->MakeInvokerPointer();
auto argument_ptr =
......
test/gemm_reduce/CMakeLists.txt
View file @ aea62819
add_test_executable(test_gemm_reduce_fp16 gemm_reduce_fp16.cpp)
target_link_libraries(test_gemm_reduce_fp16 PRIVATE host_tensor)
target_link_libraries(test_gemm_reduce_fp16 PRIVATE utility)
target_link_libraries(test_gemm_reduce_fp16 PRIVATE device_gemm_reduce_instance)
test/gemm_split_k/CMakeLists.txt
View file @ aea62819
add_test_executable(test_gemm_split_k gemm_split_k.cpp)
target_link_libraries(test_gemm_split_k PRIVATE host_tensor)
target_link_libraries(test_gemm_split_k PRIVATE utility)
target_link_libraries(test_gemm_split_k PRIVATE device_gemm_splitk_instance)
test/gemm_split_k/gemm_split_k.cpp
View file @ aea62819
......@@ -14,12 +14,12 @@
#include "ck/library/tensor_operation_instance/gpu/gemm_splitk.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/host_tensor/device_memory.hpp"
#include "ck/library/host_tensor/host_tensor.hpp"
#include "ck/library/host_tensor/host_tensor_generator.hpp"
#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/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "ck/library/host_tensor/host_gemm.hpp"
#include "ck/library/utility/host_gemm.hpp"
enum struct GemmMatrixLayout
{
......@@ -127,9 +127,9 @@ int test_gemm(const gemmArgs& args)
ck::tensor_operation::element_wise::PassThrough{},
ck::tensor_operation::element_wise::PassThrough{});
DeviceMem a_device_buf(sizeof(float) * a_m_k.mDesc.GetElementSpace());
DeviceMem b_device_buf(sizeof(float) * b_k_n.mDesc.GetElementSpace());
DeviceMem c_device_buf(sizeof(float) * c_m_n_device_result.mDesc.GetElementSpace());
DeviceMem a_device_buf(sizeof(float) * a_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(float) * b_k_n.mDesc.GetElementSpaceSize());
DeviceMem c_device_buf(sizeof(float) * c_m_n_device_result.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
......
test/grouped_convnd_fwd/CMakeLists.txt 0 → 100644
View file @ aea62819
add_gtest_executable(test_grouped_convnd_fwd grouped_convnd_fwd.cpp)
target_link_libraries(test_grouped_convnd_fwd PRIVATE utility device_grouped_conv1d_fwd_instance device_grouped_conv2d_fwd_instance device_grouped_conv3d_fwd_instance)
test/grouped_convnd_fwd/grouped_convnd_fwd.cpp 0 → 100644
View file @ aea62819
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include <iostream>
#include <initializer_list>
#include <vector>
#include <gtest/gtest.h>
#include "profiler/include/profile_grouped_conv_fwd_impl.hpp"
class TestGroupedConvNdFwd : public ::testing::Test
{
protected:
std::vector<ck::utils::conv::ConvParam> conv_params;
};
// 1d GNWC/GKXC/GNWK
TEST_F(TestGroupedConvNdFwd, GroupedConv1dFwdGNWC)
{
conv_params.clear();
conv_params.push_back({1, 2, 128, 128, 256, {1}, {14}, {2}, {1}, {0}, {0}});
conv_params.push_back({1, 2, 128, 128, 256, {3}, {28}, {1}, {1}, {1}, {1}});
conv_params.push_back({1, 2, 128, 128, 256, {1}, {3}, {1}, {1}, {0}, {0}});
for(auto& param : conv_params)
{
bool pass;
// fp32
pass = ck::profiler::profile_grouped_conv_fwd_impl<1,
ck::tensor_layout::convolution::GNWC,
ck::tensor_layout::convolution::GKXC,
ck::tensor_layout::convolution::GNWK,
float,
float,
float>(true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param);
EXPECT_TRUE(pass);
// fp16
pass = ck::profiler::profile_grouped_conv_fwd_impl<1,
ck::tensor_layout::convolution::GNWC,
ck::tensor_layout::convolution::GKXC,
ck::tensor_layout::convolution::GNWK,
ck::half_t,
ck::half_t,
ck::half_t>(true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param);
EXPECT_TRUE(pass);
// bf16
pass = ck::profiler::profile_grouped_conv_fwd_impl<1,
ck::tensor_layout::convolution::GNWC,
ck::tensor_layout::convolution::GKXC,
ck::tensor_layout::convolution::GNWK,
ck::bhalf_t,
ck::bhalf_t,
