Skip to content

GitLab

Commit 0b11569f authored by Chao Liu's avatar Chao Liu
Browse files

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

parents e8d3a0fb fa9a0a5c
Hide whitespace changes
Inline Side-by-side
Showing with 1927 additions and 2046 deletions
test/batched_gemm/batched_gemm_util.hpp deleted 100644 → 0
View file @ e8d3a0fb
#ifndef BATCHED_GEMM_UTILS_HPP
#define BATCHED_GEMM_UTILS_HPP
#include "config.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
namespace ck {
namespace batched_gemm_util {
struct GemmParams
{
GemmParams()
: M(1024), N(1024), K(1024), StrideA(1024), StrideB(1024), StrideC(1024), alpha(1), beta(0)
{
}
ck::index_t M;
ck::index_t N;
ck::index_t K;
ck::index_t StrideA;
ck::index_t StrideB;
ck::index_t StrideC;
float alpha;
float beta;
};
template <typename BatchedGemmInstance,
typename ADataType,
typename BDataType,
typename CDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation>
void RunHostBatchedGemm(const Tensor<ADataType>& A,
const Tensor<BDataType>& B,
Tensor<CDataType>& C,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op)
{
auto ref_batched_gemm = BatchedGemmInstance{};
auto ref_invoker = ref_batched_gemm.MakeInvoker();
auto ref_argument =
ref_batched_gemm.MakeArgument(A, B, C, a_element_op, b_element_op, c_element_op);
ref_invoker.Run(ref_argument);
}
template <typename DeviceGemmPtr,
typename ADataType,
typename BDataType,
typename CDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation>
void RunDeviceBatchedGemm(DeviceGemmPtr& batched_gemm_ptr,
const ck::batched_gemm_util::GemmParams& params,
const Tensor<ADataType>& A,
const Tensor<BDataType>& B,
Tensor<CDataType>& C,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op)
{
DeviceMem a_g_m_k_device_buf(sizeof(ADataType) * A.mDesc.GetElementSpace());
DeviceMem b_g_k_n_device_buf(sizeof(BDataType) * B.mDesc.GetElementSpace());
DeviceMem c_g_m_n_device_buf(sizeof(CDataType) * C.mDesc.GetElementSpace());
a_g_m_k_device_buf.ToDevice(A.mData.data());
b_g_k_n_device_buf.ToDevice(B.mData.data());
const auto batch_count = A.mDesc.GetLengths()[0];
auto invoker_ptr = batched_gemm_ptr->MakeInvokerPointer();
auto argument_ptr = batched_gemm_ptr->MakeArgumentPointer(
static_cast<ADataType*>(a_g_m_k_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_g_k_n_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_g_m_n_device_buf.GetDeviceBuffer()),
params.M,
params.N,
params.K,
params.StrideA,
params.StrideB,
params.StrideC,
a_element_op,
b_element_op,
c_element_op,
batch_count);
if(!batched_gemm_ptr->IsSupportedArgument(argument_ptr.get()))
{
throw std::runtime_error(
"wrong! device_gemm with the specified compilation parameters does "
"not support this GEMM problem");
}
invoker_ptr->Run(argument_ptr.get());
c_g_m_n_device_buf.FromDevice(C.mData.data());
}
} // namespace batched_gemm_util
} // namespace ck
#endif
test/batched_gemm_reduce/batched_gemm_reduce_fp16.cpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include "profiler/include/profile_batched_gemm_reduce_impl.hpp"
......
test/block_to_ctile_map/test_block_to_ctile_map.cpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <vector>
#include <gtest/gtest.h>
......
test/conv2d_bwd_data/conv2d_bwd_data.cpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "config.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
......
test/conv2d_bwd_weight/conv2d_bwd_weight.cpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
......
test/conv_util/conv_util.cpp
View file @ 0b11569f
#include <iostream>
#include <string>
#include <vector>
#include <gtest/gtest.h>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/conv_util.hpp"
namespace {
class TestConvUtil : public ::testing::Test
{
public:
void SetNDParams(std::size_t ndims)
{
conv_params.num_dim_spatial_ = ndims;
conv_params.filter_spatial_lengths_ = std::vector<ck::index_t>(ndims, 3);
conv_params.input_spatial_lengths_ = std::vector<ck::index_t>(ndims, 71);
conv_params.conv_filter_strides_ = std::vector<ck::index_t>(ndims, 2);
conv_params.conv_filter_dilations_ = std::vector<ck::index_t>(ndims, 1);
conv_params.input_left_pads_ = std::vector<ck::index_t>(ndims, 1);
conv_params.input_right_pads_ = std::vector<ck::index_t>(ndims, 1);
}
protected:
// ------- default 2D -------
// input NCHW {128,192,71,71},
// weights KCYX {256,192,3,3},
// stride {2,2},
// dilations {1,1},
// padding {{1,1}, {1,1}}
ck::utils::conv::ConvParams conv_params;
};
} // namespace
TEST_F(TestConvUtil, ConvParamsGetOutputSpatialLengths2D)
{
ck::utils::conv::ConvParams conv_params;
std::vector<ck::index_t> out_spatial_len = conv_params.GetOutputSpatialLengths();
EXPECT_TRUE(ck::utils::check_err(out_spatial_len,
std::vector<ck::index_t>{36, 36},
"Error: ConvParams 2D default constructor."));
conv_params.conv_filter_strides_ = std::vector<ck::index_t>{1, 1};
out_spatial_len = conv_params.GetOutputSpatialLengths();
EXPECT_TRUE(ck::utils::check_err(
out_spatial_len, std::vector<ck::index_t>{71, 71}, "Error: ConvParams 2D stride {1,1}."));
