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_of_float16_cuda
13 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE int
14 #define EIGEN_USE_GPU
15
16 #include "main.h"
17 #include <unsupported/Eigen/CXX11/Tensor>
18
19 using Eigen::Tensor;
20
21 template<typename>
test_cuda_numext()22 void test_cuda_numext() {
23 Eigen::CudaStreamDevice stream;
24 Eigen::GpuDevice gpu_device(&stream);
25 int num_elem = 101;
26
27 float* d_float = (float*)gpu_device.allocate(num_elem * sizeof(float));
28 bool* d_res_half = (bool*)gpu_device.allocate(num_elem * sizeof(bool));
29 bool* d_res_float = (bool*)gpu_device.allocate(num_elem * sizeof(bool));
30
31 Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_float(
32 d_float, num_elem);
33 Eigen::TensorMap<Eigen::Tensor<bool, 1>, Eigen::Aligned> gpu_res_half(
34 d_res_half, num_elem);
35 Eigen::TensorMap<Eigen::Tensor<bool, 1>, Eigen::Aligned> gpu_res_float(
36 d_res_float, num_elem);
37
38 gpu_float.device(gpu_device) = gpu_float.random() - gpu_float.constant(0.5f);
39 gpu_res_float.device(gpu_device) = gpu_float.unaryExpr(Eigen::internal::scalar_isnan_op<float>());
40 gpu_res_half.device(gpu_device) = gpu_float.cast<Eigen::half>().unaryExpr(Eigen::internal::scalar_isnan_op<Eigen::half>());
41
42 Tensor<bool, 1> half_prec(num_elem);
43 Tensor<bool, 1> full_prec(num_elem);
44 gpu_device.memcpyDeviceToHost(half_prec.data(), d_res_half, num_elem*sizeof(bool));
45 gpu_device.memcpyDeviceToHost(full_prec.data(), d_res_float, num_elem*sizeof(bool));
46 gpu_device.synchronize();
47
48 for (int i = 0; i < num_elem; ++i) {
49 std::cout << "Checking numext " << i << std::endl;
50 VERIFY_IS_EQUAL(full_prec(i), half_prec(i));
51 }
52
53 gpu_device.deallocate(d_float);
54 gpu_device.deallocate(d_res_half);
55 gpu_device.deallocate(d_res_float);
56 }
57
58
59 #ifdef EIGEN_HAS_CUDA_FP16
60
61 template<typename>
test_cuda_conversion()62 void test_cuda_conversion() {
63 Eigen::CudaStreamDevice stream;
64 Eigen::GpuDevice gpu_device(&stream);
65 int num_elem = 101;
66
67 float* d_float = (float*)gpu_device.allocate(num_elem * sizeof(float));
68 Eigen::half* d_half = (Eigen::half*)gpu_device.allocate(num_elem * sizeof(Eigen::half));
69 float* d_conv = (float*)gpu_device.allocate(num_elem * sizeof(float));
70
71 Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_float(
72 d_float, num_elem);
73 Eigen::TensorMap<Eigen::Tensor<Eigen::half, 1>, Eigen::Aligned> gpu_half(
74 d_half, num_elem);
75 Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_conv(
76 d_conv, num_elem);
77
78 gpu_float.device(gpu_device) = gpu_float.random();
79 gpu_half.device(gpu_device) = gpu_float.cast<Eigen::half>();
80 gpu_conv.device(gpu_device) = gpu_half.cast<float>();
81
82 Tensor<float, 1> initial(num_elem);
83 Tensor<float, 1> final(num_elem);
84 gpu_device.memcpyDeviceToHost(initial.data(), d_float, num_elem*sizeof(float));
85 gpu_device.memcpyDeviceToHost(final.data(), d_conv, num_elem*sizeof(float));
86
87 for (int i = 0; i < num_elem; ++i) {
88 VERIFY_IS_APPROX(initial(i), final(i));
89 }
90
91 gpu_device.deallocate(d_float);
92 gpu_device.deallocate(d_half);
93 gpu_device.deallocate(d_conv);
94 }
95
96 template<typename>
test_cuda_unary()97 void test_cuda_unary() {
98 Eigen::CudaStreamDevice stream;
99 Eigen::GpuDevice gpu_device(&stream);
100 int num_elem = 101;
101
102 float* d_float = (float*)gpu_device.allocate(num_elem * sizeof(float));
103 float* d_res_half = (float*)gpu_device.allocate(num_elem * sizeof(float));
104 float* d_res_float = (float*)gpu_device.allocate(num_elem * sizeof(float));
105
106 Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_float(
107 d_float, num_elem);
108 Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_res_half(
109 d_res_half, num_elem);
110 Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_res_float(
111 d_res_float, num_elem);
112
113 gpu_float.device(gpu_device) = gpu_float.random() - gpu_float.constant(0.5f);
114 gpu_res_float.device(gpu_device) = gpu_float.abs();
115 gpu_res_half.device(gpu_device) = gpu_float.cast<Eigen::half>().abs().cast<float>();
116
117 Tensor<float, 1> half_prec(num_elem);
118 Tensor<float, 1> full_prec(num_elem);
119 gpu_device.memcpyDeviceToHost(half_prec.data(), d_res_half, num_elem*sizeof(float));
120 gpu_device.memcpyDeviceToHost(full_prec.data(), d_res_float, num_elem*sizeof(float));