ck::bhalf_t>(true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param);
EXPECT_TRUE(pass);
// int8
pass = ck::profiler::profile_grouped_conv_fwd_impl<1,
ck::tensor_layout::convolution::GNWC,
ck::tensor_layout::convolution::GKXC,
ck::tensor_layout::convolution::GNWK,
int8_t,
int8_t,
int8_t>(true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param);
EXPECT_TRUE(pass);
}
}
// 2d GNHWC/GKYXC/GNHWK
TEST_F(TestGroupedConvNdFwd, GroupedConv2dFwdGNHWC)
{
conv_params.clear();
conv_params.push_back({2, 2, 128, 128, 256, {1, 1}, {7, 7}, {2, 2}, {1, 1}, {0, 0}, {0, 0}});
conv_params.push_back({2, 2, 128, 128, 256, {3, 3}, {14, 14}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
conv_params.push_back({2, 2, 128, 128, 256, {1, 1}, {3, 3}, {1, 1}, {1, 1}, {0, 0}, {0, 0}});
for(auto& param : conv_params)
{
bool pass;
// fp32
pass = ck::profiler::profile_grouped_conv_fwd_impl<2,
ck::tensor_layout::convolution::GNHWC,
ck::tensor_layout::convolution::GKYXC,
ck::tensor_layout::convolution::GNHWK,
float,
float,
float>(true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param);
EXPECT_TRUE(pass);
// fp16
pass = ck::profiler::profile_grouped_conv_fwd_impl<2,
ck::tensor_layout::convolution::GNHWC,
ck::tensor_layout::convolution::GKYXC,
ck::tensor_layout::convolution::GNHWK,
ck::half_t,
ck::half_t,
ck::half_t>(true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param);
EXPECT_TRUE(pass);
// bf16
pass = ck::profiler::profile_grouped_conv_fwd_impl<2,
ck::tensor_layout::convolution::GNHWC,
ck::tensor_layout::convolution::GKYXC,
ck::tensor_layout::convolution::GNHWK,
ck::bhalf_t,
ck::bhalf_t,
ck::bhalf_t>(true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param);
EXPECT_TRUE(pass);
// int8
pass = ck::profiler::profile_grouped_conv_fwd_impl<2,
ck::tensor_layout::convolution::GNHWC,
ck::tensor_layout::convolution::GKYXC,
ck::tensor_layout::convolution::GNHWK,
int8_t,
int8_t,
int8_t>(true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param);
EXPECT_TRUE(pass);
}
}
// 3d GNDHWC/GKZYXC/GNDHWK
TEST_F(TestGroupedConvNdFwd, GroupedConv3dFwdGNDHWC)
{
conv_params.clear();
conv_params.push_back(
{3, 2, 128, 128, 256, {1, 1, 1}, {7, 7, 7}, {2, 2, 2}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
conv_params.push_back(
{3, 2, 128, 128, 256, {3, 3, 3}, {14, 14, 3}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
conv_params.push_back(
{3, 2, 128, 128, 256, {1, 1, 1}, {3, 3, 3}, {1, 1, 1}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
for(auto& param : conv_params)
{
bool pass;
// fp32
pass = ck::profiler::profile_grouped_conv_fwd_impl<3,
ck::tensor_layout::convolution::GNDHWC,
ck::tensor_layout::convolution::GKZYXC,
ck::tensor_layout::convolution::GNDHWK,
float,
float,
float>(true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param);
EXPECT_TRUE(pass);
// fp16
pass = ck::profiler::profile_grouped_conv_fwd_impl<3,
ck::tensor_layout::convolution::GNDHWC,
ck::tensor_layout::convolution::GKZYXC,
ck::tensor_layout::convolution::GNDHWK,
ck::half_t,
ck::half_t,
ck::half_t>(true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param);
EXPECT_TRUE(pass);
// bf16
pass = ck::profiler::profile_grouped_conv_fwd_impl<3,
ck::tensor_layout::convolution::GNDHWC,
ck::tensor_layout::convolution::GKZYXC,
ck::tensor_layout::convolution::GNDHWK,
ck::bhalf_t,
ck::bhalf_t,
ck::bhalf_t>(true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param);
EXPECT_TRUE(pass);
// int8
pass = ck::profiler::profile_grouped_conv_fwd_impl<3,
ck::tensor_layout::convolution::GNDHWC,
ck::tensor_layout::convolution::GKZYXC,
ck::tensor_layout::convolution::GNDHWK,
int8_t,
int8_t,
int8_t>(true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param);
EXPECT_TRUE(pass);
}
}
// 2d NHWGC/KYXGC/NHWGK
TEST_F(TestGroupedConvNdFwd, GroupedConv2dFwdNHWGC)
{
conv_params.clear();
