cart-elc

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

cxx11_tensor_sycl.cpp (14828B)


      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 #include "main.h"
     23 #include <unsupported/Eigen/CXX11/Tensor>
     24 
     25 using Eigen::array;
     26 using Eigen::SyclDevice;
     27 using Eigen::Tensor;
     28 using Eigen::TensorMap;
     29 
     30 template <typename DataType, int DataLayout, typename IndexType>
     31 void test_sycl_mem_transfers(const Eigen::SyclDevice &sycl_device) {
     32   IndexType sizeDim1 = 5;
     33   IndexType sizeDim2 = 5;
     34   IndexType sizeDim3 = 1;
     35   array<IndexType, 3> tensorRange = {{sizeDim1, sizeDim2, sizeDim3}};
     36   Tensor<DataType, 3, DataLayout, IndexType> in1(tensorRange);
     37   Tensor<DataType, 3, DataLayout, IndexType> out1(tensorRange);
     38   Tensor<DataType, 3, DataLayout, IndexType> out2(tensorRange);
     39   Tensor<DataType, 3, DataLayout, IndexType> out3(tensorRange);
     40 
     41   in1 = in1.random();
     42 
     43   DataType* gpu_data1  = static_cast<DataType*>(sycl_device.allocate(in1.size()*sizeof(DataType)));
     44   DataType* gpu_data2  = static_cast<DataType*>(sycl_device.allocate(out1.size()*sizeof(DataType)));
     45 
     46   TensorMap<Tensor<DataType, 3, DataLayout, IndexType>> gpu1(gpu_data1, tensorRange);
     47   TensorMap<Tensor<DataType, 3, DataLayout, IndexType>> gpu2(gpu_data2, tensorRange);
     48 
     49   sycl_device.memcpyHostToDevice(gpu_data1, in1.data(),(in1.size())*sizeof(DataType));
     50   sycl_device.memcpyHostToDevice(gpu_data2, in1.data(),(in1.size())*sizeof(DataType));
     51   gpu1.device(sycl_device) = gpu1 * 3.14f;
     52   gpu2.device(sycl_device) = gpu2 * 2.7f;
     53   sycl_device.memcpyDeviceToHost(out1.data(), gpu_data1,(out1.size())*sizeof(DataType));
     54   sycl_device.memcpyDeviceToHost(out2.data(), gpu_data1,(out2.size())*sizeof(DataType));
     55   sycl_device.memcpyDeviceToHost(out3.data(), gpu_data2,(out3.size())*sizeof(DataType));
     56   sycl_device.synchronize();
     57 
     58   for (IndexType i = 0; i < in1.size(); ++i) {
     59   //  std::cout << "SYCL DATA : " << out1(i) << "  vs  CPU DATA : " << in1(i) * 3.14f << "\n";
     60     VERIFY_IS_APPROX(out1(i), in1(i) * 3.14f);
     61     VERIFY_IS_APPROX(out2(i), in1(i) * 3.14f);
     62     VERIFY_IS_APPROX(out3(i), in1(i) * 2.7f);
     63   }
     64 
     65   sycl_device.deallocate(gpu_data1);
     66   sycl_device.deallocate(gpu_data2);
     67 }
     68 
     69 template <typename DataType, int DataLayout, typename IndexType>
     70 void test_sycl_mem_sync(const Eigen::SyclDevice &sycl_device) {
     71   IndexType size = 20;
     72   array<IndexType, 1> tensorRange = {{size}};
     73   Tensor<DataType, 1, DataLayout, IndexType> in1(tensorRange);
     74   Tensor<DataType, 1, DataLayout, IndexType> in2(tensorRange);
     75   Tensor<DataType, 1, DataLayout, IndexType> out(tensorRange);
     76 
     77   in1 = in1.random();
     78   in2 = in1;
     79 
     80   DataType* gpu_data  = static_cast<DataType*>(sycl_device.allocate(in1.size()*sizeof(DataType)));
     81 
     82   TensorMap<Tensor<DataType, 1, DataLayout, IndexType>> gpu1(gpu_data, tensorRange);
     83   sycl_device.memcpyHostToDevice(gpu_data, in1.data(),(in1.size())*sizeof(DataType));
     84   sycl_device.synchronize();
     85   in1.setZero();
