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 // Copyright (C) 2014 Navdeep Jaitly <ndjaitly@google.com>
6 //
7 // This Source Code Form is subject to the terms of the Mozilla
8 // Public License v. 2.0. If a copy of the MPL was not distributed
9 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
10 
11 #define EIGEN_TEST_NO_LONGDOUBLE
12 #define EIGEN_TEST_NO_COMPLEX
13 #define EIGEN_TEST_FUNC cxx11_tensor_cuda
14 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int
15 #define EIGEN_USE_GPU
16 
17 #if defined __CUDACC_VER__ && __CUDACC_VER__ >= 70500
18 #include <cuda_fp16.h>
19 #endif
20 #include "main.h"
21 #include <unsupported/Eigen/CXX11/Tensor>
22 
23 using Eigen::Tensor;
24 typedef Tensor<float, 1>::DimensionPair DimPair;
25 
26 template<int DataLayout>
test_cuda_contraction(int m_size,int k_size,int n_size)27 void test_cuda_contraction(int m_size, int k_size, int n_size)
28 {
29   std::cout << "Testing for (" << m_size << "," << k_size << "," << n_size << ")" << std::endl;
30   // with these dimensions, the output has 300 * 140 elements, which is
31   // more than 30 * 1024, which is the number of threads in blocks on
32   // a 15 SM GK110 GPU
33   Tensor<float, 2, DataLayout> t_left(m_size, k_size);
34   Tensor<float, 2, DataLayout> t_right(k_size, n_size);
35   Tensor<float, 2, DataLayout> t_result(m_size, n_size);
36   Tensor<float, 2, DataLayout> t_result_gpu(m_size, n_size);
37   Eigen::array<DimPair, 1> dims(DimPair(1, 0));
38 
39   t_left.setRandom();
40   t_right.setRandom();
41 
42   std::size_t t_left_bytes = t_left.size()  * sizeof(float);
43   std::size_t t_right_bytes = t_right.size() * sizeof(float);
44   std::size_t t_result_bytes = t_result.size() * sizeof(float);
45 
46   float* d_t_left;
47   float* d_t_right;
48   float* d_t_result;
49 
50   cudaMalloc((void**)(&d_t_left), t_left_bytes);
51   cudaMalloc((void**)(&d_t_right), t_right_bytes);
52   cudaMalloc((void**)(&d_t_result), t_result_bytes);
53 
54   cudaMemcpy(d_t_left, t_left.data(), t_left_bytes, cudaMemcpyHostToDevice);
55   cudaMemcpy(d_t_right, t_right.data(), t_right_bytes, cudaMemcpyHostToDevice);
56 
57   Eigen::CudaStreamDevice stream;
58   Eigen::GpuDevice gpu_device(&stream);
59 
60   Eigen::TensorMap<Eigen::Tensor<float, 2, DataLayout> >
61       gpu_t_left(d_t_left, Eigen::array<int, 2>(m_size, k_size));
62   Eigen::TensorMap<Eigen::Tensor<float, 2, DataLayout> >
63       gpu_t_right(d_t_right, Eigen::array<int, 2>(k_size, n_size));
64   Eigen::TensorMap<Eigen::Tensor<float, 2, DataLayout> >
65       gpu_t_result(d_t_result, Eigen::array<int, 2>(m_size, n_size));
66 
67 
68   gpu_t_result.device(gpu_device) = gpu_t_left.contract(gpu_t_right, dims);
69   t_result = t_left.contract(t_right, dims);
70 
71   cudaMemcpy(t_result_gpu.data(), d_t_result, t_result_bytes, cudaMemcpyDeviceToHost);
72   for (DenseIndex i = 0; i < t_result.size(); i++) {
73     if (fabs(t_result(i) - t_result_gpu(i)) < 1e-4f) {
74       continue;
75     }
76     if (Eigen::internal::isApprox(t_result(i), t_result_gpu(i), 1e-4f)) {
77       continue;
78     }
79     std::cout << "mismatch detected at index " << i << ": " << t_result(i)
80               << " vs " <<  t_result_gpu(i) << std::endl;
81     assert(false);
82   }
83 
84   cudaFree((void*)d_t_left);
85   cudaFree((void*)d_t_right);
86   cudaFree((void*)d_t_result);
87 }
88 
89 
90 template<int DataLayout>
test_scalar(int m_size,int k_size,int n_size)91 void test_scalar(int m_size, int k_size, int n_size)
92 {
93   std::cout << "Testing for (" << m_size << "," << k_size << "," << n_size << ")" << std::endl;
94   // with these dimensions, the output has 300 * 140 elements, which is
95   // more than 30 * 1024, which is the number of threads in blocks on
96   // a 15 SM GK110 GPU
97   Tensor<float, 2, DataLayout> t_left(m_size, k_size);
98   Tensor<float, 2, DataLayout> t_right(k_size, n_size);
99   Tensor<float, 0, DataLayout> t_result;
100   Tensor<float, 0, DataLayout> t_result_gpu;
101   Eigen::array<DimPair, 2> dims(DimPair(0, 0), DimPair(1, 1));
102 
103   t_left.setRandom();
104   t_right.setRandom();
