1 // This file is part of Eigen, a lightweight C++ template library
2 // for linear algebra.
3 //
4 // Copyright (C) 2016
5 // Mehdi Goli Codeplay Software Ltd.
6 // Ralph Potter Codeplay Software Ltd.
7 // Luke Iwanski Codeplay Software Ltd.
8 // Contact: <eigen@codeplay.com>
9 // Benoit Steiner <benoit.steiner.goog@gmail.com>
10 //
11 // This Source Code Form is subject to the terms of the Mozilla
12 // Public License v. 2.0. If a copy of the MPL was not distributed
13 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
14
15 #define EIGEN_TEST_NO_LONGDOUBLE
16 #define EIGEN_TEST_NO_COMPLEX
17
18 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int64_t
19 #define EIGEN_USE_SYCL
20
21 #include "main.h"
22 #include <unsupported/Eigen/CXX11/Tensor>
23
24 using Eigen::array;
25 using Eigen::SyclDevice;
26 using Eigen::Tensor;
27 using Eigen::TensorMap;
28
29 template <typename DataType, int DataLayout, typename IndexType>
test_simple_shuffling_sycl(const Eigen::SyclDevice & sycl_device)30 static void test_simple_shuffling_sycl(const Eigen::SyclDevice& sycl_device) {
31 IndexType sizeDim1 = 2;
32 IndexType sizeDim2 = 3;
33 IndexType sizeDim3 = 5;
34 IndexType sizeDim4 = 7;
35 array<IndexType, 4> tensorRange = {{sizeDim1, sizeDim2, sizeDim3, sizeDim4}};
36 Tensor<DataType, 4, DataLayout, IndexType> tensor(tensorRange);
37 Tensor<DataType, 4, DataLayout, IndexType> no_shuffle(tensorRange);
38 tensor.setRandom();
39
40 const size_t buffSize = tensor.size() * sizeof(DataType);
41 array<IndexType, 4> shuffles;
42 shuffles[0] = 0;
43 shuffles[1] = 1;
44 shuffles[2] = 2;
45 shuffles[3] = 3;
46 DataType* gpu_data1 = static_cast<DataType*>(sycl_device.allocate(buffSize));
47 DataType* gpu_data2 = static_cast<DataType*>(sycl_device.allocate(buffSize));
48
49 TensorMap<Tensor<DataType, 4, DataLayout, IndexType>> gpu1(gpu_data1,
50 tensorRange);
51 TensorMap<Tensor<DataType, 4, DataLayout, IndexType>> gpu2(gpu_data2,
52 tensorRange);
53
54 sycl_device.memcpyHostToDevice(gpu_data1, tensor.data(), buffSize);
55
56 gpu2.device(sycl_device) = gpu1.shuffle(shuffles);
57 sycl_device.memcpyDeviceToHost(no_shuffle.data(), gpu_data2, buffSize);
58 sycl_device.synchronize();
59
60 VERIFY_IS_EQUAL(no_shuffle.dimension(0), sizeDim1);
61 VERIFY_IS_EQUAL(no_shuffle.dimension(1), sizeDim2);
62 VERIFY_IS_EQUAL(no_shuffle.dimension(2), sizeDim3);
63 VERIFY_IS_EQUAL(no_shuffle.dimension(3), sizeDim4);
64
65 for (IndexType i = 0; i < sizeDim1; ++i) {
66 for (IndexType j = 0; j < sizeDim2; ++j) {
67 for (IndexType k = 0; k < sizeDim3; ++k) {
68 for (IndexType l = 0; l < sizeDim4; ++l) {
69 VERIFY_IS_EQUAL(tensor(i, j, k, l), no_shuffle(i, j, k, l));
70 }
71 }
72 }
73 }
74
75 shuffles[0] = 2;
76 shuffles[1] = 3;
77 shuffles[2] = 1;
78 shuffles[3] = 0;
79 array<IndexType, 4> tensorrangeShuffle = {
80 {sizeDim3, sizeDim4, sizeDim2, sizeDim1}};
81 Tensor<DataType, 4, DataLayout, IndexType> shuffle(tensorrangeShuffle);
82 DataType* gpu_data3 = static_cast<DataType*>(sycl_device.allocate(buffSize));
83 TensorMap<Tensor<DataType, 4, DataLayout, IndexType>> gpu3(
84 gpu_data3, tensorrangeShuffle);
85
86 gpu3.device(sycl_device) = gpu1.shuffle(shuffles);
87 sycl_device.memcpyDeviceToHost(shuffle.data(), gpu_data3, buffSize);
88 sycl_device.synchronize();
89
90 VERIFY_IS_EQUAL(shuffle.dimension(0), sizeDim3);
91 VERIFY_IS_EQUAL(shuffle.dimension(1), sizeDim4);
92 VERIFY_IS_EQUAL(shuffle.dimension(2), sizeDim2);
93 VERIFY_IS_EQUAL(shuffle.dimension(3), sizeDim1);
94
95 for (IndexType i = 0; i < sizeDim1; ++i) {
96 for (IndexType j = 0; j < sizeDim2; ++j) {
97 for (IndexType k = 0; k < sizeDim3; ++k) {
98 for (IndexType l = 0; l < sizeDim4; ++l) {
99 VERIFY_IS_EQUAL(tensor(i, j, k, l), shuffle(k, l, j, i));
100 }
101 }
102 }
103 }
104 }
105
106 template <typename DataType, typename dev_Selector>
sycl_shuffling_test_per_device(dev_Selector s)107 void sycl_shuffling_test_per_device(dev_Selector s) {
108 QueueInterface queueInterface(s);
109 auto sycl_device = Eigen::SyclDevice(&queueInterface);
110 test_simple_shuffling_sycl<DataType, RowMajor, int64_t>(sycl_device);
111 test_simple_shuffling_sycl<DataType, ColMajor, int64_t>(sycl_device);
112 }
EIGEN_DECLARE_TEST(cxx11_tensor_shuffling_sycl)113 EIGEN_DECLARE_TEST(cxx11_tensor_shuffling_sycl) {
114 for (const auto& device : Eigen::get_sycl_supported_devices()) {
115 CALL_SUBTEST(sycl_shuffling_test_per_device<float>(device));
116 }
117 }
118