conv_params.conv_filter_strides_ = std::vector<ck::index_t>{2, 2};
conv_params.input_left_pads_ = std::vector<ck::index_t>{2, 2};
conv_params.input_right_pads_ = std::vector<ck::index_t>{2, 2};
out_spatial_len = conv_params.GetOutputSpatialLengths();
EXPECT_TRUE(ck::utils::check_err(out_spatial_len,
std::vector<ck::index_t>{37, 37},
"Error: ConvParams 2D padding left/right {2,2}."));
conv_params.conv_filter_dilations_ = std::vector<ck::index_t>{2, 2};
out_spatial_len = conv_params.GetOutputSpatialLengths();
EXPECT_TRUE(ck::utils::check_err(
out_spatial_len, std::vector<ck::index_t>{36, 36}, "Error: ConvParams 2D dilation {2,2}."));
conv_params.conv_filter_strides_ = std::vector<ck::index_t>{3, 3};
conv_params.input_left_pads_ = std::vector<ck::index_t>{1, 1};
conv_params.input_right_pads_ = std::vector<ck::index_t>{1, 1};
conv_params.conv_filter_dilations_ = std::vector<ck::index_t>{2, 2};
out_spatial_len = conv_params.GetOutputSpatialLengths();
EXPECT_TRUE(
ck::utils::check_err(out_spatial_len,
std::vector<ck::index_t>{23, 23},
"Error: ConvParams 2D strides{3,3}, padding {1,1}, dilations {2,2}."));
}
TEST_F(TestConvUtil, ConvParamsGetOutputSpatialLengths1D)
{
SetNDParams(1);
std::vector<ck::index_t> out_spatial_len = conv_params.GetOutputSpatialLengths();
EXPECT_TRUE(ck::utils::check_err(
out_spatial_len, std::vector<ck::index_t>{36}, "Error: ConvParams 1D."));
conv_params.conv_filter_strides_ = std::vector<ck::index_t>{1};
out_spatial_len = conv_params.GetOutputSpatialLengths();
EXPECT_TRUE(ck::utils::check_err(
out_spatial_len, std::vector<ck::index_t>{71}, "Error: ConvParams 1D stride {1}."));
conv_params.conv_filter_strides_ = std::vector<ck::index_t>{2};
conv_params.input_left_pads_ = std::vector<ck::index_t>{2};
conv_params.input_right_pads_ = std::vector<ck::index_t>{2};
out_spatial_len = conv_params.GetOutputSpatialLengths();
EXPECT_TRUE(ck::utils::check_err(out_spatial_len,
std::vector<ck::index_t>{37},
"Error: ConvParams 1D padding left/right {2}."));
conv_params.conv_filter_dilations_ = std::vector<ck::index_t>{2};
out_spatial_len = conv_params.GetOutputSpatialLengths();
EXPECT_TRUE(ck::utils::check_err(
out_spatial_len, std::vector<ck::index_t>{36}, "Error: ConvParams 1D dilation {2}."));
conv_params.conv_filter_strides_ = std::vector<ck::index_t>{3};
conv_params.input_left_pads_ = std::vector<ck::index_t>{1};
conv_params.input_right_pads_ = std::vector<ck::index_t>{1};
conv_params.conv_filter_dilations_ = std::vector<ck::index_t>{2};
out_spatial_len = conv_params.GetOutputSpatialLengths();
EXPECT_TRUE(
ck::utils::check_err(out_spatial_len,
std::vector<ck::index_t>{23},
"Error: ConvParams 1D strides{3}, padding {1}, dilations {2}."));
}
TEST_F(TestConvUtil, ConvParamsGetOutputSpatialLengths3D)
{
SetNDParams(3);
std::vector<ck::index_t> out_spatial_len = conv_params.GetOutputSpatialLengths();
EXPECT_TRUE(ck::utils::check_err(
out_spatial_len, std::vector<ck::index_t>{36, 36, 36}, "Error: ConvParams 3D."));
conv_params.conv_filter_strides_ = std::vector<ck::index_t>{1, 1, 1};
out_spatial_len = conv_params.GetOutputSpatialLengths();
EXPECT_TRUE(ck::utils::check_err(out_spatial_len,
std::vector<ck::index_t>{71, 71, 71},
"Error: ConvParams 3D stride {1, 1, 1}."));
conv_params.conv_filter_strides_ = std::vector<ck::index_t>{2, 2, 2};
conv_params.input_left_pads_ = std::vector<ck::index_t>{2, 2, 2};
conv_params.input_right_pads_ = std::vector<ck::index_t>{2, 2, 2};
out_spatial_len = conv_params.GetOutputSpatialLengths();
EXPECT_TRUE(ck::utils::check_err(out_spatial_len,
std::vector<ck::index_t>{37, 37, 37},
"Error: ConvParams 3D padding left/right {2, 2, 2}."));
conv_params.conv_filter_dilations_ = std::vector<ck::index_t>{2, 2, 2};
out_spatial_len = conv_params.GetOutputSpatialLengths();
EXPECT_TRUE(ck::utils::check_err(out_spatial_len,
std::vector<ck::index_t>{36, 36, 36},
"Error: ConvParams 3D dilation {2, 2, 2}."));
conv_params.conv_filter_strides_ = std::vector<ck::index_t>{3, 3, 3};
conv_params.input_left_pads_ = std::vector<ck::index_t>{1, 1, 1};
conv_params.input_right_pads_ = std::vector<ck::index_t>{1, 1, 1};
conv_params.conv_filter_dilations_ = std::vector<ck::index_t>{2, 2, 2};
out_spatial_len = conv_params.GetOutputSpatialLengths();
EXPECT_TRUE(ck::utils::check_err(
out_spatial_len,
std::vector<ck::index_t>{23, 23, 23},
"Error: ConvParams 3D strides{3, 3, 3}, padding {1, 1, 1}, dilations {2, 2, 2}."));
}
TEST(ConvUtil, GetHostTensorDescriptor)
{
namespace tl = ck::tensor_layout::convolution;
std::vector<std::size_t> dims{2, 3, 4, 5};
HostTensorDescriptor h = ck::utils::conv::get_host_tensor_descriptor(dims, tl::NHWC{});
EXPECT_TRUE(ck::utils::check_err(
h.GetLengths(), {2, 3, 4, 5}, "Error: wrong NHWC dimensions lengths!"));
EXPECT_TRUE(ck::utils::check_err(
h.GetStrides(), {3 * 4 * 5, 1, 3 * 5, 3}, "Error: wrong NHWC dimensions strides!"));
h = ck::utils::conv::get_host_tensor_descriptor(dims, tl::NCHW{});
EXPECT_TRUE(ck::utils::check_err(
h.GetLengths(), {2, 3, 4, 5}, "Error: wrong NCHW dimensions lengths!"));
EXPECT_TRUE(ck::utils::check_err(