     86 
     87   sycl_device.memcpyDeviceToHost(out.data(), gpu_data, out.size()*sizeof(DataType));
     88   sycl_device.synchronize();
     89 
     90   for (IndexType i = 0; i < in1.size(); ++i) {
     91     VERIFY_IS_APPROX(out(i), in2(i));
     92   }
     93 
     94   sycl_device.deallocate(gpu_data);
     95 }
     96 
     97 template <typename DataType, int DataLayout, typename IndexType>
     98 void test_sycl_mem_sync_offsets(const Eigen::SyclDevice &sycl_device) {
     99   using tensor_type = Tensor<DataType, 1, DataLayout, IndexType>;
    100   IndexType full_size = 32;
    101   IndexType half_size = full_size / 2;
    102   array<IndexType, 1> tensorRange = {{full_size}};
    103   tensor_type in1(tensorRange);
    104   tensor_type out(tensorRange);
    105 
    106   DataType* gpu_data  = static_cast<DataType*>(sycl_device.allocate(full_size * sizeof(DataType)));
    107   TensorMap<tensor_type> gpu1(gpu_data, tensorRange);
    108 
    109   in1 = in1.random();
    110   // Copy all data to device, then permute on copy back to host
    111   sycl_device.memcpyHostToDevice(gpu_data, in1.data(), full_size * sizeof(DataType));
    112   sycl_device.memcpyDeviceToHost(out.data(), gpu_data + half_size, half_size * sizeof(DataType));
    113   sycl_device.memcpyDeviceToHost(out.data() + half_size, gpu_data, half_size * sizeof(DataType));
    114 
    115   for (IndexType i = 0; i < half_size; ++i) {
    116     VERIFY_IS_APPROX(out(i), in1(i + half_size));
    117     VERIFY_IS_APPROX(out(i + half_size), in1(i));
    118   }
    119 
    120   in1 = in1.random();
    121   out.setZero();
    122   // Permute copies to device, then copy all back to host
    123   sycl_device.memcpyHostToDevice(gpu_data + half_size, in1.data(), half_size * sizeof(DataType));
    124   sycl_device.memcpyHostToDevice(gpu_data, in1.data() + half_size, half_size * sizeof(DataType));
    125   sycl_device.memcpyDeviceToHost(out.data(), gpu_data, full_size * sizeof(DataType));
    126 
    127   for (IndexType i = 0; i < half_size; ++i) {
    128     VERIFY_IS_APPROX(out(i), in1(i + half_size));
    129     VERIFY_IS_APPROX(out(i + half_size), in1(i));
    130   }
    131 
    132   in1 = in1.random();
    133   out.setZero();
    134   DataType* gpu_data_out  = static_cast<DataType*>(sycl_device.allocate(full_size * sizeof(DataType)));
    135   TensorMap<tensor_type> gpu2(gpu_data_out, tensorRange);
    136   // Copy all to device, permute copies on device, then copy all back to host
    137   sycl_device.memcpyHostToDevice(gpu_data, in1.data(), full_size * sizeof(DataType));
    138   sycl_device.memcpy(gpu_data_out + half_size, gpu_data, half_size * sizeof(DataType));
    139   sycl_device.memcpy(gpu_data_out, gpu_data + half_size, half_size * sizeof(DataType));
    140   sycl_device.memcpyDeviceToHost(out.data(), gpu_data_out, full_size * sizeof(DataType));
    141 
    142   for (IndexType i = 0; i < half_size; ++i) {
    143     VERIFY_IS_APPROX(out(i), in1(i + half_size));
    144     VERIFY_IS_APPROX(out(i + half_size), in1(i));
    145   }
    146 
    147   sycl_device.deallocate(gpu_data_out);
    148   sycl_device.deallocate(gpu_data);
    149 }
    150 
    151 template <typename DataType, int DataLayout, typename IndexType>
    152 void test_sycl_memset_offsets(const Eigen::SyclDevice &sycl_device) {