conv_params.push_back({2, 2, 128, 128, 256, {1, 1}, {7, 7}, {2, 2}, {1, 1}, {0, 0}, {0, 0}});
conv_params.push_back({2, 2, 128, 128, 256, {3, 3}, {14, 14}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
conv_params.push_back({2, 2, 128, 128, 256, {1, 1}, {3, 3}, {1, 1}, {1, 1}, {0, 0}, {0, 0}});
for(auto& param : conv_params)
{
bool pass;
// fp16
pass = ck::profiler::profile_grouped_conv_fwd_impl<2,
ck::tensor_layout::convolution::NHWGC,
ck::tensor_layout::convolution::KYXGC,
ck::tensor_layout::convolution::NHWGK,
ck::half_t,
ck::half_t,
ck::half_t>(true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param);
EXPECT_TRUE(pass);
}
}
test/grouped_gemm/CMakeLists.txt
View file @ aea62819
add_test_executable(test_grouped_gemm_fp16 grouped_gemm_fp16.cpp)
target_link_libraries(test_grouped_gemm_fp16 PRIVATE host_tensor)
target_link_libraries(test_grouped_gemm_fp16 PRIVATE utility)
target_link_libraries(test_grouped_gemm_fp16 PRIVATE device_grouped_gemm_instance)
test/grouped_gemm/grouped_gemm_fp16.cpp
View file @ aea62819
......@@ -2,39 +2,8 @@
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_gemm_xdl.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/host_tensor/device_memory.hpp"
#include "ck/library/host_tensor/host_tensor.hpp"
#include "ck/library/host_tensor/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using DeviceGroupedGemmPtr_ = ck::tensor_operation::device::DeviceGroupedGemmPtr<
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough>;
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(
std::vector<DeviceGroupedGemmPtr_>&);
}
} // namespace device
} // namespace tensor_operation
} // namespace ck
#include "profiler/include/profile_grouped_gemm_impl.hpp"
namespace {
......@@ -43,169 +12,52 @@ using BDataType = ck::half_t;
using CDataType = ck::half_t;
using AccDataType = float;
using ALayout = ck::tensor_layout::gemm::RowMajor;
using BLayout = ck::tensor_layout::gemm::ColumnMajor;
using CLayout = ck::tensor_layout::gemm::RowMajor;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
bool TestGroupedGemm(DeviceGroupedGemmPtr_& groupedGemmPtr)
template <typename ALayout, typename BLayout, typename CLayout>
bool TestGroupedGemm()
{
int group_count = rand() % 10 + 1;
// GEMM shape
std::vector<ck::tensor_operation::device::GemmShape> gemm_shapes;
std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
std::vector<const void*> p_a, p_b;
std::vector<void*> p_c;
gemm_shapes.reserve(group_count);
std::vector<int> Ms, Ns, Ks, StrideAs, StrideBs, StrideCs;
for(int i = 0; i < group_count; i++)
{
int M = 256 + 256 * (rand() % 10);
int N = 256 + 256 * (rand() % 10);
int K = 128 + 128 * (rand() % 10);
int AStride = std::is_same<ck::tensor_layout::gemm::RowMajor, ALayout>::value ? K : M;
int BStride = std::is_same<ck::tensor_layout::gemm::RowMajor, BLayout>::value ? N : K;
int CStride = std::is_same<ck::tensor_layout::gemm::RowMajor, CLayout>::value ? N : M;
gemm_shapes.push_back({M, N, K, AStride, BStride, CStride});
}
Ms.push_back(256 + 256 * (rand() % 10));
Ns.push_back(256 + 256 * (rand() % 10));
Ks.push_back(128 + 128 * (rand() % 10));
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
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}));
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({1, stride}));
}
};
std::vector<Tensor<ADataType>> a_tensors;
;
std::vector<Tensor<BDataType>> b_tensors;
std::vector<Tensor<CDataType>> c_host_tensors;
std::vector<Tensor<CDataType>> c_device_tensors;
a_tensors.reserve(group_count);
b_tensors.reserve(group_count);
c_host_tensors.reserve(group_count);
c_device_tensors.reserve(group_count);
using DeviceMemPtr = std::unique_ptr<DeviceMem>;
std::vector<DeviceMemPtr> a_tensors_device, b_tensors_device, c_tensors_device;
a_tensors_device.reserve(group_count);
