cxx11_tensor_broadcast_sycl.cpp (5708B)
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::array; 24 using Eigen::SyclDevice; 25 using Eigen::Tensor; 26 using Eigen::TensorMap; 27 28 template <typename DataType, int DataLayout, typename IndexType> 29 static void test_broadcast_sycl_fixed(const Eigen::SyclDevice &sycl_device){ 30 31 // BROADCAST test: 32 IndexType inDim1=2; 33 IndexType inDim2=3; 34 IndexType inDim3=5; 35 IndexType inDim4=7; 36 IndexType bDim1=2; 37 IndexType bDim2=3; 38 IndexType bDim3=1; 39 IndexType bDim4=4; 40 array<IndexType, 4> in_range = {{inDim1, inDim2, inDim3, inDim4}}; 41 array<IndexType, 4> broadcasts = {{bDim1, bDim2, bDim3, bDim4}}; 42 array<IndexType, 4> out_range; // = in_range * broadcasts 43 for (size_t i = 0; i < out_range.size(); ++i) 44 out_range[i] = in_range[i] * broadcasts[i]; 45 46 Tensor<DataType, 4, DataLayout, IndexType> input(in_range); 47 Tensor<DataType, 4, DataLayout, IndexType> out(out_range); 48 49 for (size_t i = 0; i < in_range.size(); ++i) 50 VERIFY_IS_EQUAL(out.dimension(i), out_range[i]); 51 52 53 for (IndexType i = 0; i < input.size(); ++i) 54 input(i) = static_cast<DataType>(i); 55 56 DataType * gpu_in_data = static_cast<DataType*>(sycl_device.allocate(input.dimensions().TotalSize()*sizeof(DataType))); 57 DataType * gpu_out_data = static_cast<DataType*>(sycl_device.allocate(out.dimensions().TotalSize()*sizeof(DataType))); 58 59 TensorMap<TensorFixedSize<DataType, Sizes<2, 3, 5, 7>, DataLayout, IndexType>> gpu_in(gpu_in_data, in_range); 60 TensorMap<Tensor<DataType, 4, DataLayout, IndexType>> gpu_out(gpu_out_data, out_range); 61 sycl_device.memcpyHostToDevice(gpu_in_data, input.data(),(input.dimensions().TotalSize())*sizeof(DataType)); 62 gpu_out.device(sycl_device) = gpu_in.broadcast(broadcasts); 63 sycl_device.memcpyDeviceToHost(out.data(), gpu_out_data,(out.dimensions().TotalSize())*sizeof(DataType)); 64 65 for (IndexType i = 0; i < inDim1*bDim1; ++i) { 66 for (IndexType j = 0; j < inDim2*bDim2; ++j) { 67 for (IndexType k = 0; k < inDim3*bDim3; ++k) { 68 for (IndexType l = 0; l < inDim4*bDim4; ++l) { 69 VERIFY_IS_APPROX(input(i%2,j%3,k%5,l%7), out(i,j,k,l)); 70 } 71 } 72 } 73 } 74 printf("Broadcast Test with fixed size Passed\n"); 75 sycl_device.deallocate(gpu_in_data); 76 sycl_device.deallocate(gpu_out_data); 77 } 78 79 template <typename DataType, int DataLayout, typename IndexType> 80 static void test_broadcast_sycl(const Eigen::SyclDevice &sycl_device){ 81 82 // BROADCAST test: 83 IndexType inDim1=2; 84 IndexType inDim2=3; 85 IndexType inDim3=5; 86 IndexType inDim4=7; 87 IndexType bDim1=2; 88 IndexType bDim2=3; 89 IndexType bDim3=1; 90 IndexType bDim4=4; 91 array<IndexType, 4> in_range = {{inDim1, inDim2, inDim3, inDim4}}; 92 array<IndexType, 4> broadcasts = {{bDim1, bDim2, bDim3, bDim4}}; 93 array<IndexType, 4> out_range; // = in_range * broadcasts 94 for (size_t i = 0; i < out_range.size(); ++i) 95 out_range[i] = in_range[i] * broadcasts[i]; 96 97 Tensor<DataType, 4, DataLayout, IndexType> input(in_range); 98 Tensor<DataType, 4, DataLayout, IndexType> out(out_range); 99 100 for (size_t i = 0; i < in_range.size(); ++i) 101 VERIFY_IS_EQUAL(out.dimension(i), out_range[i]); 102 103 104 for (IndexType i = 0; i < input.size(); ++i) 105 input(i) = static_cast<DataType>(i); 106 107 DataType * gpu_in_data = static_cast<DataType*>(sycl_device.allocate(input.dimensions().TotalSize()*sizeof(DataType))); 108 DataType * gpu_out_data = static_cast<DataType*>(sycl_device.allocate(out.dimensions().TotalSize()*sizeof(DataType))); 109 110 TensorMap<Tensor<DataType, 4, DataLayout, IndexType>> gpu_in(gpu_in_data, in_range); 111 TensorMap<Tensor<DataType, 4, DataLayout, IndexType>> gpu_out(gpu_out_data, out_range); 112 sycl_device.memcpyHostToDevice(gpu_in_data, input.data(),(input.dimensions().TotalSize())*sizeof(DataType)); 113 gpu_out.device(sycl_device) = gpu_in.broadcast(broadcasts); 114 sycl_device.memcpyDeviceToHost(out.data(), gpu_out_data,(out.dimensions().TotalSize())*sizeof(DataType)); 115 116 for (IndexType i = 0; i < inDim1*bDim1; ++i) { 117 for (IndexType j = 0; j < inDim2*bDim2; ++j) { 118 for (IndexType k = 0; k < inDim3*bDim3; ++k) { 119 for (IndexType l = 0; l < inDim4*bDim4; ++l) { 120 VERIFY_IS_APPROX(input(i%inDim1,j%inDim2,k%inDim3,l%inDim4), out(i,j,k,l)); 121 } 122 } 123 } 124 } 125 printf("Broadcast Test Passed\n"); 126 sycl_device.deallocate(gpu_in_data); 127 sycl_device.deallocate(gpu_out_data); 128 } 129 130 template<typename DataType> void sycl_broadcast_test_per_device(const cl::sycl::device& d){ 131 std::cout << "Running on " << d.template get_info<cl::sycl::info::device::name>() << std::endl; 132 QueueInterface queueInterface(d); 133 auto sycl_device = Eigen::SyclDevice(&queueInterface); 134 test_broadcast_sycl<DataType, RowMajor, int64_t>(sycl_device); 135 test_broadcast_sycl<DataType, ColMajor, int64_t>(sycl_device); 136 test_broadcast_sycl_fixed<DataType, RowMajor, int64_t>(sycl_device); 137 test_broadcast_sycl_fixed<DataType, ColMajor, int64_t>(sycl_device); 138 } 139 140 EIGEN_DECLARE_TEST(cxx11_tensor_broadcast_sycl) { 141 for (const auto& device :Eigen::get_sycl_supported_devices()) { 142 CALL_SUBTEST(sycl_broadcast_test_per_device<float>(device)); 143 } 144 }