devtools/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2026-04-10 11:34:33 +08:00
Code Issues Packages Projects Releases Wiki Activity

Adding sycl backend for TensorPadding.h; disbaling __unit128 for sycl in TensorIntDiv.h; disabling cashsize for sycl in tensorDeviceDefault.h; adding sycl backend for StrideSliceOP ; removing sycl compiler warning for creating an array of size 0 in CXX11Meta.h; cleaning up the sycl backend code.

Browse Source
This commit is contained in:
Mehdi Goli
2016-12-01 13:02:27 +00:00
parent a70393fd02
commit 79aa2b784e
17 changed files with 506 additions and 81 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
unsupported/test/CMakeLists.txt
View File

@@ -148,6 +148,7 @@ if(EIGEN_TEST_CXX11)
ei_add_test_sycl(cxx11_tensor_reduction_sycl "-std=c++11")
ei_add_test_sycl(cxx11_tensor_morphing_sycl "-std=c++11")
ei_add_test_sycl(cxx11_tensor_shuffling_sycl "-std=c++11")
ei_add_test_sycl(cxx11_tensor_padding_sycl "-std=c++11")
ei_add_test_sycl(cxx11_tensor_builtins_sycl "-std=c++11")
endif(EIGEN_TEST_SYCL)
# It should be safe to always run these tests as there is some fallback code for

49
unsupported/test/cxx11_tensor_morphing_sycl.cpp
View File

@@ -180,6 +180,53 @@ static void test_simple_slice(const Eigen::SyclDevice &sycl_device)
sycl_device.deallocate(gpu_data3);
}
template<typename DataType, int DataLayout, typename IndexType>
static void test_strided_slice_write_sycl(const Eigen::SyclDevice& sycl_device)
{
typedef Tensor<DataType, 2, DataLayout, IndexType> Tensor2f;
typedef Eigen::DSizes<IndexType, 2> Index2;
IndexType sizeDim1 = 7L;
IndexType sizeDim2 = 11L;
array<IndexType, 2> tensorRange = {{sizeDim1, sizeDim2}};
Tensor<DataType, 2, DataLayout, IndexType> tensor(tensorRange),tensor2(tensorRange);
IndexType sliceDim1 = 2;
IndexType sliceDim2 = 3;
array<IndexType, 2> sliceRange = {{sliceDim1, sliceDim2}};
Tensor2f slice(sliceRange);
Index2 strides(1L,1L);
Index2 indicesStart(3L,4L);
Index2 indicesStop(5L,7L);
Index2 lengths(2L,3L);
DataType* gpu_data1 = static_cast<DataType*>(sycl_device.allocate(tensor.size()*sizeof(DataType)));
DataType* gpu_data2 = static_cast<DataType*>(sycl_device.allocate(tensor2.size()*sizeof(DataType)));
DataType* gpu_data3 = static_cast<DataType*>(sycl_device.allocate(slice.size()*sizeof(DataType)));
TensorMap<Tensor<DataType, 2,DataLayout,IndexType>> gpu1(gpu_data1, tensorRange);
TensorMap<Tensor<DataType, 2,DataLayout,IndexType>> gpu2(gpu_data2, tensorRange);
TensorMap<Tensor<DataType, 2,DataLayout,IndexType>> gpu3(gpu_data3, sliceRange);
tensor.setRandom();
sycl_device.memcpyHostToDevice(gpu_data1, tensor.data(),(tensor.size())*sizeof(DataType));
gpu2.device(sycl_device)=gpu1;
slice.setRandom();
sycl_device.memcpyHostToDevice(gpu_data3, slice.data(),(slice.size())*sizeof(DataType));
gpu1.slice(indicesStart,lengths).device(sycl_device)=gpu3;
gpu2.stridedSlice(indicesStart,indicesStop,strides).device(sycl_device)=gpu3;
sycl_device.memcpyDeviceToHost(tensor.data(), gpu_data1,(tensor.size())*sizeof(DataType));
sycl_device.memcpyDeviceToHost(tensor2.data(), gpu_data2,(tensor2.size())*sizeof(DataType));
for(int i=0;i<sizeDim1;i++) for(int j=0;j<sizeDim2;j++){
VERIFY_IS_EQUAL(tensor(i,j), tensor2(i,j));
}
sycl_device.deallocate(gpu_data1);
sycl_device.deallocate(gpu_data2);
sycl_device.deallocate(gpu_data3);
}
template<typename DataType, typename dev_Selector> void sycl_morphing_test_per_device(dev_Selector s){
QueueInterface queueInterface(s);
auto sycl_device = Eigen::SyclDevice(&queueInterface);
@@ -189,6 +236,8 @@ template<typename DataType, typename dev_Selector> void sycl_morphing_test_per_d
test_simple_reshape<DataType, ColMajor>(sycl_device);
test_reshape_as_lvalue<DataType, RowMajor>(sycl_device);
test_reshape_as_lvalue<DataType, ColMajor>(sycl_device);
test_strided_slice_write_sycl<DataType, ColMajor, int64_t>(sycl_device);
test_strided_slice_write_sycl<DataType, RowMajor, int64_t>(sycl_device);
}
void test_cxx11_tensor_morphing_sycl()
{

