• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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_device
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 using Eigen::RowMajor;
24 
25 // Context for evaluation on cpu
26 struct CPUContext {
CPUContextCPUContext27   CPUContext(const Eigen::Tensor<float, 3>& in1, Eigen::Tensor<float, 3>& in2, Eigen::Tensor<float, 3>& out) : in1_(in1), in2_(in2), out_(out), kernel_1d_(2), kernel_2d_(2,2), kernel_3d_(2,2,2) {
28     kernel_1d_(0) = 3.14f;
29     kernel_1d_(1) = 2.7f;
30 
31     kernel_2d_(0,0) = 3.14f;
32     kernel_2d_(1,0) = 2.7f;
33     kernel_2d_(0,1) = 0.2f;
34     kernel_2d_(1,1) = 7.0f;
35 
36     kernel_3d_(0,0,0) = 3.14f;
37     kernel_3d_(0,1,0) = 2.7f;
38     kernel_3d_(0,0,1) = 0.2f;
39     kernel_3d_(0,1,1) = 7.0f;
40     kernel_3d_(1,0,0) = -1.0f;
41     kernel_3d_(1,1,0) = -0.3f;
42     kernel_3d_(1,0,1) = -0.7f;
43     kernel_3d_(1,1,1) = -0.5f;
44   }
45 
deviceCPUContext46   const Eigen::DefaultDevice& device() const { return cpu_device_; }
47 
in1CPUContext48   const Eigen::Tensor<float, 3>& in1() const { return in1_; }
in2CPUContext49   const Eigen::Tensor<float, 3>& in2() const { return in2_; }
outCPUContext50   Eigen::Tensor<float, 3>& out() { return out_; }
kernel1dCPUContext51   const Eigen::Tensor<float, 1>& kernel1d() const { return kernel_1d_; }
kernel2dCPUContext52   const Eigen::Tensor<float, 2>& kernel2d() const { return kernel_2d_; }
kernel3dCPUContext53   const Eigen::Tensor<float, 3>& kernel3d() const { return kernel_3d_; }
54 
55  private:
56   const Eigen::Tensor<float, 3>& in1_;
57   const Eigen::Tensor<float, 3>& in2_;
58   Eigen::Tensor<float, 3>& out_;
59 
60   Eigen::Tensor<float, 1> kernel_1d_;
61   Eigen::Tensor<float, 2> kernel_2d_;
62   Eigen::Tensor<float, 3> kernel_3d_;
63 
64   Eigen::DefaultDevice cpu_device_;
65 };
66 
67 
68 // Context for evaluation on GPU
69 struct GPUContext {
GPUContextGPUContext70   GPUContext(const Eigen::TensorMap<Eigen::Tensor<float, 3> >& in1, Eigen::TensorMap<Eigen::Tensor<float, 3> >& in2, Eigen::TensorMap<Eigen::Tensor<float, 3> >& out) : in1_(in1), in2_(in2), out_(out), gpu_device_(&stream_) {
71     assert(cudaMalloc((void**)(&kernel_1d_), 2*sizeof(float)) == cudaSuccess);
72     float kernel_1d_val[] = {3.14f, 2.7f};
73     assert(cudaMemcpy(kernel_1d_, kernel_1d_val, 2*sizeof(float), cudaMemcpyHostToDevice) == cudaSuccess);
74 
75     assert(cudaMalloc((void**)(&kernel_2d_), 4*sizeof(float)) == cudaSuccess);
76     float kernel_2d_val[] = {3.14f, 2.7f, 0.2f, 7.0f};
77     assert(cudaMemcpy(kernel_2d_, kernel_2d_val, 4*sizeof(float), cudaMemcpyHostToDevice) == cudaSuccess);
78 
79     assert(cudaMalloc((void**)(&kernel_3d_), 8*sizeof(float)) == cudaSuccess);
80     float kernel_3d_val[] = {3.14f, -1.0f, 2.7f, -0.3f, 0.2f, -0.7f, 7.0f, -0.5f};
81     assert(cudaMemcpy(kernel_3d_, kernel_3d_val, 8*sizeof(float), cudaMemcpyHostToDevice) == cudaSuccess);
82   }
~GPUContextGPUContext83   ~GPUContext() {
84     assert(cudaFree(kernel_1d_) == cudaSuccess);
85     assert(cudaFree(kernel_2d_) == cudaSuccess);
86     assert(cudaFree(kernel_3d_) == cudaSuccess);
87   }
88 
deviceGPUContext89   const Eigen::GpuDevice& device() const { return gpu_device_; }
90 
in1GPUContext91   const Eigen::TensorMap<Eigen::Tensor<float, 3> >& in1() const { return in1_; }
in2GPUContext92   const Eigen::TensorMap<Eigen::Tensor<float, 3> >& in2() const { return in2_; }
outGPUContext93   Eigen::TensorMap<Eigen::Tensor<float, 3> >& out() { return out_; }
kernel1dGPUContext94   Eigen::TensorMap<Eigen::Tensor<float, 1> > kernel1d() const { return Eigen::TensorMap<Eigen::Tensor<float, 1> >(kernel_1d_, 2); }
