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 #define EIGEN_TEST_FUNC cxx11_tensor_scan_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 typedef Tensor<float, 1>::DimensionPair DimPair;
24 
25 template<int DataLayout>
26 void test_cuda_cumsum(int m_size, int k_size, int n_size)
27 {
28   std::cout << "Testing for (" << m_size << "," << k_size << "," << n_size << ")" << std::endl;
29   Tensor<float, 3, DataLayout> t_input(m_size, k_size, n_size);
30   Tensor<float, 3, DataLayout> t_result(m_size, k_size, n_size);
31   Tensor<float, 3, DataLayout> t_result_gpu(m_size, k_size, n_size);
32 
33   t_input.setRandom();
34 
35   std::size_t t_input_bytes = t_input.size()  * sizeof(float);
36   std::size_t t_result_bytes = t_result.size() * sizeof(float);
37 
38   float* d_t_input;
39   float* d_t_result;
40 
41   cudaMalloc((void**)(&d_t_input), t_input_bytes);
42   cudaMalloc((void**)(&d_t_result), t_result_bytes);
43 
44   cudaMemcpy(d_t_input, t_input.data(), t_input_bytes, cudaMemcpyHostToDevice);
45 
46   Eigen::CudaStreamDevice stream;
47   Eigen::GpuDevice gpu_device(&stream);
48 
49   Eigen::TensorMap<Eigen::Tensor<float, 3, DataLayout> >
50       gpu_t_input(d_t_input, Eigen::array<int, 3>(m_size, k_size, n_size));
51   Eigen::TensorMap<Eigen::Tensor<float, 3, DataLayout> >
52       gpu_t_result(d_t_result, Eigen::array<int, 3>(m_size, k_size, n_size));
53 
54   gpu_t_result.device(gpu_device) = gpu_t_input.cumsum(1);
55   t_result = t_input.cumsum(1);
56 
57   cudaMemcpy(t_result_gpu.data(), d_t_result, t_result_bytes, cudaMemcpyDeviceToHost);
58   for (DenseIndex i = 0; i < t_result.size(); i++) {
59     if (fabs(t_result(i) - t_result_gpu(i)) < 1e-4f) {
60       continue;
61     }
62     if (Eigen::internal::isApprox(t_result(i), t_result_gpu(i), 1e-4f)) {
63       continue;
64     }
65     std::cout << "mismatch detected at index " << i << ": " << t_result(i)
66               << " vs " <<  t_result_gpu(i) << std::endl;
67     assert(false);
68   }
69 
70   cudaFree((void*)d_t_input);
71   cudaFree((void*)d_t_result);
72 }
73 
74 
75 void test_cxx11_tensor_scan_cuda()
76 {
77   CALL_SUBTEST_1(test_cuda_cumsum<ColMajor>(128, 128, 128));
78   CALL_SUBTEST_2(test_cuda_cumsum<RowMajor>(128, 128, 128));
79 }
80