cart-elc

Source code for CART-ELC
git clone git://git.laack.co/cart-elc.git
Log | Files | Refs | README | LICENSE

cxx11_tensor_generator_sycl.cpp (5732B)


      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 static const float error_threshold =1e-8f;
     20 
     21 #include "main.h"
     22 #include <unsupported/Eigen/CXX11/Tensor>
     23 
     24 using Eigen::Tensor;
     25 struct Generator1D {
     26   Generator1D() { }
     27 
     28   float operator()(const array<Eigen::DenseIndex, 1>& coordinates) const {
     29     return coordinates[0];
     30   }
     31 };
     32 
     33 template <typename DataType, int DataLayout, typename IndexType>
     34 static void test_1D_sycl(const Eigen::SyclDevice& sycl_device)
     35 {
     36 
     37   IndexType sizeDim1 = 6;
     38   array<IndexType, 1> tensorRange = {{sizeDim1}};
     39   Tensor<DataType, 1, DataLayout,IndexType> vec(tensorRange);
     40   Tensor<DataType, 1, DataLayout,IndexType> result(tensorRange);
     41 
     42   const size_t tensorBuffSize =vec.size()*sizeof(DataType);
     43   DataType* gpu_data_vec  = static_cast<DataType*>(sycl_device.allocate(tensorBuffSize));
     44   DataType* gpu_data_result  = static_cast<DataType*>(sycl_device.allocate(tensorBuffSize));
     45 
     46   TensorMap<Tensor<DataType, 1, DataLayout,IndexType>> gpu_vec(gpu_data_vec, tensorRange);
     47   TensorMap<Tensor<DataType, 1, DataLayout,IndexType>> gpu_result(gpu_data_result, tensorRange);
     48 
     49   sycl_device.memcpyHostToDevice(gpu_data_vec, vec.data(), tensorBuffSize);
     50   gpu_result.device(sycl_device)=gpu_vec.generate(Generator1D());
     51   sycl_device.memcpyDeviceToHost(result.data(), gpu_data_result, tensorBuffSize);
     52 
     53   for (IndexType i = 0; i < 6; ++i) {
     54     VERIFY_IS_EQUAL(result(i), i);
     55   }
     56 }
     57 
     58 
     59 struct Generator2D {
     60   Generator2D() { }
     61 
     62   float operator()(const array<Eigen::DenseIndex, 2>& coordinates) const {
     63     return 3 * coordinates[0] + 11 * coordinates[1];
     64   }
     65 };
     66 
     67 template <typename DataType, int DataLayout, typename IndexType>
     68 static void test_2D_sycl(const Eigen::SyclDevice& sycl_device)
     69 {
     70   IndexType sizeDim1 = 5;
     71   IndexType sizeDim2 = 7;
     72   array<IndexType, 2> tensorRange = {{sizeDim1, sizeDim2}};
     73   Tensor<DataType, 2, DataLayout,IndexType> matrix(tensorRange);
     74   Tensor<DataType, 2, DataLayout,IndexType> result(tensorRange);
     75 
     76   const size_t tensorBuffSize =matrix.size()*sizeof(DataType);
     77   DataType* gpu_data_matrix  = static_cast<DataType*>(sycl_device.allocate(tensorBuffSize));
     78   DataType* gpu_data_result  = static_cast<DataType*>(sycl_device.allocate(tensorBuffSize));
     79 
     80   TensorMap<Tensor<DataType, 2, DataLayout,IndexType>> gpu_matrix(gpu_data_matrix, tensorRange);
     81   TensorMap<Tensor<DataType, 2, DataLayout,IndexType>> gpu_result(gpu_data_result, tensorRange);
     82 
     83   sycl_device.memcpyHostToDevice(gpu_data_matrix, matrix.data(), tensorBuffSize);