h.GetStrides(), {3 * 4 * 5, 4 * 5, 5, 1}, "Error: wrong NCHW dimensions strides!"));
dims = std::vector<std::size_t>{2, 3, 4};
h = ck::utils::conv::get_host_tensor_descriptor(dims, tl::NWC{});
EXPECT_TRUE(
ck::utils::check_err(h.GetLengths(), {2, 3, 4}, "Error: wrong NWC dimensions lengths!"));
EXPECT_TRUE(ck::utils::check_err(
h.GetStrides(), {3 * 4, 1, 3}, "Error: wrong NWC dimensions strides!"));
h = ck::utils::conv::get_host_tensor_descriptor(dims, tl::NCW{});
EXPECT_TRUE(
ck::utils::check_err(h.GetLengths(), {2, 3, 4}, "Error: wrong NCW dimensions lengths!"));
EXPECT_TRUE(ck::utils::check_err(
h.GetStrides(), {3 * 4, 4, 1}, "Error: wrong NCW dimensions strides!"));
dims = std::vector<std::size_t>{2, 3, 4, 5, 6};
h = ck::utils::conv::get_host_tensor_descriptor(dims, tl::NDHWC{});
EXPECT_TRUE(
ck::utils::check_err(h.GetLengths(), dims, "Error: wrong NDHWC dimensions lengths!"));
EXPECT_TRUE(ck::utils::check_err(h.GetStrides(),
{3 * 4 * 5 * 6, // N
1, // C
3 * 5 * 6, // D
3 * 6, // H
3}, // W
"Error: wrong NDHWC dimensions strides!"));
h = ck::utils::conv::get_host_tensor_descriptor(dims, tl::NCDHW{});
EXPECT_TRUE(
ck::utils::check_err(h.GetLengths(), dims, "Error: wrong NCDHW dimensions lengths!"));
EXPECT_TRUE(ck::utils::check_err(h.GetStrides(),
{3 * 4 * 5 * 6, // N
4 * 5 * 6, // C
5 * 6, // D
6, // H
1}, // W
"Error: wrong NCDHW dimensions strides!"));
}
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <string>
#include <vector>
#include <gtest/gtest.h>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/conv_util.hpp"
namespace {
class TestConvUtil : public ::testing::Test
{
public:
void SetNDParams(std::size_t ndims)
{
conv_params.num_dim_spatial_ = ndims;
conv_params.filter_spatial_lengths_ = std::vector<ck::index_t>(ndims, 3);
conv_params.input_spatial_lengths_ = std::vector<ck::index_t>(ndims, 71);
conv_params.conv_filter_strides_ = std::vector<ck::index_t>(ndims, 2);
conv_params.conv_filter_dilations_ = std::vector<ck::index_t>(ndims, 1);
conv_params.input_left_pads_ = std::vector<ck::index_t>(ndims, 1);
conv_params.input_right_pads_ = std::vector<ck::index_t>(ndims, 1);
}
protected:
// ------- default 2D -------
// input NCHW {128,192,71,71},
// weights KCYX {256,192,3,3},
// stride {2,2},
// dilations {1,1},
// padding {{1,1}, {1,1}}
ck::utils::conv::ConvParams conv_params;
};
} // namespace
TEST_F(TestConvUtil, ConvParamsGetOutputSpatialLengths2D)
{
ck::utils::conv::ConvParams conv_params;
std::vector<ck::index_t> out_spatial_len = conv_params.GetOutputSpatialLengths();
EXPECT_TRUE(ck::utils::check_err(out_spatial_len,
std::vector<ck::index_t>{36, 36},
"Error: ConvParams 2D default constructor."));
conv_params.conv_filter_strides_ = std::vector<ck::index_t>{1, 1};
out_spatial_len = conv_params.GetOutputSpatialLengths();
EXPECT_TRUE(ck::utils::check_err(
out_spatial_len, std::vector<ck::index_t>{71, 71}, "Error: ConvParams 2D stride {1,1}."));
conv_params.conv_filter_strides_ = std::vector<ck::index_t>{2, 2};
conv_params.input_left_pads_ = std::vector<ck::index_t>{2, 2};
conv_params.input_right_pads_ = std::vector<ck::index_t>{2, 2};
out_spatial_len = conv_params.GetOutputSpatialLengths();
EXPECT_TRUE(ck::utils::check_err(out_spatial_len,
std::vector<ck::index_t>{37, 37},
"Error: ConvParams 2D padding left/right {2,2}."));
conv_params.conv_filter_dilations_ = std::vector<ck::index_t>{2, 2};
out_spatial_len = conv_params.GetOutputSpatialLengths();
EXPECT_TRUE(ck::utils::check_err(
out_spatial_len, std::vector<ck::index_t>{36, 36}, "Error: ConvParams 2D dilation {2,2}."));
conv_params.conv_filter_strides_ = std::vector<ck::index_t>{3, 3};
conv_params.input_left_pads_ = std::vector<ck::index_t>{1, 1};
conv_params.input_right_pads_ = std::vector<ck::index_t>{1, 1};
conv_params.conv_filter_dilations_ = std::vector<ck::index_t>{2, 2};
out_spatial_len = conv_params.GetOutputSpatialLengths();
EXPECT_TRUE(
ck::utils::check_err(out_spatial_len,
std::vector<ck::index_t>{23, 23},
"Error: ConvParams 2D strides{3,3}, padding {1,1}, dilations {2,2}."));
}
TEST_F(TestConvUtil, ConvParamsGetOutputSpatialLengths1D)
{
SetNDParams(1);
std::vector<ck::index_t> out_spatial_len = conv_params.GetOutputSpatialLengths();
EXPECT_TRUE(ck::utils::check_err(
out_spatial_len, std::vector<ck::index_t>{36}, "Error: ConvParams 1D."));
conv_params.conv_filter_strides_ = std::vector<ck::index_t>{1};
out_spatial_len = conv_params.GetOutputSpatialLengths();
EXPECT_TRUE(ck::utils::check_err(
out_spatial_len, std::vector<ck::index_t>{71}, "Error: ConvParams 1D stride {1}."));
conv_params.conv_filter_strides_ = std::vector<ck::index_t>{2};
conv_params.input_left_pads_ = std::vector<ck::index_t>{2};
conv_params.input_right_pads_ = std::vector<ck::index_t>{2};
out_spatial_len = conv_params.GetOutputSpatialLengths();
EXPECT_TRUE(ck::utils::check_err(out_spatial_len,
std::vector<ck::index_t>{37},
"Error: ConvParams 1D padding left/right {2}."));
conv_params.conv_filter_dilations_ = std::vector<ck::index_t>{2};
out_spatial_len = conv_params.GetOutputSpatialLengths();
EXPECT_TRUE(ck::utils::check_err(