kernel2dGPUContext95   Eigen::TensorMap<Eigen::Tensor<float, 2> > kernel2d() const { return Eigen::TensorMap<Eigen::Tensor<float, 2> >(kernel_2d_, 2, 2); }
kernel3dGPUContext96   Eigen::TensorMap<Eigen::Tensor<float, 3> > kernel3d() const { return Eigen::TensorMap<Eigen::Tensor<float, 3> >(kernel_3d_, 2, 2, 2); }
97 
98  private:
99   const Eigen::TensorMap<Eigen::Tensor<float, 3> >& in1_;
100   const Eigen::TensorMap<Eigen::Tensor<float, 3> >& in2_;
101   Eigen::TensorMap<Eigen::Tensor<float, 3> >& out_;
102 
103   float* kernel_1d_;
104   float* kernel_2d_;
105   float* kernel_3d_;
106 
107   Eigen::CudaStreamDevice stream_;
108   Eigen::GpuDevice gpu_device_;
109 };
110 
111 
112 // The actual expression to evaluate
113 template <typename Context>
test_contextual_eval(Context * context)114 void test_contextual_eval(Context* context)
115 {
116   context->out().device(context->device()) = context->in1() + context->in2() * 3.14f + context->in1().constant(2.718f);
117 }
118 
119 template <typename Context>
test_forced_contextual_eval(Context * context)120 void test_forced_contextual_eval(Context* context)
121 {
122   context->out().device(context->device()) = (context->in1() + context->in2()).eval() * 3.14f + context->in1().constant(2.718f);
123 }
124 
125 template <typename Context>
test_compound_assignment(Context * context)126 void test_compound_assignment(Context* context)
127 {
128   context->out().device(context->device()) = context->in1().constant(2.718f);
129   context->out().device(context->device()) += context->in1() + context->in2() * 3.14f;
130 }
131 
132 
133 template <typename Context>
test_contraction(Context * context)134 void test_contraction(Context* context)
135 {
136   Eigen::array<std::pair<int, int>, 2> dims;
137   dims[0] = std::make_pair(1, 1);
138   dims[1] = std::make_pair(2, 2);
139 
140   Eigen::array<int, 2> shape(40, 50*70);
141 
142   Eigen::DSizes<int, 2> indices(0,0);
143   Eigen::DSizes<int, 2> sizes(40,40);
144 
145   context->out().reshape(shape).slice(indices, sizes).device(context->device()) = context->in1().contract(context->in2(), dims);
146 }
147 
148 
149 template <typename Context>
test_1d_convolution(Context * context)150 void test_1d_convolution(Context* context)
151 {
152   Eigen::DSizes<int, 3> indices(0,0,0);
153   Eigen::DSizes<int, 3> sizes(40,49,70);
154 
155   Eigen::array<int, 1> dims(1);
156   context->out().slice(indices, sizes).device(context->device()) = context->in1().convolve(context->kernel1d(), dims);
157 }
158 
159 template <typename Context>
test_2d_convolution(Context * context)160 void test_2d_convolution(Context* context)
161 {
162   Eigen::DSizes<int, 3> indices(0,0,0);
163   Eigen::DSizes<int, 3> sizes(40,49,69);
164 
165   Eigen::array<int, 2> dims(1,2);
166   context->out().slice(indices, sizes).device(context->device()) = context->in1().convolve(context->kernel2d(), dims);
167 }
168 
169 template <typename Context>
test_3d_convolution(Context * context)170 void test_3d_convolution(Context* context)
171 {
172   Eigen::DSizes<int, 3> indices(0,0,0);
173   Eigen::DSizes<int, 3> sizes(39,49,69);
174 
175   Eigen::array<int, 3> dims(0,1,2);
176   context->out().slice(indices, sizes).device(context->device()) = context->in1().convolve(context->kernel3d(), dims);
177 }
178 
179 
test_cpu()180 void test_cpu() {
181   Eigen::Tensor<float, 3> in1(40,50,70);
182   Eigen::Tensor<float, 3> in2(40,50,70);
183   Eigen::Tensor<float, 3> out(40,50,70);
184 
185   in1 = in1.random() + in1.constant(10.0f);
186   in2 = in2.random() + in2.constant(10.0f);
187 
188   CPUContext context(in1, in2, out);
189   test_contextual_eval(&context);
190   for (int i = 0; i < 40; ++i) {
191     for (int j = 0; j < 50; ++j) {
192       for (int k = 0; k < 70; ++k) {
