cxx11_tensor_concatenation_sycl.cpp (8411B)
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 // 10 // This Source Code Form is subject to the terms of the Mozilla 11 // Public License v. 2.0. If a copy of the MPL was not distributed 12 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/. 13 14 #define EIGEN_TEST_NO_LONGDOUBLE 15 #define EIGEN_TEST_NO_COMPLEX 16 17 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int64_t 18 #define EIGEN_USE_SYCL 19 20 #include "main.h" 21 #include <unsupported/Eigen/CXX11/Tensor> 22 23 using Eigen::Tensor; 24 25 template<typename DataType, int DataLayout, typename IndexType> 26 static void test_simple_concatenation(const Eigen::SyclDevice& sycl_device) 27 { 28 IndexType leftDim1 = 2; 29 IndexType leftDim2 = 3; 30 IndexType leftDim3 = 1; 31 Eigen::array<IndexType, 3> leftRange = {{leftDim1, leftDim2, leftDim3}}; 32 IndexType rightDim1 = 2; 33 IndexType rightDim2 = 3; 34 IndexType rightDim3 = 1; 35 Eigen::array<IndexType, 3> rightRange = {{rightDim1, rightDim2, rightDim3}}; 36 37 //IndexType concatDim1 = 3; 38 // IndexType concatDim2 = 3; 39 // IndexType concatDim3 = 1; 40 //Eigen::array<IndexType, 3> concatRange = {{concatDim1, concatDim2, concatDim3}}; 41 42 Tensor<DataType, 3, DataLayout, IndexType> left(leftRange); 43 Tensor<DataType, 3, DataLayout, IndexType> right(rightRange); 44 left.setRandom(); 45 right.setRandom(); 46 47 DataType * gpu_in1_data = static_cast<DataType*>(sycl_device.allocate(left.dimensions().TotalSize()*sizeof(DataType))); 48 DataType * gpu_in2_data = static_cast<DataType*>(sycl_device.allocate(right.dimensions().TotalSize()*sizeof(DataType))); 49 50 Eigen::TensorMap<Eigen::Tensor<DataType, 3, DataLayout, IndexType>> gpu_in1(gpu_in1_data, leftRange); 51 Eigen::TensorMap<Eigen::Tensor<DataType, 3, DataLayout, IndexType>> gpu_in2(gpu_in2_data, rightRange); 52 sycl_device.memcpyHostToDevice(gpu_in1_data, left.data(),(left.dimensions().TotalSize())*sizeof(DataType)); 53 sycl_device.memcpyHostToDevice(gpu_in2_data, right.data(),(right.dimensions().TotalSize())*sizeof(DataType)); 54 /// 55 Tensor<DataType, 3, DataLayout, IndexType> concatenation1(leftDim1+rightDim1, leftDim2, leftDim3); 56 DataType * gpu_out_data1 = static_cast<DataType*>(sycl_device.allocate(concatenation1.dimensions().TotalSize()*sizeof(DataType))); 57 Eigen::TensorMap<Eigen::Tensor<DataType, 3, DataLayout, IndexType>> gpu_out1(gpu_out_data1, concatenation1.dimensions()); 58 59 //concatenation = left.concatenate(right, 0); 60 gpu_out1.device(sycl_device) =gpu_in1.concatenate(gpu_in2, 0); 61 sycl_device.memcpyDeviceToHost(concatenation1.data(), gpu_out_data1,(concatenation1.dimensions().TotalSize())*sizeof(DataType)); 62 63 VERIFY_IS_EQUAL(concatenation1.dimension(0), 4); 64 VERIFY_IS_EQUAL(concatenation1.dimension(1), 3); 65 VERIFY_IS_EQUAL(concatenation1.dimension(2), 1); 66 for (IndexType j = 0; j < 3; ++j) { 67 for (IndexType i = 0; i < 2; ++i) { 68 VERIFY_IS_EQUAL(concatenation1(i, j, 0), left(i, j, 0)); 69 } 70 for (IndexType i = 2; i < 4; ++i) { 71 VERIFY_IS_EQUAL(concatenation1(i, j, 0), right(i - 2, j, 0)); 72 } 73 } 74 75 sycl_device.deallocate(gpu_out_data1); 76 Tensor<DataType, 3, DataLayout, IndexType> concatenation2(leftDim1, leftDim2 +rightDim2, leftDim3); 77 DataType * gpu_out_data2 = static_cast<DataType*>(sycl_device.allocate(concatenation2.dimensions().TotalSize()*sizeof(DataType))); 78 Eigen::TensorMap<Eigen::Tensor<DataType, 3, DataLayout, IndexType>> gpu_out2(gpu_out_data2, concatenation2.dimensions()); 79 gpu_out2.device(sycl_device) =gpu_in1.concatenate(gpu_in2, 1); 80 sycl_device.memcpyDeviceToHost(concatenation2.data(), gpu_out_data2,(concatenation2.dimensions().TotalSize())*sizeof(DataType)); 81 82 //concatenation = left.concatenate(right, 1); 83 VERIFY_IS_EQUAL(concatenation2.dimension(0), 2); 84 VERIFY_IS_EQUAL(concatenation2.dimension(1), 6); 85 VERIFY_IS_EQUAL(concatenation2.dimension(2), 1); 86 for (IndexType i = 0; i < 2; ++i) { 87 for (IndexType j = 0; j < 3; ++j) { 88 VERIFY_IS_EQUAL(concatenation2(i, j, 0), left(i, j, 0)); 89 } 90 for (IndexType j = 3; j < 6; ++j) { 91 VERIFY_IS_EQUAL(concatenation2(i, j, 0), right(i, j - 3, 0)); 92 } 93 } 94 sycl_device.deallocate(gpu_out_data2); 95 Tensor<DataType, 