out_spatial_len, std::vector<ck::index_t>{36}, "Error: ConvParams 1D dilation {2}."));
conv_params.conv_filter_strides_ = std::vector<ck::index_t>{3};
conv_params.input_left_pads_ = std::vector<ck::index_t>{1};
conv_params.input_right_pads_ = std::vector<ck::index_t>{1};
conv_params.conv_filter_dilations_ = std::vector<ck::index_t>{2};
out_spatial_len = conv_params.GetOutputSpatialLengths();
EXPECT_TRUE(
ck::utils::check_err(out_spatial_len,
std::vector<ck::index_t>{23},
"Error: ConvParams 1D strides{3}, padding {1}, dilations {2}."));
}
TEST_F(TestConvUtil, ConvParamsGetOutputSpatialLengths3D)
{
SetNDParams(3);
std::vector<ck::index_t> out_spatial_len = conv_params.GetOutputSpatialLengths();
EXPECT_TRUE(ck::utils::check_err(
out_spatial_len, std::vector<ck::index_t>{36, 36, 36}, "Error: ConvParams 3D."));
conv_params.conv_filter_strides_ = std::vector<ck::index_t>{1, 1, 1};
out_spatial_len = conv_params.GetOutputSpatialLengths();
EXPECT_TRUE(ck::utils::check_err(out_spatial_len,
std::vector<ck::index_t>{71, 71, 71},
"Error: ConvParams 3D stride {1, 1, 1}."));
conv_params.conv_filter_strides_ = std::vector<ck::index_t>{2, 2, 2};
conv_params.input_left_pads_ = std::vector<ck::index_t>{2, 2, 2};
conv_params.input_right_pads_ = std::vector<ck::index_t>{2, 2, 2};
out_spatial_len = conv_params.GetOutputSpatialLengths();
EXPECT_TRUE(ck::utils::check_err(out_spatial_len,
std::vector<ck::index_t>{37, 37, 37},
"Error: ConvParams 3D padding left/right {2, 2, 2}."));
conv_params.conv_filter_dilations_ = std::vector<ck::index_t>{2, 2, 2};
out_spatial_len = conv_params.GetOutputSpatialLengths();
EXPECT_TRUE(ck::utils::check_err(out_spatial_len,
std::vector<ck::index_t>{36, 36, 36},
"Error: ConvParams 3D dilation {2, 2, 2}."));
conv_params.conv_filter_strides_ = std::vector<ck::index_t>{3, 3, 3};
conv_params.input_left_pads_ = std::vector<ck::index_t>{1, 1, 1};
conv_params.input_right_pads_ = std::vector<ck::index_t>{1, 1, 1};
conv_params.conv_filter_dilations_ = std::vector<ck::index_t>{2, 2, 2};
out_spatial_len = conv_params.GetOutputSpatialLengths();
EXPECT_TRUE(ck::utils::check_err(
out_spatial_len,
std::vector<ck::index_t>{23, 23, 23},
"Error: ConvParams 3D strides{3, 3, 3}, padding {1, 1, 1}, dilations {2, 2, 2}."));
}
TEST(ConvUtil, GetHostTensorDescriptor)
{
namespace tl = ck::tensor_layout::convolution;
std::vector<std::size_t> dims{2, 3, 4, 5};
HostTensorDescriptor h = ck::utils::conv::get_host_tensor_descriptor(dims, tl::NHWC{});
EXPECT_TRUE(ck::utils::check_err(
h.GetLengths(), {2, 3, 4, 5}, "Error: wrong NHWC dimensions lengths!"));
EXPECT_TRUE(ck::utils::check_err(
h.GetStrides(), {3 * 4 * 5, 1, 3 * 5, 3}, "Error: wrong NHWC dimensions strides!"));
h = ck::utils::conv::get_host_tensor_descriptor(dims, tl::NCHW{});
EXPECT_TRUE(ck::utils::check_err(
h.GetLengths(), {2, 3, 4, 5}, "Error: wrong NCHW dimensions lengths!"));
EXPECT_TRUE(ck::utils::check_err(
h.GetStrides(), {3 * 4 * 5, 4 * 5, 5, 1}, "Error: wrong NCHW dimensions strides!"));
dims = std::vector<std::size_t>{2, 3, 4};
h = ck::utils::conv::get_host_tensor_descriptor(dims, tl::NWC{});
EXPECT_TRUE(
ck::utils::check_err(h.GetLengths(), {2, 3, 4}, "Error: wrong NWC dimensions lengths!"));
EXPECT_TRUE(ck::utils::check_err(
h.GetStrides(), {3 * 4, 1, 3}, "Error: wrong NWC dimensions strides!"));
h = ck::utils::conv::get_host_tensor_descriptor(dims, tl::NCW{});
EXPECT_TRUE(
ck::utils::check_err(h.GetLengths(), {2, 3, 4}, "Error: wrong NCW dimensions lengths!"));
EXPECT_TRUE(ck::utils::check_err(
h.GetStrides(), {3 * 4, 4, 1}, "Error: wrong NCW dimensions strides!"));
dims = std::vector<std::size_t>{2, 3, 4, 5, 6};
h = ck::utils::conv::get_host_tensor_descriptor(dims, tl::NDHWC{});
EXPECT_TRUE(
ck::utils::check_err(h.GetLengths(), dims, "Error: wrong NDHWC dimensions lengths!"));
EXPECT_TRUE(ck::utils::check_err(h.GetStrides(),
{3 * 4 * 5 * 6, // N
1, // C
3 * 5 * 6, // D
3 * 6, // H
3}, // W
"Error: wrong NDHWC dimensions strides!"));
h = ck::utils::conv::get_host_tensor_descriptor(dims, tl::NCDHW{});
EXPECT_TRUE(
ck::utils::check_err(h.GetLengths(), dims, "Error: wrong NCDHW dimensions lengths!"));
EXPECT_TRUE(ck::utils::check_err(h.GetStrides(),
{3 * 4 * 5 * 6, // N
4 * 5 * 6, // C
5 * 6, // D
6, // H
1}, // W
"Error: wrong NCDHW dimensions strides!"));
}
test/convnd_bwd_data/convnd_bwd_data.cpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
......
test/convnd_fwd/conv1d_fwd.cpp
View file @ 0b11569f
#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>());
}
// 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
View file @ 0b11569f
#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));
}
// 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
View file @ 0b11569f
#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>());
}
// 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
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <tuple>
......
test/gemm/gemm_dl_fp16.cpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <algorithm>
#include <cstdlib>
#include <iostream>
......