    153   using tensor_type = Tensor<DataType, 1, DataLayout, IndexType>;
    154   IndexType full_size = 32;
    155   IndexType half_size = full_size / 2;
    156   array<IndexType, 1> tensorRange = {{full_size}};
    157   tensor_type cpu_out(tensorRange);
    158   tensor_type out(tensorRange);
    159 
    160   cpu_out.setZero();
    161 
    162   std::memset(cpu_out.data(), 0, half_size * sizeof(DataType));
    163   std::memset(cpu_out.data() + half_size, 1, half_size * sizeof(DataType));
    164 
    165   DataType* gpu_data  = static_cast<DataType*>(sycl_device.allocate(full_size * sizeof(DataType)));
    166   TensorMap<tensor_type> gpu1(gpu_data, tensorRange);
    167 
    168   sycl_device.memset(gpu_data, 0, half_size * sizeof(DataType));
    169   sycl_device.memset(gpu_data + half_size, 1, half_size * sizeof(DataType));
    170   sycl_device.memcpyDeviceToHost(out.data(), gpu_data, full_size * sizeof(DataType));
    171 
    172   for (IndexType i = 0; i < full_size; ++i) {
    173     VERIFY_IS_APPROX(out(i), cpu_out(i));
    174   }
    175 
    176   sycl_device.deallocate(gpu_data);
    177 }
    178 
    179 template <typename DataType, int DataLayout, typename IndexType>
    180 void test_sycl_computations(const Eigen::SyclDevice &sycl_device) {
    181 
    182   IndexType sizeDim1 = 100;
    183   IndexType sizeDim2 = 10;
    184   IndexType sizeDim3 = 20;
    185   array<IndexType, 3> tensorRange = {{sizeDim1, sizeDim2, sizeDim3}};
    186   Tensor<DataType, 3,DataLayout, IndexType> in1(tensorRange);
    187   Tensor<DataType, 3,DataLayout, IndexType> in2(tensorRange);
    188   Tensor<DataType, 3,DataLayout, IndexType> in3(tensorRange);
    189   Tensor<DataType, 3,DataLayout, IndexType> out(tensorRange);
    190 
    191   in2 = in2.random();
    192   in3 = in3.random();
    193 
    194   DataType * gpu_in1_data  = static_cast<DataType*>(sycl_device.allocate(in1.size()*sizeof(DataType)));
    195   DataType * gpu_in2_data  = static_cast<DataType*>(sycl_device.allocate(in2.size()*sizeof(DataType)));
    196   DataType * gpu_in3_data  = static_cast<DataType*>(sycl_device.allocate(in3.size()*sizeof(DataType)));
    197   DataType * gpu_out_data =  static_cast<DataType*>(sycl_device.allocate(out.size()*sizeof(DataType)));
    198 
    199   TensorMap<Tensor<DataType, 3, DataLayout, IndexType>> gpu_in1(gpu_in1_data, tensorRange);
    200   TensorMap<Tensor<DataType, 3, DataLayout, IndexType>> gpu_in2(gpu_in2_data, tensorRange);
    201   TensorMap<Tensor<DataType, 3, DataLayout, IndexType>> gpu_in3(gpu_in3_data, tensorRange);
    202   TensorMap<Tensor<DataType, 3, DataLayout, IndexType>> gpu_out(gpu_out_data, tensorRange);
    203 
    204   /// a=1.2f
    205   gpu_in1.device(sycl_device) = gpu_in1.constant(1.2f);
    206   sycl_device.memcpyDeviceToHost(in1.data(), gpu_in1_data ,(in1.size())*sizeof(DataType));
    207   sycl_device.synchronize();
    208 
    209   for (IndexType i = 0; i < sizeDim1; ++i) {
    210     for (IndexType j = 0; j < sizeDim2; ++j) {
    211       for (IndexType k = 0; k < sizeDim3; ++k) {
    212         VERIFY_IS_APPROX(in1(i,j,k), 1.2f);
    213       }
    214     }
    215   }
    216   printf("a=1.2f Test passed\n");
    217 
    218   /// a=b*1.2f
    219   gpu_out.device(sycl_device) = gpu_in1 * 1.2f;
    220   sycl_device.memcpyDeviceToHost(out.data(), gpu_out_data ,(out.size())*sizeof(DataType));
