cxx11_tensor_scan_gpu.cu (2576B)
1 // This file is part of Eigen, a lightweight C++ template library 2 // for linear algebra. 3 // 4 // Copyright (C) 2016 Benoit Steiner <benoit.steiner.goog@gmail.com> 5 // 6 // This Source Code Form is subject to the terms of the Mozilla 7 // Public License v. 2.0. If a copy of the MPL was not distributed 8 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/. 9 10 #define EIGEN_TEST_NO_LONGDOUBLE 11 #define EIGEN_TEST_NO_COMPLEX 12 13 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int 14 #define EIGEN_USE_GPU 15 16 #include "main.h" 17 #include <unsupported/Eigen/CXX11/Tensor> 18 19 #include <Eigen/CXX11/src/Tensor/TensorGpuHipCudaDefines.h> 20 21 using Eigen::Tensor; 22 typedef Tensor<float, 1>::DimensionPair DimPair; 23 24 template<int DataLayout> 25 void test_gpu_cumsum(int m_size, int k_size, int n_size) 26 { 27 std::cout << "Testing for (" << m_size << "," << k_size << "," << n_size << ")" << std::endl; 28 Tensor<float, 3, DataLayout> t_input(m_size, k_size, n_size); 29 Tensor<float, 3, DataLayout> t_result(m_size, k_size, n_size); 30 Tensor<float, 3, DataLayout> t_result_gpu(m_size, k_size, n_size); 31 32 t_input.setRandom(); 33 34 std::size_t t_input_bytes = t_input.size() * sizeof(float); 35 std::size_t t_result_bytes = t_result.size() * sizeof(float); 36 37 float* d_t_input; 38 float* d_t_result; 39 40 gpuMalloc((void**)(&d_t_input), t_input_bytes); 41 gpuMalloc((void**)(&d_t_result), t_result_bytes); 42 43 gpuMemcpy(d_t_input, t_input.data(), t_input_bytes, gpuMemcpyHostToDevice); 44 45 Eigen::GpuStreamDevice stream; 46 Eigen::GpuDevice gpu_device(&stream); 47 48 Eigen::TensorMap<Eigen::Tensor<float, 3, DataLayout> > 49 gpu_t_input(d_t_input, Eigen::array<int, 3>(m_size, k_size, n_size)); 50 Eigen::TensorMap<Eigen::Tensor<float, 3, DataLayout> > 51 gpu_t_result(d_t_result, Eigen::array<int, 3>(m_size, k_size, n_size)); 52 53 gpu_t_result.device(gpu_device) = gpu_t_input.cumsum(1); 54 t_result = t_input.cumsum(1); 55 56 gpuMemcpy(t_result_gpu.data(), d_t_result, t_result_bytes, gpuMemcpyDeviceToHost); 57 for (DenseIndex i = 0; i < t_result.size(); i++) { 58 if (fabs(t_result(i) - t_result_gpu(i)) < 1e-4f) { 59 continue; 60 } 61 if (Eigen::internal::isApprox(t_result(i), t_result_gpu(i), 1e-4f)) { 62 continue; 63 } 64 std::cout << "mismatch detected at index " << i << ": " << t_result(i) 65 << " vs " << t_result_gpu(i) << std::endl; 66 assert(false); 67 } 68 69 gpuFree((void*)d_t_input); 70 gpuFree((void*)d_t_result); 71 } 72 73 74 EIGEN_DECLARE_TEST(cxx11_tensor_scan_gpu) 75 { 76 CALL_SUBTEST_1(test_gpu_cumsum<ColMajor>(128, 128, 128)); 77 CALL_SUBTEST_2(test_gpu_cumsum<RowMajor>(128, 128, 128)); 78 }