test/gemm/gemm_dl_fp32.cpp
View file @ 0b11569f
#include <algorithm>
#include <cstdlib>
#include <iostream>
#include <numeric>
#include <tuple>
#include <vector>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm.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"
#include "test/gemm/gemm_util.hpp"
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using DeviceGemmNoOpPtr =
ck::tensor_operation::device::DeviceGemmPtr<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 device_gemm_instance {
void add_device_gemm_dl_f32_f32_f32_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_dl_f32_f32_f32_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_dl_f32_f32_f32_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_dl_f32_f32_f32_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
} // namespace device_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
int main()
{
using ADataType = float;
using BDataType = float;
using CDataType = float;
using AccDataType = float;
using RowMajor = ck::tensor_layout::gemm::RowMajor;
using ColumnMajor = ck::tensor_layout::gemm::ColumnMajor;
bool res = true;
std::vector<DeviceGemmNoOpPtr> gemmPtrs;
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_dl_f32_f32_f32_km_kn_mn_instances(gemmPtrs);
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemm<DeviceGemmNoOpPtr,
ADataType,
BDataType,
CDataType,
AccDataType,
ColumnMajor,
RowMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
gemmPtrs.clear();
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_dl_f32_f32_f32_km_nk_mn_instances(gemmPtrs);
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemm<DeviceGemmNoOpPtr,
ADataType,
BDataType,
CDataType,
AccDataType,
ColumnMajor,
ColumnMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
gemmPtrs.clear();
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_dl_f32_f32_f32_mk_kn_mn_instances(gemmPtrs);
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemm<DeviceGemmNoOpPtr,
ADataType,
BDataType,
CDataType,
AccDataType,
RowMajor,
RowMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
gemmPtrs.clear();
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_dl_f32_f32_f32_mk_nk_mn_instances(gemmPtrs);
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemm<DeviceGemmNoOpPtr,
ADataType,
BDataType,
CDataType,
AccDataType,
RowMajor,
ColumnMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
std::cout << "TestGemm ..... " << (res ? "SUCCESS" : "FAILURE") << std::endl;
return res ? 0 : 1;
}
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <algorithm>
#include <cstdlib>
#include <iostream>
#include <numeric>
#include <tuple>
#include <vector>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm.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"
#include "test/gemm/gemm_util.hpp"
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using DeviceGemmNoOpPtr =
ck::tensor_operation::device::DeviceGemmPtr<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 device_gemm_instance {
void add_device_gemm_dl_f32_f32_f32_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_dl_f32_f32_f32_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_dl_f32_f32_f32_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_dl_f32_f32_f32_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
} // namespace device_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
int main()
{
using ADataType = float;
using BDataType = float;
using CDataType = float;
using AccDataType = float;
using RowMajor = ck::tensor_layout::gemm::RowMajor;
using ColumnMajor = ck::tensor_layout::gemm::ColumnMajor;
bool res = true;
std::vector<DeviceGemmNoOpPtr> gemmPtrs;
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_dl_f32_f32_f32_km_kn_mn_instances(gemmPtrs);
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemm<DeviceGemmNoOpPtr,
ADataType,
BDataType,
CDataType,
AccDataType,
ColumnMajor,
RowMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
gemmPtrs.clear();
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_dl_f32_f32_f32_km_nk_mn_instances(gemmPtrs);
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemm<DeviceGemmNoOpPtr,
ADataType,
BDataType,
CDataType,
AccDataType,
ColumnMajor,
ColumnMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
gemmPtrs.clear();
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_dl_f32_f32_f32_mk_kn_mn_instances(gemmPtrs);
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemm<DeviceGemmNoOpPtr,
ADataType,
BDataType,
CDataType,
AccDataType,
RowMajor,
RowMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
gemmPtrs.clear();
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_dl_f32_f32_f32_mk_nk_mn_instances(gemmPtrs);
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemm<DeviceGemmNoOpPtr,
ADataType,
BDataType,
CDataType,
AccDataType,
RowMajor,
ColumnMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
std::cout << "TestGemm ..... " << (res ? "SUCCESS" : "FAILURE") << std::endl;
return res ? 0 : 1;
}
test/gemm/gemm_dl_int8.cpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <algorithm>
#include <cstdlib>
#include <iostream>
......
test/gemm/gemm_util.hpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/ck.hpp"
......@@ -211,6 +214,11 @@ struct TestGemm
res = ck::utils::check_err(c_device.mData, c_host.mData);
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);
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);
......@@ -231,121 +239,5 @@ struct TestGemm
}
};
template <typename DeviceGemmPtr_,
typename ALayout,
typename BLayout,
typename CLayout,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation>
struct TestGemmBF16
{
using BF16 = ck::bhalf_t;
auto PrepareGemmTensorBF16(const ck::gemm_util::GemmParams& params)
{
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}));
}
};
// use fp32 host kernel to verify bf16 device kernel
Tensor<BF16> a_m_k_bf16(
f_host_tensor_descriptor(params.M, params.K, params.StrideA, ALayout{}));
Tensor<BF16> b_k_n_bf16(
f_host_tensor_descriptor(params.K, params.N, params.StrideB, BLayout{}));
Tensor<BF16> c_m_n_device_bf16(
f_host_tensor_descriptor(params.M, params.N, params.StrideC, CLayout{}));
Tensor<float> a_m_k_fp32(
f_host_tensor_descriptor(params.M, params.K, params.StrideA, ALayout{}));
Tensor<float> b_k_n_fp32(
f_host_tensor_descriptor(params.K, params.N, params.StrideB, BLayout{}));
Tensor<float> c_m_n_host_fp32(
f_host_tensor_descriptor(params.M, params.N, params.StrideC, CLayout{}));
Tensor<float> c_m_n_device_fp32(
f_host_tensor_descriptor(params.M, params.N, params.StrideC, CLayout{}));
a_m_k_bf16.GenerateTensorValue(GeneratorTensor_3<BF16>{-0.5, 0.5});
b_k_n_bf16.GenerateTensorValue(GeneratorTensor_3<BF16>{-0.5, 0.5});
bf16_to_f32_(a_m_k_bf16, a_m_k_fp32);
bf16_to_f32_(b_k_n_bf16, b_k_n_fp32);
return std::make_tuple(a_m_k_bf16,
b_k_n_bf16,
c_m_n_device_bf16,
a_m_k_fp32,
b_k_n_fp32,
c_m_n_host_fp32,
c_m_n_device_fp32);
}
auto operator()(DeviceGemmPtr_& gemmPtr)
{
// Arrange
ck::gemm_util::GemmParams params;
params.M = 1024;
params.N = 1024;
params.K = 1024;
params.StrideA = 1024;
params.StrideB = 1024;
params.StrideC = 1024;
auto host_tensors = PrepareGemmTensorBF16(params);
const Tensor<BF16>& a_bf16 = std::get<0>(host_tensors);
const Tensor<BF16>& b_bf16 = std::get<1>(host_tensors);
Tensor<BF16>& c_device_bf16 = std::get<2>(host_tensors);
Tensor<float>& a_fp32 = std::get<3>(host_tensors);
Tensor<float>& b_fp32 = std::get<4>(host_tensors);
Tensor<float>& c_host_fp32 = std::get<5>(host_tensors);
Tensor<float>& c_device_fp32 = std::get<6>(host_tensors);
auto a_element_op = AElementwiseOperation{};
auto b_element_op = BElementwiseOperation{};
auto c_element_op = CElementwiseOperation{};
// use fp32 host kernel to verify bf16 device kernel
using ReferenceGemmInstance =
ck::tensor_operation::host::ReferenceGemm<float,
float,
float,
float,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation>;
ck::gemm_util::RunHostGEMM<ReferenceGemmInstance>(
a_fp32, b_fp32, c_host_fp32, a_element_op, b_element_op, c_element_op);
// Act
ck::gemm_util::RunDeviceGEMM(gemmPtr,
params,
a_bf16,
b_bf16,
c_device_bf16,
a_element_op,
b_element_op,
c_element_op);
bf16_to_f32_(c_device_bf16, c_device_fp32);
// Assert
bool res = ck::utils::check_err(
c_device_fp32.mData, c_host_fp32.mData, "Error: incorrect results!", 1e-2f, 1e-3f);
std::cout << (res ? "SUCCESS" : "FAILURE") << std::endl;
return res;
};
};
} // namespace gemm_util
} // namespace ck
test/gemm/gemm_xdl_bf16.cpp
View file @ 0b11569f
#include <algorithm>
#include <cstdlib>
#include <iostream>
#include <numeric>
#include <tuple>
#include <vector>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm.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"
#include "test/gemm/gemm_util.hpp"
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using DeviceGemmNoOpPtr =
ck::tensor_operation::device::DeviceGemmPtr<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 device_gemm_instance {