105 
106   std::size_t t_left_bytes = t_left.size()  * sizeof(float);
107   std::size_t t_right_bytes = t_right.size() * sizeof(float);
108   std::size_t t_result_bytes = sizeof(float);
109 
110   float* d_t_left;
111   float* d_t_right;
112   float* d_t_result;
113 
114   cudaMalloc((void**)(&d_t_left), t_left_bytes);
115   cudaMalloc((void**)(&d_t_right), t_right_bytes);
116   cudaMalloc((void**)(&d_t_result), t_result_bytes);
117 
118   cudaMemcpy(d_t_left, t_left.data(), t_left_bytes, cudaMemcpyHostToDevice);
119   cudaMemcpy(d_t_right, t_right.data(), t_right_bytes, cudaMemcpyHostToDevice);
120 
121   Eigen::CudaStreamDevice stream;
122   Eigen::GpuDevice gpu_device(&stream);
123 
124   Eigen::TensorMap<Eigen::Tensor<float, 2, DataLayout> >
125       gpu_t_left(d_t_left, m_size, k_size);
126   Eigen::TensorMap<Eigen::Tensor<float, 2, DataLayout> >
127       gpu_t_right(d_t_right, k_size, n_size);
128   Eigen::TensorMap<Eigen::Tensor<float, 0, DataLayout> >
129       gpu_t_result(d_t_result);
130 
131   gpu_t_result.device(gpu_device) = gpu_t_left.contract(gpu_t_right, dims);
132   t_result = t_left.contract(t_right, dims);
133 
134   cudaMemcpy(t_result_gpu.data(), d_t_result, t_result_bytes, cudaMemcpyDeviceToHost);
135   if (fabs(t_result() - t_result_gpu()) > 1e-4f &&
136       !Eigen::internal::isApprox(t_result(), t_result_gpu(), 1e-4f)) {
137     std::cout << "mismatch detected: " << t_result()
138               << " vs " <<  t_result_gpu() << std::endl;
139     assert(false);
140   }
141 
142   cudaFree((void*)d_t_left);
143   cudaFree((void*)d_t_right);
144   cudaFree((void*)d_t_result);
145 }
146 
147 
148 template<int DataLayout>
test_cuda_contraction_m()149 void test_cuda_contraction_m() {
150   for (int k = 32; k < 256; k++) {
151     test_cuda_contraction<ColMajor>(k, 128, 128);
152     test_cuda_contraction<RowMajor>(k, 128, 128);
153   }
154 }
155 
156 template<int DataLayout>
test_cuda_contraction_k()157 void test_cuda_contraction_k() {
158   for (int k = 32; k < 256; k++) {
159     test_cuda_contraction<ColMajor>(128, k, 128);
160     test_cuda_contraction<RowMajor>(128, k, 128);
161   }
162 }
163 
164 template<int DataLayout>
test_cuda_contraction_n()165 void test_cuda_contraction_n() {
166   for (int k = 32; k < 256; k++) {
167     test_cuda_contraction<ColMajor>(128, 128, k);
168     test_cuda_contraction<RowMajor>(128, 128, k);
169   }
170 }
171 
172 
173 template<int DataLayout>
test_cuda_contraction_sizes()174 void test_cuda_contraction_sizes() {
175   int m_sizes[] = { 31,  39,   63,   64,   65,
176                    127, 129,  255,  257 , 511,
177                    512, 513, 1023, 1024, 1025};
178 
179   int n_sizes[] = { 31,  39,   63,   64,   65,
180                    127, 129,  255,  257,  511,
181                    512, 513, 1023, 1024, 1025};
182 
183   int k_sizes[] = {  31,   39,  63,  64,   65,
184                      95,   96, 127, 129,  255,
185                     257,  511, 512, 513, 1023,
186                    1024, 1025};
187 
188   for (int i = 0; i < 15; i++) {
189     for (int j = 0; j < 15; j++) {
190       for (int k = 0; k < 17; k++) {
191         test_cuda_contraction<DataLayout>(m_sizes[i], n_sizes[j], k_sizes[k]);
192       }
193     }
194   }
195 }
196 
test_cxx11_tensor_cuda()197 void test_cxx11_tensor_cuda()
198 {
199   CALL_SUBTEST_1(test_cuda_contraction<ColMajor>(128, 128, 128));
200   CALL_SUBTEST_1(test_cuda_contraction<RowMajor>(128, 128, 128));
201 
202   CALL_SUBTEST_1(test_scalar<ColMajor>(128, 128, 128));
203   CALL_SUBTEST_1(test_scalar<RowMajor>(128, 128, 128));
204 
205   CALL_SUBTEST_2(test_cuda_contraction_m<ColMajor>());
206   CALL_SUBTEST_3(test_cuda_contraction_m<RowMajor>());
207 
208   CALL_SUBTEST_4(test_cuda_contraction_k<ColMajor>());
209   CALL_SUBTEST_5(test_cuda_contraction_k<RowMajor>());
210 
211   CALL_SUBTEST_6(test_cuda_contraction_n<ColMajor>());
212   CALL_SUBTEST_7(test_cuda_contraction_n<RowMajor>());
213 
214   CALL_SUBTEST_8(test_cuda_contraction_sizes<ColMajor>());
215   CALL_SUBTEST_9(test_cuda_contraction_sizes<RowMajor>());
216 }
217