     84   gpu_result.device(sycl_device)=gpu_matrix.generate(Generator2D());
     85   sycl_device.memcpyDeviceToHost(result.data(), gpu_data_result, tensorBuffSize);
     86 
     87   for (IndexType i = 0; i < 5; ++i) {
     88     for (IndexType j = 0; j < 5; ++j) {
     89       VERIFY_IS_EQUAL(result(i, j), 3*i + 11*j);
     90     }
     91   }
     92 }
     93 
     94 template <typename DataType, int DataLayout, typename IndexType>
     95 static void test_gaussian_sycl(const Eigen::SyclDevice& sycl_device)
     96 {
     97   IndexType rows = 32;
     98   IndexType cols = 48;
     99   array<DataType, 2> means;
    100   means[0] = rows / 2.0f;
    101   means[1] = cols / 2.0f;
    102   array<DataType, 2> std_devs;
    103   std_devs[0] = 3.14f;
    104   std_devs[1] = 2.7f;
    105   internal::GaussianGenerator<DataType, Eigen::DenseIndex, 2> gaussian_gen(means, std_devs);
    106 
    107   array<IndexType, 2> tensorRange = {{rows, cols}};
    108   Tensor<DataType, 2, DataLayout,IndexType> matrix(tensorRange);
    109   Tensor<DataType, 2, DataLayout,IndexType> result(tensorRange);
    110 
    111   const size_t tensorBuffSize =matrix.size()*sizeof(DataType);
    112   DataType* gpu_data_matrix  = static_cast<DataType*>(sycl_device.allocate(tensorBuffSize));
    113   DataType* gpu_data_result  = static_cast<DataType*>(sycl_device.allocate(tensorBuffSize));
    114 
    115   TensorMap<Tensor<DataType, 2, DataLayout,IndexType>> gpu_matrix(gpu_data_matrix, tensorRange);
    116   TensorMap<Tensor<DataType, 2, DataLayout,IndexType>> gpu_result(gpu_data_result, tensorRange);
    117 
    118   sycl_device.memcpyHostToDevice(gpu_data_matrix, matrix.data(), tensorBuffSize);
    119   gpu_result.device(sycl_device)=gpu_matrix.generate(gaussian_gen);
    120   sycl_device.memcpyDeviceToHost(result.data(), gpu_data_result, tensorBuffSize);
    121 
    122   for (IndexType i = 0; i < rows; ++i) {
    123     for (IndexType j = 0; j < cols; ++j) {
    124       DataType g_rows = powf(rows/2.0f - i, 2) / (3.14f * 3.14f) * 0.5f;
    125       DataType g_cols = powf(cols/2.0f - j, 2) / (2.7f * 2.7f) * 0.5f;
    126       DataType gaussian = expf(-g_rows - g_cols);
    127       Eigen::internal::isApprox(result(i, j), gaussian, error_threshold);
    128     }
    129   }
    130 }
    131 
    132 template<typename DataType, typename dev_Selector> void sycl_generator_test_per_device(dev_Selector s){
    133   QueueInterface queueInterface(s);
    134   auto sycl_device = Eigen::SyclDevice(&queueInterface);
    135   test_1D_sycl<DataType, RowMajor, int64_t>(sycl_device);
    136   test_1D_sycl<DataType, ColMajor, int64_t>(sycl_device);
    137   test_2D_sycl<DataType, RowMajor, int64_t>(sycl_device);
    138   test_2D_sycl<DataType, ColMajor, int64_t>(sycl_device);
    139   test_gaussian_sycl<DataType, RowMajor, int64_t>(sycl_device);
    140   test_gaussian_sycl<DataType, ColMajor, int64_t>(sycl_device);
    141 }
    142 EIGEN_DECLARE_TEST(cxx11_tensor_generator_sycl)
    143 {
    144   for (const auto& device :Eigen::get_sycl_supported_devices()) {
    145     CALL_SUBTEST(sycl_generator_test_per_device<float>(device));
    146   }
    147 }