cart-elc

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

cxx11_tensor_padding_sycl.cpp (5677B)


      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 
     16 #define EIGEN_TEST_NO_LONGDOUBLE
     17 #define EIGEN_TEST_NO_COMPLEX
     18 
     19 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int64_t
     20 #define EIGEN_USE_SYCL
     21 
     22 
     23 #include "main.h"
     24 #include <unsupported/Eigen/CXX11/Tensor>
     25 
     26 using Eigen::array;
     27 using Eigen::SyclDevice;
     28 using Eigen::Tensor;
     29 using Eigen::TensorMap;
     30 
     31 
     32 template<typename DataType, int DataLayout, typename IndexType>
     33 static void test_simple_padding(const Eigen::SyclDevice& sycl_device)
     34 {
     35 
     36   IndexType sizeDim1 = 2;
     37   IndexType sizeDim2 = 3;
     38   IndexType sizeDim3 = 5;
     39   IndexType sizeDim4 = 7;
     40   array<IndexType, 4> tensorRange = {{sizeDim1, sizeDim2, sizeDim3, sizeDim4}};
     41 
     42   Tensor<DataType, 4, DataLayout, IndexType> tensor(tensorRange);
     43   tensor.setRandom();
     44 
     45   array<std::pair<IndexType, IndexType>, 4> paddings;
     46   paddings[0] = std::make_pair(0, 0);
     47   paddings[1] = std::make_pair(2, 1);
     48   paddings[2] = std::make_pair(3, 4);
     49   paddings[3] = std::make_pair(0, 0);
     50 
     51   IndexType padedSizeDim1 = 2;
     52   IndexType padedSizeDim2 = 6;
     53   IndexType padedSizeDim3 = 12;
     54   IndexType padedSizeDim4 = 7;
     55   array<IndexType, 4> padedtensorRange = {{padedSizeDim1, padedSizeDim2, padedSizeDim3, padedSizeDim4}};
     56 
     57   Tensor<DataType, 4, DataLayout, IndexType> padded(padedtensorRange);
     58 
     59 
     60   DataType* gpu_data1  = static_cast<DataType*>(sycl_device.allocate(tensor.size()*sizeof(DataType)));
     61   DataType* gpu_data2  = static_cast<DataType*>(sycl_device.allocate(padded.size()*sizeof(DataType)));
     62   TensorMap<Tensor<DataType, 4,DataLayout,IndexType>> gpu1(gpu_data1, tensorRange);
     63   TensorMap<Tensor<DataType, 4,DataLayout,IndexType>> gpu2(gpu_data2, padedtensorRange);
     64 
     65   VERIFY_IS_EQUAL(padded.dimension(0), 2+0);
     66   VERIFY_IS_EQUAL(padded.dimension(1), 3+3);
     67   VERIFY_IS_EQUAL(padded.dimension(2), 5+7);
     68   VERIFY_IS_EQUAL(padded.dimension(3), 7+0);
     69   sycl_device.memcpyHostToDevice(gpu_data1, tensor.data(),(tensor.size())*sizeof(DataType));
     70   gpu2.device(sycl_device)=gpu1.pad(paddings);
     71   sycl_device.memcpyDeviceToHost(padded.data(), gpu_data2,(padded.size())*sizeof(DataType));
     72   for (IndexType i = 0; i < padedSizeDim1; ++i) {
     73     for (IndexType j = 0; j < padedSizeDim2; ++j) {
     74       for (IndexType k = 0; k < padedSizeDim3; ++k) {
     75         for (IndexType l = 0; l < padedSizeDim4; ++l) {
     76           if (j >= 2 && j < 5 && k >= 3 && k < 8) {
     77             VERIFY_IS_EQUAL(padded(i,j,k,l), tensor(i,j-2,k-3,l));
     78           } else {
     79             VERIFY_IS_EQUAL(padded(i,j,k,l), 0.0f);
     80           }
     81         }
     82       }
     83     }
     84   }
     85   sycl_device.deallocate(gpu_data1);