void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instances(
std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_nk_mn_instances(
std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instances(
std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_kn_mn_instances(
std::vector<DeviceGemmNoOpPtr>&);
} // namespace device_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
int main()
{
using RowMajor = ck::tensor_layout::gemm::RowMajor;
using ColumnMajor = ck::tensor_layout::gemm::ColumnMajor;
bool res = true;
std::vector<DeviceGemmNoOpPtr> gemmPtrs;
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instances(gemmPtrs);
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemmBF16<DeviceGemmNoOpPtr,
ColumnMajor,
RowMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
gemmPtrs.clear();
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_nk_mn_instances(gemmPtrs);
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemmBF16<DeviceGemmNoOpPtr,
ColumnMajor,
ColumnMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
gemmPtrs.clear();
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_kn_mn_instances(gemmPtrs);
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemmBF16<DeviceGemmNoOpPtr,
RowMajor,
RowMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
gemmPtrs.clear();
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instances(gemmPtrs);
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemmBF16<DeviceGemmNoOpPtr,
RowMajor,
ColumnMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
std::cout << "TestGemm ..... " << (res ? "SUCCESS" : "FAILURE") << std::endl;
return res ? 0 : 1;
}
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <algorithm>
#include <cstdlib>
#include <iostream>
#include <numeric>
#include <tuple>
#include <vector>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm.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"
#include "test/gemm/gemm_util.hpp"
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using DeviceGemmNoOpPtr =
ck::tensor_operation::device::DeviceGemmPtr<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 device_gemm_instance {
void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instances(
std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_nk_mn_instances(
std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instances(
std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_kn_mn_instances(
std::vector<DeviceGemmNoOpPtr>&);
} // namespace device_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
int main()
{
using ADataType = ck::bhalf_t;
using BDataType = ck::bhalf_t;
using CDataType = ck::bhalf_t;
using AccDataType = float;
using RowMajor = ck::tensor_layout::gemm::RowMajor;
using ColumnMajor = ck::tensor_layout::gemm::ColumnMajor;
bool res = true;
std::vector<DeviceGemmNoOpPtr> gemmPtrs;
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instances(gemmPtrs);
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemm<DeviceGemmNoOpPtr,
ADataType,
BDataType,
CDataType,
AccDataType,
ColumnMajor,
RowMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
gemmPtrs.clear();
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_nk_mn_instances(gemmPtrs);
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemm<DeviceGemmNoOpPtr,
ADataType,
BDataType,
CDataType,
AccDataType,
ColumnMajor,
ColumnMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
gemmPtrs.clear();
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_kn_mn_instances(gemmPtrs);
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemm<DeviceGemmNoOpPtr,
ADataType,
BDataType,
CDataType,
AccDataType,
RowMajor,
RowMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
gemmPtrs.clear();
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instances(gemmPtrs);
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemm<DeviceGemmNoOpPtr,
ADataType,
BDataType,
CDataType,
AccDataType,
RowMajor,
ColumnMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
std::cout << "TestGemm ..... " << (res ? "SUCCESS" : "FAILURE") << std::endl;
return res ? 0 : 1;
}
test/gemm/gemm_xdl_fp16.cpp
View file @ 0b11569f
#include <algorithm>
#include <cstdlib>
#include <iostream>
#include <numeric>
#include <tuple>
#include <vector>
#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_gemm.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"
#include "test/gemm/gemm_util.hpp"
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using DeviceGemmNoOpPtr =
ck::tensor_operation::device::DeviceGemmPtr<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 device_gemm_instance {
void add_device_gemm_xdl_f16_f16_f16_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_f16_f16_f16_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instances(
std::vector<DeviceGemmNoOpPtr>&);
} // namespace device_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
int main()
{
using ADataType = ck::half_t;
using BDataType = ck::half_t;
using CDataType = ck::half_t;
using AccDataType = float;
using RowMajor = ck::tensor_layout::gemm::RowMajor;
using ColumnMajor = ck::tensor_layout::gemm::ColumnMajor;
bool res = true;
std::vector<DeviceGemmNoOpPtr> gemmPtrs;
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f16_f16_f16_km_kn_mn_instances(gemmPtrs);
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instances(gemmPtrs);
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(gemmPtrs);
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemm<DeviceGemmNoOpPtr,
ADataType,
BDataType,
CDataType,
AccDataType,
ColumnMajor,
RowMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
gemmPtrs.clear();
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f16_f16_f16_km_nk_mn_instances(gemmPtrs);
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instances(gemmPtrs);
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(gemmPtrs);
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemm<DeviceGemmNoOpPtr,
ADataType,
BDataType,
CDataType,
AccDataType,
ColumnMajor,
ColumnMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
gemmPtrs.clear();
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(gemmPtrs);
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances(gemmPtrs);
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(gemmPtrs);
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemm<DeviceGemmNoOpPtr,
ADataType,
BDataType,
CDataType,
AccDataType,
RowMajor,
RowMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
gemmPtrs.clear();
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(gemmPtrs);
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances(gemmPtrs);
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(gemmPtrs);
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instances(gemmPtrs);
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemm<DeviceGemmNoOpPtr,
ADataType,
BDataType,
CDataType,
AccDataType,
RowMajor,
ColumnMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
std::cout << "TestGemm ..... " << (res ? "SUCCESS" : "FAILURE") << std::endl;
return res ? 0 : 1;
}
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <algorithm>
#include <cstdlib>
#include <iostream>
#include <numeric>
#include <tuple>
#include <vector>
#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_gemm.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"
#include "test/gemm/gemm_util.hpp"
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using DeviceGemmNoOpPtr =
ck::tensor_operation::device::DeviceGemmPtr<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 device_gemm_instance {
void add_device_gemm_xdl_f16_f16_f16_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_f16_f16_f16_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
#if 0
void add_device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
#endif
void add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instances(
std::vector<DeviceGemmNoOpPtr>&);
} // namespace device_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
int main()
{
using ADataType = ck::half_t;
using BDataType = ck::half_t;
using CDataType = ck::half_t;
using AccDataType = float;
using RowMajor = ck::tensor_layout::gemm::RowMajor;
using ColumnMajor = ck::tensor_layout::gemm::ColumnMajor;
bool res = true;
std::vector<DeviceGemmNoOpPtr> gemmPtrs;
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f16_f16_f16_km_kn_mn_instances(gemmPtrs);
#if 0
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instances(gemmPtrs);
#endif
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(gemmPtrs);
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemm<DeviceGemmNoOpPtr,
ADataType,
BDataType,
CDataType,
AccDataType,
ColumnMajor,
RowMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
gemmPtrs.clear();
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f16_f16_f16_km_nk_mn_instances(gemmPtrs);
#if 0
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instances(gemmPtrs);