    221   sycl_device.synchronize();
    222 
    223   for (IndexType i = 0; i < sizeDim1; ++i) {
    224     for (IndexType j = 0; j < sizeDim2; ++j) {
    225       for (IndexType k = 0; k < sizeDim3; ++k) {
    226         VERIFY_IS_APPROX(out(i,j,k),
    227                          in1(i,j,k) * 1.2f);
    228       }
    229     }
    230   }
    231   printf("a=b*1.2f Test Passed\n");
    232 
    233   /// c=a*b
    234   sycl_device.memcpyHostToDevice(gpu_in2_data, in2.data(),(in2.size())*sizeof(DataType));
    235   gpu_out.device(sycl_device) = gpu_in1 * gpu_in2;
    236   sycl_device.memcpyDeviceToHost(out.data(), gpu_out_data,(out.size())*sizeof(DataType));
    237   sycl_device.synchronize();
    238 
    239   for (IndexType i = 0; i < sizeDim1; ++i) {
    240     for (IndexType j = 0; j < sizeDim2; ++j) {
    241       for (IndexType k = 0; k < sizeDim3; ++k) {
    242         VERIFY_IS_APPROX(out(i,j,k),
    243                          in1(i,j,k) *
    244                              in2(i,j,k));
    245       }
    246     }
    247   }
    248   printf("c=a*b Test Passed\n");
    249 
    250   /// c=a+b
    251   gpu_out.device(sycl_device) = gpu_in1 + gpu_in2;
    252   sycl_device.memcpyDeviceToHost(out.data(), gpu_out_data,(out.size())*sizeof(DataType));
    253   sycl_device.synchronize();
    254   for (IndexType i = 0; i < sizeDim1; ++i) {
    255     for (IndexType j = 0; j < sizeDim2; ++j) {
    256       for (IndexType k = 0; k < sizeDim3; ++k) {
    257         VERIFY_IS_APPROX(out(i,j,k),
    258                          in1(i,j,k) +
    259                              in2(i,j,k));
    260       }
    261     }
    262   }
    263   printf("c=a+b Test Passed\n");
    264 
    265   /// c=a*a
    266   gpu_out.device(sycl_device) = gpu_in1 * gpu_in1;
    267   sycl_device.memcpyDeviceToHost(out.data(), gpu_out_data,(out.size())*sizeof(DataType));
    268   sycl_device.synchronize();
    269   for (IndexType i = 0; i < sizeDim1; ++i) {
    270     for (IndexType j = 0; j < sizeDim2; ++j) {
    271       for (IndexType k = 0; k < sizeDim3; ++k) {
    272         VERIFY_IS_APPROX(out(i,j,k),
    273                          in1(i,j,k) *
    274                              in1(i,j,k));
    275       }
    276     }
    277   }
    278   printf("c= a*a Test Passed\n");
    279 
    280   //a*3.14f + b*2.7f
    281   gpu_out.device(sycl_device) =  gpu_in1 * gpu_in1.constant(3.14f) + gpu_in2 * gpu_in2.constant(2.7f);
    282   sycl_device.memcpyDeviceToHost(out.data(),gpu_out_data,(out.size())*sizeof(DataType));
    283   sycl_device.synchronize();
    284   for (IndexType i = 0; i < sizeDim1; ++i) {
    285     for (IndexType j = 0; j < sizeDim2; ++j) {
    286       for (IndexType k = 0; k < sizeDim3; ++k) {
    287         VERIFY_IS_APPROX(out(i,j,k),
    288                          in1(i,j,k) * 3.14f
    289                        + in2(i,j,k) * 2.7f);
    290       }
    291     }
    292   }
    293   printf("a*3.14f + b*2.7f Test Passed\n");
    294 
    295   ///d= (a>0.5? b:c)
    296   sycl_device.memcpyHostToDevice(gpu_in3_data, in3.data(),(in3.size())*sizeof(DataType));
    297   gpu_out.device(sycl_device) =(gpu_in1 > gpu_in1.constant(0.5f)).select(gpu_in2, gpu_in3);
    298   sycl_device.memcpyDeviceToHost(out.data(), gpu_out_data,(out.size())*sizeof(DataType));
    299   sycl_device.synchronize();