     86   sycl_device.deallocate(gpu_data2);
     87 }
     88 
     89 template<typename DataType, int DataLayout, typename IndexType>
     90 static void test_padded_expr(const Eigen::SyclDevice& sycl_device)
     91 {
     92   IndexType sizeDim1 = 2;
     93   IndexType sizeDim2 = 3;
     94   IndexType sizeDim3 = 5;
     95   IndexType sizeDim4 = 7;
     96   array<IndexType, 4> tensorRange = {{sizeDim1, sizeDim2, sizeDim3, sizeDim4}};
     97 
     98   Tensor<DataType, 4, DataLayout, IndexType> tensor(tensorRange);
     99   tensor.setRandom();
    100 
    101   array<std::pair<IndexType, IndexType>, 4> paddings;
    102   paddings[0] = std::make_pair(0, 0);
    103   paddings[1] = std::make_pair(2, 1);
    104   paddings[2] = std::make_pair(3, 4);
    105   paddings[3] = std::make_pair(0, 0);
    106 
    107   Eigen::DSizes<IndexType, 2> reshape_dims;
    108   reshape_dims[0] = 12;
    109   reshape_dims[1] = 84;
    110 
    111 
    112   Tensor<DataType, 2, DataLayout, IndexType>  result(reshape_dims);
    113 
    114   DataType* gpu_data1  = static_cast<DataType*>(sycl_device.allocate(tensor.size()*sizeof(DataType)));
    115   DataType* gpu_data2  = static_cast<DataType*>(sycl_device.allocate(result.size()*sizeof(DataType)));
    116   TensorMap<Tensor<DataType, 4,DataLayout,IndexType>> gpu1(gpu_data1, tensorRange);
    117   TensorMap<Tensor<DataType, 2,DataLayout,IndexType>> gpu2(gpu_data2, reshape_dims);
    118 
    119 
    120   sycl_device.memcpyHostToDevice(gpu_data1, tensor.data(),(tensor.size())*sizeof(DataType));
    121   gpu2.device(sycl_device)=gpu1.pad(paddings).reshape(reshape_dims);
    122   sycl_device.memcpyDeviceToHost(result.data(), gpu_data2,(result.size())*sizeof(DataType));
    123 
    124   for (IndexType i = 0; i < 2; ++i) {
    125     for (IndexType j = 0; j < 6; ++j) {
    126       for (IndexType k = 0; k < 12; ++k) {
    127         for (IndexType l = 0; l < 7; ++l) {
    128           const float result_value = DataLayout == ColMajor ?
    129               result(i+2*j,k+12*l) : result(j+6*i,l+7*k);
    130           if (j >= 2 && j < 5 && k >= 3 && k < 8) {
    131             VERIFY_IS_EQUAL(result_value, tensor(i,j-2,k-3,l));
    132           } else {
    133             VERIFY_IS_EQUAL(result_value, 0.0f);
    134           }
    135         }
    136       }
    137     }
    138   }
    139   sycl_device.deallocate(gpu_data1);
    140   sycl_device.deallocate(gpu_data2);
    141 }
    142 
    143 template<typename DataType, typename dev_Selector> void sycl_padding_test_per_device(dev_Selector s){
    144   QueueInterface queueInterface(s);
    145   auto sycl_device = Eigen::SyclDevice(&queueInterface);
    146   test_simple_padding<DataType, RowMajor, int64_t>(sycl_device);
    147   test_simple_padding<DataType, ColMajor, int64_t>(sycl_device);
    148   test_padded_expr<DataType, RowMajor, int64_t>(sycl_device);
    149   test_padded_expr<DataType, ColMajor, int64_t>(sycl_device);
    150 
    151 }
    152 EIGEN_DECLARE_TEST(cxx11_tensor_padding_sycl)
    153 {
    154   for (const auto& device :Eigen::get_sycl_supported_devices()) {
    155     CALL_SUBTEST(sycl_padding_test_per_device<float>(device));
    156   }
    157 }