b_tensors_device.reserve(group_count);
c_tensors_device.reserve(group_count);
for(std::size_t i = 0; i < gemm_shapes.size(); i++)
{
a_tensors.emplace_back(Tensor<ADataType>(f_host_tensor_descriptor(
gemm_shapes[i].M, gemm_shapes[i].K, gemm_shapes[i].StrideA, ALayout{})));
b_tensors.emplace_back(Tensor<BDataType>(f_host_tensor_descriptor(
gemm_shapes[i].K, gemm_shapes[i].N, gemm_shapes[i].StrideB, BLayout{})));
c_host_tensors.emplace_back(Tensor<CDataType>(f_host_tensor_descriptor(
gemm_shapes[i].M, gemm_shapes[i].N, gemm_shapes[i].StrideC, CLayout{})));
c_device_tensors.emplace_back(Tensor<CDataType>(f_host_tensor_descriptor(
gemm_shapes[i].M, gemm_shapes[i].N, gemm_shapes[i].StrideC, CLayout{})));
a_tensors[i].GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_tensors[i].GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
StrideAs.push_back(std::is_same<Row, ALayout>::value ? Ks[i] : Ms[i]);
StrideBs.push_back(std::is_same<Row, BLayout>::value ? Ns[i] : Ks[i]);
StrideCs.push_back(std::is_same<Row, CLayout>::value ? Ns[i] : Ms[i]);
}
for(std::size_t i = 0; i < gemm_shapes.size(); i++)
{
a_tensors_device.emplace_back(
std::make_unique<DeviceMem>(sizeof(ADataType) * a_tensors[i].mDesc.GetElementSize()));
b_tensors_device.emplace_back(
std::make_unique<DeviceMem>(sizeof(BDataType) * b_tensors[i].mDesc.GetElementSize()));
c_tensors_device.emplace_back(std::make_unique<DeviceMem>(
sizeof(CDataType) * c_device_tensors[i].mDesc.GetElementSize()));
a_tensors_device[i]->ToDevice(a_tensors[i].mData.data());
b_tensors_device[i]->ToDevice(b_tensors[i].mData.data());
p_a.push_back(a_tensors_device[i]->GetDeviceBuffer());
p_b.push_back(b_tensors_device[i]->GetDeviceBuffer());
p_c.push_back(c_tensors_device[i]->GetDeviceBuffer());
}
auto a_element_op = PassThrough{};
auto b_element_op = PassThrough{};
auto c_element_op = PassThrough{};
// do GEMM
auto invoker_ptr = groupedGemmPtr->MakeInvokerPointer();
auto argument_ptr = groupedGemmPtr->MakeArgumentPointer(
p_a, p_b, p_c, gemm_shapes, a_element_op, b_element_op, c_element_op);
DeviceMem gemm_desc_workspace(groupedGemmPtr->GetWorkSpaceSize(argument_ptr.get()));
groupedGemmPtr->SetWorkSpacePointer(argument_ptr.get(), gemm_desc_workspace.GetDeviceBuffer());
invoker_ptr->Run(argument_ptr.get());
for(std::size_t i = 0; i < gemm_shapes.size(); i++)
{
c_tensors_device[i]->FromDevice(c_device_tensors[i].mData.data());
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
CDataType,
AccDataType,
PassThrough,
PassThrough,
PassThrough>;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(a_tensors[i],
b_tensors[i],
c_host_tensors[i],
a_element_op,
b_element_op,
c_element_op);
if(!groupedGemmPtr->IsSupportedArgument(argument_ptr.get()))
{
return false;
}
ref_invoker.Run(ref_argument);
bool res = ck::utils::check_err(c_host_tensors[i].mData, c_device_tensors[i].mData);
std::cout << "group_id: " << i << (res ? " SUCCESS" : " FAILURE") << std::endl;
if(!res)
return false;
}
return true;
return ck::profiler::profile_grouped_gemm_impl<ADataType,
BDataType,
CDataType,
AccDataType,
ALayout,
BLayout,
CLayout>(
true, 1, false, 1, Ms, Ns, Ks, StrideAs, StrideBs, StrideCs);
}
} // anonymous namespace
int main()
{
std::vector<DeviceGroupedGemmPtr_> groupedGemmPtrs;
ck::tensor_operation::device::instance::
add_device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(groupedGemmPtrs);
bool res = true;
for(auto& gemmPtr : groupedGemmPtrs)
{
res &= TestGroupedGemm(gemmPtr);
}
res = res && TestGroupedGemm<Row, Row, Row>();
res = res && TestGroupedGemm<Row, Col, Row>();
res = res && TestGroupedGemm<Col, Row, Row>();
res = res && TestGroupedGemm<Col, Col, Row>();
std::cout << "TestGroupedGemm ..... " << (res ? "SUCCESS" : "FAILURE") << std::endl;
......
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