    300   for (IndexType i = 0; i < sizeDim1; ++i) {
    301     for (IndexType j = 0; j < sizeDim2; ++j) {
    302       for (IndexType k = 0; k < sizeDim3; ++k) {
    303         VERIFY_IS_APPROX(out(i, j, k), (in1(i, j, k) > 0.5f)
    304                                                 ? in2(i, j, k)
    305                                                 : in3(i, j, k));
    306       }
    307     }
    308   }
    309   printf("d= (a>0.5? b:c) Test Passed\n");
    310   sycl_device.deallocate(gpu_in1_data);
    311   sycl_device.deallocate(gpu_in2_data);
    312   sycl_device.deallocate(gpu_in3_data);
    313   sycl_device.deallocate(gpu_out_data);
    314 }
    315 template<typename Scalar1, typename Scalar2,  int DataLayout, typename IndexType>
    316 static void test_sycl_cast(const Eigen::SyclDevice& sycl_device){
    317     IndexType size = 20;
    318     array<IndexType, 1> tensorRange = {{size}};
    319     Tensor<Scalar1, 1, DataLayout, IndexType> in(tensorRange);
    320     Tensor<Scalar2, 1, DataLayout, IndexType> out(tensorRange);
    321     Tensor<Scalar2, 1, DataLayout, IndexType> out_host(tensorRange);
    322 
    323     in = in.random();
    324 
    325     Scalar1* gpu_in_data  = static_cast<Scalar1*>(sycl_device.allocate(in.size()*sizeof(Scalar1)));
    326     Scalar2 * gpu_out_data =  static_cast<Scalar2*>(sycl_device.allocate(out.size()*sizeof(Scalar2)));
    327 
    328     TensorMap<Tensor<Scalar1, 1, DataLayout, IndexType>> gpu_in(gpu_in_data, tensorRange);
    329     TensorMap<Tensor<Scalar2, 1, DataLayout, IndexType>> gpu_out(gpu_out_data, tensorRange);
    330     sycl_device.memcpyHostToDevice(gpu_in_data, in.data(),(in.size())*sizeof(Scalar1));
    331     gpu_out.device(sycl_device) = gpu_in. template cast<Scalar2>();
    332     sycl_device.memcpyDeviceToHost(out.data(), gpu_out_data, out.size()*sizeof(Scalar2));
    333     out_host = in. template cast<Scalar2>();
    334     for(IndexType i=0; i< size; i++)
    335     {
    336       VERIFY_IS_APPROX(out(i), out_host(i));
    337     }
    338     printf("cast Test Passed\n");
    339     sycl_device.deallocate(gpu_in_data);
    340     sycl_device.deallocate(gpu_out_data);
    341 }
    342 template<typename DataType, typename dev_Selector> void sycl_computing_test_per_device(dev_Selector s){
    343   QueueInterface queueInterface(s);
    344   auto sycl_device = Eigen::SyclDevice(&queueInterface);
    345   test_sycl_mem_transfers<DataType, RowMajor, int64_t>(sycl_device);
    346   test_sycl_computations<DataType, RowMajor, int64_t>(sycl_device);
    347   test_sycl_mem_sync<DataType, RowMajor, int64_t>(sycl_device);
    348   test_sycl_mem_sync_offsets<DataType, RowMajor, int64_t>(sycl_device);
    349   test_sycl_memset_offsets<DataType, RowMajor, int64_t>(sycl_device);
    350   test_sycl_mem_transfers<DataType, ColMajor, int64_t>(sycl_device);
    351   test_sycl_computations<DataType, ColMajor, int64_t>(sycl_device);
    352   test_sycl_mem_sync<DataType, ColMajor, int64_t>(sycl_device);
    353   test_sycl_cast<DataType, int, RowMajor, int64_t>(sycl_device);
    354   test_sycl_cast<DataType, int, ColMajor, int64_t>(sycl_device);
    355 }
    356 
    357 EIGEN_DECLARE_TEST(cxx11_tensor_sycl) {
    358   for (const auto& device :Eigen::get_sycl_supported_devices()) {
    359     CALL_SUBTEST(sycl_computing_test_per_device<float>(device));
    360   }
    361 }