161
unsupported/test/cxx11_tensor_padding_sycl.cpp Normal file
View File

@@ -0,0 +1,161 @@
// This file is part of Eigen, a lightweight C++ template library
// for linear algebra.
//
// Copyright (C) 2016
// Mehdi Goli Codeplay Software Ltd.
// Ralph Potter Codeplay Software Ltd.
// Luke Iwanski Codeplay Software Ltd.
// Contact: <eigen@codeplay.com>
// Benoit Steiner <benoit.steiner.goog@gmail.com>
//
// This Source Code Form is subject to the terms of the Mozilla
// Public License v. 2.0. If a copy of the MPL was not distributed
// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
#define EIGEN_TEST_NO_LONGDOUBLE
#define EIGEN_TEST_NO_COMPLEX
#define EIGEN_TEST_FUNC cxx11_tensor_padding_sycl
#define EIGEN_DEFAULT_DENSE_INDEX_TYPE int
#define EIGEN_USE_SYCL
#include "main.h"
#include <unsupported/Eigen/CXX11/Tensor>
using Eigen::array;
using Eigen::SyclDevice;
using Eigen::Tensor;
using Eigen::TensorMap;
template<typename DataType, int DataLayout, typename IndexType>
static void test_simple_padding(const Eigen::SyclDevice& sycl_device)
{
IndexType sizeDim1 = 2;
IndexType sizeDim2 = 3;
IndexType sizeDim3 = 5;
IndexType sizeDim4 = 7;
array<IndexType, 4> tensorRange = {{sizeDim1, sizeDim2, sizeDim3, sizeDim4}};
Tensor<DataType, 4, DataLayout, IndexType> tensor(tensorRange);
tensor.setRandom();
array<std::pair<IndexType, IndexType>, 4> paddings;
paddings[0] = std::make_pair(0, 0);
paddings[1] = std::make_pair(2, 1);
paddings[2] = std::make_pair(3, 4);
paddings[3] = std::make_pair(0, 0);
IndexType padedSizeDim1 = 2;
IndexType padedSizeDim2 = 6;
IndexType padedSizeDim3 = 12;
IndexType padedSizeDim4 = 7;
array<IndexType, 4> padedtensorRange = {{padedSizeDim1, padedSizeDim2, padedSizeDim3, padedSizeDim4}};
Tensor<DataType, 4, DataLayout, IndexType> padded(padedtensorRange);
DataType* gpu_data1 = static_cast<DataType*>(sycl_device.allocate(tensor.size()*sizeof(DataType)));
DataType* gpu_data2 = static_cast<DataType*>(sycl_device.allocate(padded.size()*sizeof(DataType)));
TensorMap<Tensor<DataType, 4,DataLayout,IndexType>> gpu1(gpu_data1, tensorRange);
TensorMap<Tensor<DataType, 4,DataLayout,IndexType>> gpu2(gpu_data2, padedtensorRange);
VERIFY_IS_EQUAL(padded.dimension(0), 2+0);
VERIFY_IS_EQUAL(padded.dimension(1), 3+3);
VERIFY_IS_EQUAL(padded.dimension(2), 5+7);
VERIFY_IS_EQUAL(padded.dimension(3), 7+0);
sycl_device.memcpyHostToDevice(gpu_data1, tensor.data(),(tensor.size())*sizeof(DataType));
gpu2.device(sycl_device)=gpu1.pad(paddings);
sycl_device.memcpyDeviceToHost(padded.data(), gpu_data2,(padded.size())*sizeof(DataType));
for (int i = 0; i < padedSizeDim1; ++i) {
for (int j = 0; j < padedSizeDim2; ++j) {
for (int k = 0; k < padedSizeDim3; ++k) {
for (int l = 0; l < padedSizeDim4; ++l) {
if (j >= 2 && j < 5 && k >= 3 && k < 8) {