193         VERIFY_IS_APPROX(out(i,j,k), in1(i,j,k) + in2(i,j,k) * 3.14f + 2.718f);
194       }
195     }
196   }
197 
198   test_forced_contextual_eval(&context);
199   for (int i = 0; i < 40; ++i) {
200     for (int j = 0; j < 50; ++j) {
201       for (int k = 0; k < 70; ++k) {
202         VERIFY_IS_APPROX(out(i,j,k), (in1(i,j,k) + in2(i,j,k)) * 3.14f + 2.718f);
203       }
204     }
205   }
206 
207   test_compound_assignment(&context);
208   for (int i = 0; i < 40; ++i) {
209     for (int j = 0; j < 50; ++j) {
210       for (int k = 0; k < 70; ++k) {
211         VERIFY_IS_APPROX(out(i,j,k), in1(i,j,k) + in2(i,j,k) * 3.14f + 2.718f);
212       }
213     }
214   }
215 
216   test_contraction(&context);
217   for (int i = 0; i < 40; ++i) {
218     for (int j = 0; j < 40; ++j) {
219       const float result = out(i,j,0);
220       float expected = 0;
221       for (int k = 0; k < 50; ++k) {
222         for (int l = 0; l < 70; ++l) {
223           expected += in1(i, k, l) * in2(j, k, l);
224         }
225       }
226       VERIFY_IS_APPROX(expected, result);
227     }
228   }
229 
230   test_1d_convolution(&context);
231   for (int i = 0; i < 40; ++i) {
232     for (int j = 0; j < 49; ++j) {
233       for (int k = 0; k < 70; ++k) {
234         VERIFY_IS_APPROX(out(i,j,k), (in1(i,j,k) * 3.14f + in1(i,j+1,k) * 2.7f));
235       }
236     }
237   }
238 
239   test_2d_convolution(&context);
240   for (int i = 0; i < 40; ++i) {
241     for (int j = 0; j < 49; ++j) {
242       for (int k = 0; k < 69; ++k) {
243         const float result = out(i,j,k);
244         const float expected = (in1(i,j,k) * 3.14f + in1(i,j+1,k) * 2.7f) +
245                                (in1(i,j,k+1) * 0.2f + in1(i,j+1,k+1) * 7.0f);
246         if (fabs(expected) < 1e-4f && fabs(result) < 1e-4f) {
247           continue;
248         }
249         VERIFY_IS_APPROX(expected, result);
250       }
251     }
252   }
253 
254   test_3d_convolution(&context);
255   for (int i = 0; i < 39; ++i) {
256     for (int j = 0; j < 49; ++j) {
257       for (int k = 0; k < 69; ++k) {
258         const float result = out(i,j,k);
259         const float expected = (in1(i,j,k) * 3.14f + in1(i,j+1,k) * 2.7f +
260                                 in1(i,j,k+1) * 0.2f + in1(i,j+1,k+1) * 7.0f) +
261                                (in1(i+1,j,k) * -1.0f + in1(i+1,j+1,k) * -0.3f +
262                                 in1(i+1,j,k+1) * -0.7f + in1(i+1,j+1,k+1) * -0.5f);
263         if (fabs(expected) < 1e-4f && fabs(result) < 1e-4f) {
264           continue;
265         }
266         VERIFY_IS_APPROX(expected, result);
267       }
268     }
269   }
270 }
271 
test_gpu()272 void test_gpu() {
273   Eigen::Tensor<float, 3> in1(40,50,70);
274   Eigen::Tensor<float, 3> in2(40,50,70);
275   Eigen::Tensor<float, 3> out(40,50,70);
276   in1 = in1.random() + in1.constant(10.0f);
277   in2 = in2.random() + in2.constant(10.0f);
278 
279   std::size_t in1_bytes = in1.size() * sizeof(float);
280   std::size_t in2_bytes = in2.size() * sizeof(float);
281   std::size_t out_bytes = out.size() * sizeof(float);
282 
283   float* d_in1;
284   float* d_in2;
285   float* d_out;
286   cudaMalloc((void**)(&d_in1), in1_bytes);
287   cudaMalloc((void**)(&d_in2), in2_bytes);
288   cudaMalloc((void**)(&d_out), out_bytes);
289 
290   cudaMemcpy(d_in1, in1.data(), in1_bytes, cudaMemcpyHostToDevice);
291   cudaMemcpy(d_in2, in2.data(), in2_bytes, cudaMemcpyHostToDevice);
292 
293   Eigen::TensorMap<Eigen::Tensor<float, 3> > gpu_in1(d_in1, 40,50,70);
294   Eigen::TensorMap<Eigen::Tensor<float, 3> > gpu_in2(d_in2, 40,50,70);
295   Eigen::TensorMap<Eigen::Tensor<float, 3> > gpu_out(d_out, 40,50,70);
296 
297   GPUContext context(gpu_in1, gpu_in2, gpu_out);
298   test_contextual_eval(&context);