#endif
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(gemmPtrs);
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemm<DeviceGemmNoOpPtr,
ADataType,
BDataType,
CDataType,
AccDataType,
ColumnMajor,
ColumnMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
gemmPtrs.clear();
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(gemmPtrs);
#if 0
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances(gemmPtrs);
#endif
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(gemmPtrs);
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemm<DeviceGemmNoOpPtr,
ADataType,
BDataType,
CDataType,
AccDataType,
RowMajor,
RowMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
gemmPtrs.clear();
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(gemmPtrs);
#if 0
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances(gemmPtrs);
#endif
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(gemmPtrs);
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instances(gemmPtrs);
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemm<DeviceGemmNoOpPtr,
ADataType,
BDataType,
CDataType,
AccDataType,
RowMajor,
ColumnMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
std::cout << "TestGemm ..... " << (res ? "SUCCESS" : "FAILURE") << std::endl;
return res ? 0 : 1;
}
test/gemm/gemm_xdl_fp32.cpp
View file @ 0b11569f
#include <algorithm>
#include <cstdlib>
#include <iostream>
#include <numeric>
#include <tuple>
#include <vector>
#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_gemm.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"
#include "test/gemm/gemm_util.hpp"
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using DeviceGemmNoOpPtr =
ck::tensor_operation::device::DeviceGemmPtr<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 device_gemm_instance {
void add_device_gemm_xdl_f32_f32_f32_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_f32_f32_f32_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_f32_f32_f32_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_f32_f32_f32_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
} // namespace device_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
int main()
{
using ADataType = float;
using BDataType = float;
using CDataType = float;
using AccDataType = float;
using RowMajor = ck::tensor_layout::gemm::RowMajor;
using ColumnMajor = ck::tensor_layout::gemm::ColumnMajor;
bool res = true;
std::vector<DeviceGemmNoOpPtr> gemmPtrs;
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f32_f32_f32_km_kn_mn_instances(gemmPtrs);
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instances(gemmPtrs);
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instances(gemmPtrs);
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemm<DeviceGemmNoOpPtr,
ADataType,
BDataType,
CDataType,
AccDataType,
ColumnMajor,
RowMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
gemmPtrs.clear();
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f32_f32_f32_km_nk_mn_instances(gemmPtrs);
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instances(gemmPtrs);
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instances(gemmPtrs);
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemm<DeviceGemmNoOpPtr,
ADataType,
BDataType,
CDataType,
AccDataType,
ColumnMajor,
ColumnMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
gemmPtrs.clear();
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f32_f32_f32_mk_kn_mn_instances(gemmPtrs);
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instances(gemmPtrs);
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instances(gemmPtrs);
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemm<DeviceGemmNoOpPtr,
ADataType,
BDataType,
CDataType,
AccDataType,
RowMajor,
RowMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
gemmPtrs.clear();
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f32_f32_f32_mk_nk_mn_instances(gemmPtrs);
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instances(gemmPtrs);
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instances(gemmPtrs);
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemm<DeviceGemmNoOpPtr,
ADataType,
BDataType,
CDataType,
AccDataType,
RowMajor,
ColumnMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
std::cout << "TestGemm ..... " << (res ? "SUCCESS" : "FAILURE") << std::endl;
return res ? 0 : 1;
}
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <algorithm>
#include <cstdlib>
#include <iostream>
#include <numeric>
#include <tuple>
#include <vector>
#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_gemm.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"
#include "test/gemm/gemm_util.hpp"
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using DeviceGemmNoOpPtr =
ck::tensor_operation::device::DeviceGemmPtr<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 device_gemm_instance {
void add_device_gemm_xdl_f32_f32_f32_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_f32_f32_f32_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_f32_f32_f32_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_f32_f32_f32_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
#if 0
void add_device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
#endif
void add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
} // namespace device_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
int main()
{
using ADataType = float;
using BDataType = float;
using CDataType = float;
using AccDataType = float;
using RowMajor = ck::tensor_layout::gemm::RowMajor;
using ColumnMajor = ck::tensor_layout::gemm::ColumnMajor;
bool res = true;
std::vector<DeviceGemmNoOpPtr> gemmPtrs;
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f32_f32_f32_km_kn_mn_instances(gemmPtrs);
#if 0
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instances(gemmPtrs);
#endif
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instances(gemmPtrs);
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemm<DeviceGemmNoOpPtr,
ADataType,
BDataType,
CDataType,
AccDataType,
ColumnMajor,
RowMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
gemmPtrs.clear();
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f32_f32_f32_km_nk_mn_instances(gemmPtrs);
#if 0
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instances(gemmPtrs);
#endif
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instances(gemmPtrs);
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemm<DeviceGemmNoOpPtr,
ADataType,
BDataType,
CDataType,
AccDataType,
ColumnMajor,
ColumnMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
gemmPtrs.clear();
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f32_f32_f32_mk_kn_mn_instances(gemmPtrs);
#if 0
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instances(gemmPtrs);
#endif
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instances(gemmPtrs);
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemm<DeviceGemmNoOpPtr,
ADataType,
BDataType,
CDataType,
AccDataType,
RowMajor,
RowMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
gemmPtrs.clear();
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f32_f32_f32_mk_nk_mn_instances(gemmPtrs);
#if 0
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instances(gemmPtrs);
#endif
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instances(gemmPtrs);
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemm<DeviceGemmNoOpPtr,
ADataType,
BDataType,
CDataType,
AccDataType,
RowMajor,
ColumnMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
std::cout << "TestGemm ..... " << (res ? "SUCCESS" : "FAILURE") << std::endl;
return res ? 0 : 1;
}
test/gemm/gemm_xdl_fp64.cpp
View file @ 0b11569f
#include <algorithm>
#include <cstdlib>
#include <iostream>
#include <numeric>
#include <tuple>
#include <vector>
#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_gemm.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"
#include "test/gemm/gemm_util.hpp"
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using DeviceGemmNoOpPtr =
ck::tensor_operation::device::DeviceGemmPtr<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 device_gemm_instance {
void add_device_gemm_xdl_f64_f64_f64_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_f64_f64_f64_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_f64_f64_f64_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_f64_f64_f64_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
} // namespace device_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
inline std::string get_device_name()
{
hipDeviceProp_t props{};
int device;
auto status = hipGetDevice(&device);
if(status != hipSuccess)
{
return std::string();
}
status = hipGetDeviceProperties(&props, device);
if(status != hipSuccess)
{
return std::string();
}
const std::string name(props.gcnArchName);
return name;
}
int main()
{
if(get_device_name().find("gfx90a") == std::string::npos)
{