VERIFY_IS_EQUAL(padded(i,j,k,l), tensor(i,j-2,k-3,l));
} else {
VERIFY_IS_EQUAL(padded(i,j,k,l), 0.0f);
}
}
}
}
}
sycl_device.deallocate(gpu_data1);
sycl_device.deallocate(gpu_data2);
}
template<typename DataType, int DataLayout, typename IndexType>
static void test_padded_expr(const Eigen::SyclDevice& sycl_device)
{
IndexType sizeDim1 = 2;
IndexType sizeDim2 = 3;
IndexType sizeDim3 = 5;
IndexType sizeDim4 = 7;
array<IndexType, 4> tensorRange = {{sizeDim1, sizeDim2, sizeDim3, sizeDim4}};
Tensor<DataType, 4, DataLayout, IndexType> tensor(tensorRange);
tensor.setRandom();
array<std::pair<IndexType, IndexType>, 4> paddings;
paddings[0] = std::make_pair(0, 0);
paddings[1] = std::make_pair(2, 1);
paddings[2] = std::make_pair(3, 4);
paddings[3] = std::make_pair(0, 0);
Eigen::DSizes<IndexType, 2> reshape_dims;
reshape_dims[0] = 12;
reshape_dims[1] = 84;
Tensor<DataType, 2, DataLayout, IndexType> result(reshape_dims);
DataType* gpu_data1 = static_cast<DataType*>(sycl_device.allocate(tensor.size()*sizeof(DataType)));
DataType* gpu_data2 = static_cast<DataType*>(sycl_device.allocate(result.size()*sizeof(DataType)));
TensorMap<Tensor<DataType, 4,DataLayout,IndexType>> gpu1(gpu_data1, tensorRange);
TensorMap<Tensor<DataType, 2,DataLayout,IndexType>> gpu2(gpu_data2, reshape_dims);
sycl_device.memcpyHostToDevice(gpu_data1, tensor.data(),(tensor.size())*sizeof(DataType));
gpu2.device(sycl_device)=gpu1.pad(paddings).reshape(reshape_dims);
sycl_device.memcpyDeviceToHost(result.data(), gpu_data2,(result.size())*sizeof(DataType));
for (int i = 0; i < 2; ++i) {
for (int j = 0; j < 6; ++j) {
for (int k = 0; k < 12; ++k) {
for (int l = 0; l < 7; ++l) {
const float result_value = DataLayout == ColMajor ?
result(i+2*j,k+12*l) : result(j+6*i,l+7*k);
if (j >= 2 && j < 5 && k >= 3 && k < 8) {
VERIFY_IS_EQUAL(result_value, tensor(i,j-2,k-3,l));
} else {
VERIFY_IS_EQUAL(result_value, 0.0f);
}
}
}
}
}
sycl_device.deallocate(gpu_data1);
sycl_device.deallocate(gpu_data2);
}
template<typename DataType, typename dev_Selector> void sycl_padding_test_per_device(dev_Selector s){
QueueInterface queueInterface(s);
auto sycl_device = Eigen::SyclDevice(&queueInterface);
test_simple_padding<DataType, RowMajor, int>(sycl_device);
test_simple_padding<DataType, ColMajor, int>(sycl_device);
test_padded_expr<DataType, RowMajor, int>(sycl_device);
test_padded_expr<DataType, ColMajor, int>(sycl_device);
test_simple_padding<DataType, RowMajor, int64_t>(sycl_device);
test_simple_padding<DataType, ColMajor, int64_t>(sycl_device);
test_padded_expr<DataType, RowMajor, int64_t>(sycl_device);
test_padded_expr<DataType, ColMajor, int64_t>(sycl_device);
}
void test_cxx11_tensor_padding_sycl()
{
for (const auto& device :Eigen::get_sycl_supported_devices()) {
CALL_SUBTEST(sycl_padding_test_per_device<float>(device));
}
}