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 (a] 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 #define EIGEN_TEST_FUNC cxx11_tensor_cast_float16_cuda 13 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int 14 #define EIGEN_USE_GPU 15 16 #if defined __CUDACC_VER__ && __CUDACC_VER__ >= 70500 17 #include <cuda_fp16.h> 18 #endif 19 #include "main.h" 20 #include <unsupported/Eigen/CXX11/Tensor> 21 22 using Eigen::Tensor; 23 24 void test_cuda_conversion() { 25 Eigen::CudaStreamDevice stream; 26 Eigen::GpuDevice gpu_device(&stream); 27 int num_elem = 101; 28 29 Tensor<float, 1> floats(num_elem); 30 floats.setRandom(); 31 32 float* d_float = (float*)gpu_device.allocate(num_elem * sizeof(float)); 33 Eigen::half* d_half = (Eigen::half*)gpu_device.allocate(num_elem * sizeof(Eigen::half)); 34 float* d_conv = (float*)gpu_device.allocate(num_elem * sizeof(float)); 35 36 Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_float( 37 d_float, num_elem); 38 Eigen::TensorMap<Eigen::Tensor<Eigen::half, 1>, Eigen::Aligned> gpu_half( 39 d_half, num_elem); 40 Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_conv( 41 d_conv, num_elem); 42 43 gpu_device.memcpyHostToDevice(d_float, floats.data(), num_elem*sizeof(float)); 44 45 gpu_half.device(gpu_device) = gpu_float.cast<Eigen::half>(); 46 gpu_conv.device(gpu_device) = gpu_half.cast<float>(); 47 48 Tensor<float, 1> initial(num_elem); 49 Tensor<float, 1> final(num_elem); 50 gpu_device.memcpyDeviceToHost(initial.data(), d_float, num_elem*sizeof(float)); 51 gpu_device.memcpyDeviceToHost(final.data(), d_conv, num_elem*sizeof(float)); 52 gpu_device.synchronize(); 53 54 for (int i = 0; i < num_elem; ++i) { 55 VERIFY_IS_APPROX(initial(i), final(i)); 56 } 57 58 gpu_device.deallocate(d_float); 59 gpu_device.deallocate(d_half); 60 gpu_device.deallocate(d_conv); 61 } 62 63 64 void test_fallback_conversion() { 65 int num_elem = 101; 66 Tensor<float, 1> floats(num_elem); 67 floats.setRandom(); 68 69 Eigen::Tensor<Eigen::half, 1> halfs = floats.cast<Eigen::half>(); 70 Eigen::Tensor<float, 1> conv = halfs.cast<float>(); 71 72 for (int i = 0; i < num_elem; ++i) { 73 VERIFY_IS_APPROX(floats(i), conv(i)); 74 } 75 } 76 77 78 void test_cxx11_tensor_cast_float16_cuda() 79 { 80 CALL_SUBTEST(test_cuda_conversion()); 81 CALL_SUBTEST(test_fallback_conversion()); 82 } 83