299   assert(cudaMemcpy(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost) == cudaSuccess);
300   for (int i = 0; i < 40; ++i) {
301     for (int j = 0; j < 50; ++j) {
302       for (int k = 0; k < 70; ++k) {
303         VERIFY_IS_APPROX(out(i,j,k), in1(i,j,k) + in2(i,j,k) * 3.14f + 2.718f);
304       }
305     }
306   }
307 
308   test_forced_contextual_eval(&context);
309   assert(cudaMemcpy(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost) == cudaSuccess);
310   for (int i = 0; i < 40; ++i) {
311     for (int j = 0; j < 50; ++j) {
312       for (int k = 0; k < 70; ++k) {
313         VERIFY_IS_APPROX(out(i,j,k), (in1(i,j,k) + in2(i,j,k)) * 3.14f + 2.718f);
314       }
315     }
316   }
317 
318   test_compound_assignment(&context);
319   assert(cudaMemcpy(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost) == cudaSuccess);
320   for (int i = 0; i < 40; ++i) {
321     for (int j = 0; j < 50; ++j) {
322       for (int k = 0; k < 70; ++k) {
323         VERIFY_IS_APPROX(out(i,j,k), in1(i,j,k) + in2(i,j,k) * 3.14f + 2.718f);
324       }
325     }
326   }
327 
328   test_contraction(&context);
329   assert(cudaMemcpy(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost) == cudaSuccess);
330   for (int i = 0; i < 40; ++i) {
331     for (int j = 0; j < 40; ++j) {
332       const float result = out(i,j,0);
333       float expected = 0;
334       for (int k = 0; k < 50; ++k) {
335         for (int l = 0; l < 70; ++l) {
336           expected += in1(i, k, l) * in2(j, k, l);
337         }
338       }
339       VERIFY_IS_APPROX(expected, result);
340     }
341   }
342 
343   test_1d_convolution(&context);
344   assert(cudaMemcpyAsync(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost, context.device().stream()) == cudaSuccess);
345   assert(cudaStreamSynchronize(context.device().stream()) == cudaSuccess);
346   for (int i = 0; i < 40; ++i) {
347     for (int j = 0; j < 49; ++j) {
348       for (int k = 0; k < 70; ++k) {
349         VERIFY_IS_APPROX(out(i,j,k), (in1(i,j,k) * 3.14f + in1(i,j+1,k) * 2.7f));
350       }
351     }
352   }
353 
354   test_2d_convolution(&context);
355   assert(cudaMemcpyAsync(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost, context.device().stream()) == cudaSuccess);
356   assert(cudaStreamSynchronize(context.device().stream()) == cudaSuccess);
357   for (int i = 0; i < 40; ++i) {
358     for (int j = 0; j < 49; ++j) {
359       for (int k = 0; k < 69; ++k) {
360         const float result = out(i,j,k);
361         const float expected = (in1(i,j,k) * 3.14f + in1(i,j+1,k) * 2.7f +
362                                 in1(i,j,k+1) * 0.2f + in1(i,j+1,k+1) * 7.0f);
363         VERIFY_IS_APPROX(expected, result);
364       }
365     }
366   }
367 
368   test_3d_convolution(&context);
369   assert(cudaMemcpyAsync(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost, context.device().stream()) == cudaSuccess);
370   assert(cudaStreamSynchronize(context.device().stream()) == cudaSuccess);
371   for (int i = 0; i < 39; ++i) {
372     for (int j = 0; j < 49; ++j) {
373       for (int k = 0; k < 69; ++k) {
374        const float result = out(i,j,k);
375         const float expected = (in1(i,j,k) * 3.14f + in1(i,j+1,k) * 2.7f +
376                                 in1(i,j,k+1) * 0.2f + in1(i,j+1,k+1) * 7.0f +
377                                 in1(i+1,j,k) * -1.0f + in1(i+1,j+1,k) * -0.3f +
378                                 in1(i+1,j,k+1) * -0.7f + in1(i+1,j+1,k+1) * -0.5f);
379         VERIFY_IS_APPROX(expected, result);
380       }
381     }
382   }
383 }
384 
385 
test_cxx11_tensor_device()386 void test_cxx11_tensor_device()
387 {
388   CALL_SUBTEST_1(test_cpu());
389   CALL_SUBTEST_2(test_gpu());
390 }
391