std::cout << "TestGemm ..... SUCCESS" << std::endl;
return 0;
}
using ADataType = double;
using BDataType = double;
using CDataType = double;
using AccDataType = double;
using RowMajor = ck::tensor_layout::gemm::RowMajor;
using ColumnMajor = ck::tensor_layout::gemm::ColumnMajor;
bool res = true;
std::vector<DeviceGemmNoOpPtr> gemmPtrs;
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f64_f64_f64_km_kn_mn_instances(gemmPtrs);
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemm<DeviceGemmNoOpPtr,
ADataType,
BDataType,
CDataType,
AccDataType,
ColumnMajor,
RowMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
gemmPtrs.clear();
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f64_f64_f64_km_nk_mn_instances(gemmPtrs);
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemm<DeviceGemmNoOpPtr,
ADataType,
BDataType,
CDataType,
AccDataType,
ColumnMajor,
ColumnMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
gemmPtrs.clear();
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f64_f64_f64_mk_kn_mn_instances(gemmPtrs);
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemm<DeviceGemmNoOpPtr,
ADataType,
BDataType,
CDataType,
AccDataType,
RowMajor,
RowMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
gemmPtrs.clear();
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f64_f64_f64_mk_nk_mn_instances(gemmPtrs);
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemm<DeviceGemmNoOpPtr,
ADataType,
BDataType,
CDataType,
AccDataType,
RowMajor,
ColumnMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
std::cout << "TestGemm ..... " << (res ? "SUCCESS" : "FAILURE") << std::endl;
return res ? 0 : 1;
}
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <algorithm>
#include <cstdlib>
#include <iostream>
#include <numeric>
#include <tuple>
#include <vector>
#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_gemm.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"
#include "test/gemm/gemm_util.hpp"
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using DeviceGemmNoOpPtr =
ck::tensor_operation::device::DeviceGemmPtr<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 device_gemm_instance {
void add_device_gemm_xdl_f64_f64_f64_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_f64_f64_f64_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_f64_f64_f64_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_f64_f64_f64_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
} // namespace device_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
inline std::string get_device_name()
{
hipDeviceProp_t props{};
int device;
auto status = hipGetDevice(&device);
if(status != hipSuccess)
{
return std::string();
}
status = hipGetDeviceProperties(&props, device);
if(status != hipSuccess)
{
return std::string();
}
const std::string name(props.gcnArchName);
return name;
}
int main()
{
if(get_device_name().find("gfx90a") == std::string::npos)
{
std::cout << "TestGemm ..... SUCCESS" << std::endl;
return 0;
}
using ADataType = double;
using BDataType = double;
using CDataType = double;
using AccDataType = double;
using RowMajor = ck::tensor_layout::gemm::RowMajor;
using ColumnMajor = ck::tensor_layout::gemm::ColumnMajor;
bool res = true;
std::vector<DeviceGemmNoOpPtr> gemmPtrs;
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f64_f64_f64_km_kn_mn_instances(gemmPtrs);
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemm<DeviceGemmNoOpPtr,
ADataType,
BDataType,
CDataType,
AccDataType,
ColumnMajor,
RowMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
gemmPtrs.clear();
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f64_f64_f64_km_nk_mn_instances(gemmPtrs);
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemm<DeviceGemmNoOpPtr,
ADataType,
BDataType,
CDataType,
AccDataType,
ColumnMajor,
ColumnMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
gemmPtrs.clear();
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f64_f64_f64_mk_kn_mn_instances(gemmPtrs);
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemm<DeviceGemmNoOpPtr,
ADataType,
BDataType,
CDataType,
AccDataType,
RowMajor,
RowMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
gemmPtrs.clear();
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f64_f64_f64_mk_nk_mn_instances(gemmPtrs);
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemm<DeviceGemmNoOpPtr,
ADataType,
BDataType,
CDataType,
AccDataType,
RowMajor,
ColumnMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
std::cout << "TestGemm ..... " << (res ? "SUCCESS" : "FAILURE") << std::endl;
return res ? 0 : 1;
}
test/gemm/gemm_xdl_int8.cpp
View file @ 0b11569f
#include <algorithm>
#include <cstdlib>
#include <iostream>
#include <numeric>
#include <tuple>
#include <vector>
#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_gemm.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"
#include "test/gemm/gemm_util.hpp"
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using DeviceGemmNoOpPtr =
ck::tensor_operation::device::DeviceGemmPtr<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 device_gemm_instance {
void add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
} // namespace device_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
int main()
{
using ADataType = int8_t;
using BDataType = int8_t;
using CDataType = int8_t;
using AccDataType = int32_t;
using RowMajor = ck::tensor_layout::gemm::RowMajor;
using ColumnMajor = ck::tensor_layout::gemm::ColumnMajor;
std::vector<DeviceGemmNoOpPtr> gemmPtrs;
bool res = true;
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instances(gemmPtrs);
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemm<DeviceGemmNoOpPtr,
ADataType,
BDataType,
CDataType,
AccDataType,
ColumnMajor,
RowMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
gemmPtrs.clear();
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instances(gemmPtrs);
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemm<DeviceGemmNoOpPtr,
ADataType,
BDataType,
CDataType,
AccDataType,
ColumnMajor,
ColumnMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
gemmPtrs.clear();
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instances(gemmPtrs);
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemm<DeviceGemmNoOpPtr,
ADataType,
BDataType,
CDataType,
AccDataType,
RowMajor,
RowMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
gemmPtrs.clear();
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instances(gemmPtrs);
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemm<DeviceGemmNoOpPtr,
ADataType,
BDataType,
CDataType,
AccDataType,
RowMajor,
ColumnMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
std::cout << "TestGemm ..... " << (res ? "SUCCESS" : "FAILURE") << std::endl;
return res ? 0 : 1;
}
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <algorithm>
#include <cstdlib>
#include <iostream>
#include <numeric>
#include <tuple>
#include <vector>
#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_gemm.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"
#include "test/gemm/gemm_util.hpp"
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using DeviceGemmNoOpPtr =
ck::tensor_operation::device::DeviceGemmPtr<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 device_gemm_instance {
void add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
} // namespace device_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
int main()
{
using ADataType = int8_t;
using BDataType = int8_t;
using CDataType = int8_t;
using AccDataType = int32_t;
using RowMajor = ck::tensor_layout::gemm::RowMajor;
using ColumnMajor = ck::tensor_layout::gemm::ColumnMajor;
std::vector<DeviceGemmNoOpPtr> gemmPtrs;
bool res = true;
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instances(gemmPtrs);
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemm<DeviceGemmNoOpPtr,
ADataType,
BDataType,
CDataType,
AccDataType,
ColumnMajor,
RowMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
gemmPtrs.clear();
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instances(gemmPtrs);
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemm<DeviceGemmNoOpPtr,
ADataType,
BDataType,
CDataType,
AccDataType,
ColumnMajor,
ColumnMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
gemmPtrs.clear();
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instances(gemmPtrs);
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemm<DeviceGemmNoOpPtr,
ADataType,
BDataType,
CDataType,
AccDataType,
RowMajor,
RowMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
gemmPtrs.clear();
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instances(gemmPtrs);
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemm<DeviceGemmNoOpPtr,
ADataType,
BDataType,
CDataType,
AccDataType,
RowMajor,
ColumnMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
std::cout << "TestGemm ..... " << (res ? "SUCCESS" : "FAILURE") << std::endl;
return res ? 0 : 1;
}
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