1 // This file is part of Eigen, a lightweight C++ template library
2 // for linear algebra.
3 //
4 // Copyright (C) 2014 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 #define EIGEN_TEST_FUNC cxx11_tensor_random_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 <Eigen/CXX11/Tensor>
21 
22 
23 void test_cuda_random_uniform()
24 {
25   Tensor<float, 2> out(72,97);
26   out.setZero();
27 
28   std::size_t out_bytes = out.size() * sizeof(float);
29 
30   float* d_out;
31   cudaMalloc((void**)(&d_out), out_bytes);
32 
33   Eigen::CudaStreamDevice stream;
34   Eigen::GpuDevice gpu_device(&stream);
35 
36   Eigen::TensorMap<Eigen::Tensor<float, 2> > gpu_out(d_out, 72,97);
37 
38   gpu_out.device(gpu_device) = gpu_out.random();
39 
40   assert(cudaMemcpyAsync(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost, gpu_device.stream()) == cudaSuccess);
41   assert(cudaStreamSynchronize(gpu_device.stream()) == cudaSuccess);
42 
43   // For now we just check thes code doesn't crash.
44   // TODO: come up with a valid test of randomness
45 }
46 
47 
48 void test_cuda_random_normal()
49 {
50   Tensor<float, 2> out(72,97);
51   out.setZero();
52 
53   std::size_t out_bytes = out.size() * sizeof(float);
54 
55   float* d_out;
56   cudaMalloc((void**)(&d_out), out_bytes);
57 
58   Eigen::CudaStreamDevice stream;
59   Eigen::GpuDevice gpu_device(&stream);
60 
61   Eigen::TensorMap<Eigen::Tensor<float, 2> > gpu_out(d_out, 72,97);
62 
63   Eigen::internal::NormalRandomGenerator<float> gen(true);
64   gpu_out.device(gpu_device) = gpu_out.random(gen);
65 
66   assert(cudaMemcpyAsync(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost, gpu_device.stream()) == cudaSuccess);
67   assert(cudaStreamSynchronize(gpu_device.stream()) == cudaSuccess);
68 }
69 
70 static void test_complex()
71 {
72   Tensor<std::complex<float>, 1> vec(6);
73   vec.setRandom();
74 
75   // Fixme: we should check that the generated numbers follow a uniform
76   // distribution instead.
77   for (int i = 1; i < 6; ++i) {
78     VERIFY_IS_NOT_EQUAL(vec(i), vec(i-1));
79   }
80 }
81 
82 
83 void test_cxx11_tensor_random_cuda()
84 {
85   CALL_SUBTEST(test_cuda_random_uniform());
86   CALL_SUBTEST(test_cuda_random_normal());
87   CALL_SUBTEST(test_complex());
88 }
89