3, DataLayout, IndexType> concatenation3(leftDim1, leftDim2, leftDim3+rightDim3); 96 DataType * gpu_out_data3 = static_cast<DataType*>(sycl_device.allocate(concatenation3.dimensions().TotalSize()*sizeof(DataType))); 97 Eigen::TensorMap<Eigen::Tensor<DataType, 3, DataLayout, IndexType>> gpu_out3(gpu_out_data3, concatenation3.dimensions()); 98 gpu_out3.device(sycl_device) =gpu_in1.concatenate(gpu_in2, 2); 99 sycl_device.memcpyDeviceToHost(concatenation3.data(), gpu_out_data3,(concatenation3.dimensions().TotalSize())*sizeof(DataType)); 100 101 //concatenation = left.concatenate(right, 2); 102 VERIFY_IS_EQUAL(concatenation3.dimension(0), 2); 103 VERIFY_IS_EQUAL(concatenation3.dimension(1), 3); 104 VERIFY_IS_EQUAL(concatenation3.dimension(2), 2); 105 for (IndexType i = 0; i < 2; ++i) { 106 for (IndexType j = 0; j < 3; ++j) { 107 VERIFY_IS_EQUAL(concatenation3(i, j, 0), left(i, j, 0)); 108 VERIFY_IS_EQUAL(concatenation3(i, j, 1), right(i, j, 0)); 109 } 110 } 111 sycl_device.deallocate(gpu_out_data3); 112 sycl_device.deallocate(gpu_in1_data); 113 sycl_device.deallocate(gpu_in2_data); 114 } 115 template<typename DataType, int DataLayout, typename IndexType> 116 static void test_concatenation_as_lvalue(const Eigen::SyclDevice& sycl_device) 117 { 118 119 IndexType leftDim1 = 2; 120 IndexType leftDim2 = 3; 121 Eigen::array<IndexType, 2> leftRange = {{leftDim1, leftDim2}}; 122 123 IndexType rightDim1 = 2; 124 IndexType rightDim2 = 3; 125 Eigen::array<IndexType, 2> rightRange = {{rightDim1, rightDim2}}; 126 127 IndexType concatDim1 = 4; 128 IndexType concatDim2 = 3; 129 Eigen::array<IndexType, 2> resRange = {{concatDim1, concatDim2}}; 130 131 Tensor<DataType, 2, DataLayout, IndexType> left(leftRange); 132 Tensor<DataType, 2, DataLayout, IndexType> right(rightRange); 133 Tensor<DataType, 2, DataLayout, IndexType> result(resRange); 134 135 left.setRandom(); 136 right.setRandom(); 137 result.setRandom(); 138 139 DataType * gpu_in1_data = static_cast<DataType*>(sycl_device.allocate(left.dimensions().TotalSize()*sizeof(DataType))); 140 DataType * gpu_in2_data = static_cast<DataType*>(sycl_device.allocate(right.dimensions().TotalSize()*sizeof(DataType))); 141 DataType * gpu_out_data = static_cast<DataType*>(sycl_device.allocate(result.dimensions().TotalSize()*sizeof(DataType))); 142 143 144 Eigen::TensorMap<Eigen::Tensor<DataType, 2, DataLayout, IndexType>> gpu_in1(gpu_in1_data, leftRange); 145 Eigen::TensorMap<Eigen::Tensor<DataType, 2, DataLayout, IndexType>> gpu_in2(gpu_in2_data, rightRange); 146 Eigen::TensorMap<Eigen::Tensor<DataType, 2, DataLayout, IndexType>> gpu_out(gpu_out_data, resRange); 147 148 sycl_device.memcpyHostToDevice(gpu_in1_data, left.data(),(left.dimensions().TotalSize())*sizeof(DataType)); 149 sycl_device.memcpyHostToDevice(gpu_in2_data, right.data(),(right.dimensions().TotalSize())*sizeof(DataType)); 150 sycl_device.memcpyHostToDevice(gpu_out_data, result.data(),(result.dimensions().TotalSize())*sizeof(DataType)); 151 152 // t1.concatenate(t2, 0) = result; 153 gpu_in1.concatenate(gpu_in2, 0).device(sycl_device) =gpu_out; 154 sycl_device.memcpyDeviceToHost(left.data(), gpu_in1_data,(left.dimensions().TotalSize())*sizeof(DataType)); 155 sycl_device.memcpyDeviceToHost(right.data(), gpu_in2_data,(right.dimensions().TotalSize())*sizeof(DataType)); 156 157 for (IndexType i = 0; i < 2; ++i) { 158 for (IndexType j = 0; j < 3; ++j) { 159 VERIFY_IS_EQUAL(left(i, j), result(i, j)); 160 VERIFY_IS_EQUAL(right(i, j), result(i+2, j)); 161 } 162 } 163 sycl_device.deallocate(gpu_in1_data); 164 sycl_device.deallocate(gpu_in2_data); 165 sycl_device.deallocate(gpu_out_data); 166 } 167 168 169 template <typename DataType, typename Dev_selector> void tensorConcat_perDevice(Dev_selector s){ 170 QueueInterface queueInterface(s); 171 auto sycl_device = Eigen::SyclDevice(&queueInterface); 172 test_simple_concatenation<DataType, RowMajor, int64_t>(sycl_device); 173 test_simple_concatenation<DataType, ColMajor, int64_t>(sycl_device); 174 test_concatenation_as_lvalue<DataType, ColMajor, int64_t>(sycl_device); 175 } 176 EIGEN_DECLARE_TEST(cxx11_tensor_concatenation_sycl) { 177 for (const auto& device :Eigen::get_sycl_supported_devices()) { 178 CALL_SUBTEST(tensorConcat_perDevice<float>(device)); 179 } 180 }