121 gpu_device.synchronize();
122
123 for (int i = 0; i < num_elem; ++i) {
124 std::cout << "Checking unary " << i << std::endl;
125 VERIFY_IS_APPROX(full_prec(i), half_prec(i));
126 }
127
128 gpu_device.deallocate(d_float);
129 gpu_device.deallocate(d_res_half);
130 gpu_device.deallocate(d_res_float);
131 }
132
133 template<typename>
test_cuda_elementwise()134 void test_cuda_elementwise() {
135 Eigen::CudaStreamDevice stream;
136 Eigen::GpuDevice gpu_device(&stream);
137 int num_elem = 101;
138
139 float* d_float1 = (float*)gpu_device.allocate(num_elem * sizeof(float));
140 float* d_float2 = (float*)gpu_device.allocate(num_elem * sizeof(float));
141 float* d_res_half = (float*)gpu_device.allocate(num_elem * sizeof(float));
142 float* d_res_float = (float*)gpu_device.allocate(num_elem * sizeof(float));
143
144 Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_float1(
145 d_float1, num_elem);
146 Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_float2(
147 d_float2, num_elem);
148 Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_res_half(
149 d_res_half, num_elem);
150 Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_res_float(
151 d_res_float, num_elem);
152
153 gpu_float1.device(gpu_device) = gpu_float1.random();
154 gpu_float2.device(gpu_device) = gpu_float2.random();
155 gpu_res_float.device(gpu_device) = (gpu_float1 + gpu_float2) * gpu_float1;
156 gpu_res_half.device(gpu_device) = ((gpu_float1.cast<Eigen::half>() + gpu_float2.cast<Eigen::half>()) * gpu_float1.cast<Eigen::half>()).cast<float>();
157
158 Tensor<float, 1> half_prec(num_elem);
159 Tensor<float, 1> full_prec(num_elem);
160 gpu_device.memcpyDeviceToHost(half_prec.data(), d_res_half, num_elem*sizeof(float));
161 gpu_device.memcpyDeviceToHost(full_prec.data(), d_res_float, num_elem*sizeof(float));
162 gpu_device.synchronize();
163
164 for (int i = 0; i < num_elem; ++i) {
165 std::cout << "Checking elemwise " << i << ": full prec = " << full_prec(i) << " vs half prec = " << half_prec(i) << std::endl;
166 VERIFY_IS_APPROX(static_cast<Eigen::half>(full_prec(i)), static_cast<Eigen::half>(half_prec(i)));
167 }
168
169 gpu_device.deallocate(d_float1);
170 gpu_device.deallocate(d_float2);
171 gpu_device.deallocate(d_res_half);
172 gpu_device.deallocate(d_res_float);
173 }
174
175 template<typename>
test_cuda_trancendental()176 void test_cuda_trancendental() {
177 Eigen::CudaStreamDevice stream;
178 Eigen::GpuDevice gpu_device(&stream);
179 int num_elem = 101;
180
181 float* d_float1 = (float*)gpu_device.allocate(num_elem * sizeof(float));
182 float* d_float2 = (float*)gpu_device.allocate(num_elem * sizeof(float));
183 float* d_float3 = (float*)gpu_device.allocate(num_elem * sizeof(float));
184 Eigen::half* d_res1_half = (Eigen::half*)gpu_device.allocate(num_elem * sizeof(Eigen::half));
185 Eigen::half* d_res1_float = (Eigen::half*)gpu_device.allocate(num_elem * sizeof(Eigen::half));
186 Eigen::half* d_res2_half = (Eigen::half*)gpu_device.allocate(num_elem * sizeof(Eigen::half));
187 Eigen::half* d_res2_float = (Eigen::half*)gpu_device.allocate(num_elem * sizeof(Eigen::half));
188 Eigen::half* d_res3_half = (Eigen::half*)gpu_device.allocate(num_elem * sizeof(Eigen::half));
189 Eigen::half* d_res3_float = (Eigen::half*)gpu_device.allocate(num_elem * sizeof(Eigen::half));
190
191 Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_float1(d_float1, num_elem);
192 Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_float2(d_float2, num_elem);
193 Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_float3(d_float3, num_elem);
194 Eigen::TensorMap<Eigen::Tensor<Eigen::half, 1>, Eigen::Aligned> gpu_res1_half(d_res1_half, num_elem);
195 Eigen::TensorMap<Eigen::Tensor<Eigen::half, 1>, Eigen::Aligned> gpu_res1_float(d_res1_float, num_elem);
196 Eigen::TensorMap<Eigen::Tensor<Eigen::half, 1>, Eigen::Aligned> gpu_res2_half(d_res2_half, num_elem);
197 Eigen::TensorMap<Eigen::Tensor<Eigen::half, 1>, Eigen::Aligned> gpu_res2_float(d_res2_float, num_elem);
198 Eigen::TensorMap<Eigen::Tensor<Eigen::half, 1>, Eigen::Aligned> gpu_res3_half(d_res3_half, num_elem);
199 Eigen::TensorMap<Eigen::Tensor<Eigen::half, 1>, Eigen::Aligned> gpu_res3_float(d_res3_float, num_elem);
200
201 gpu_float1.device(gpu_device) = gpu_float1.random() - gpu_float1.constant(0.5f);
202 gpu_float2.device(gpu_device) = gpu_float2.random() + gpu_float1.constant(0.5f);
203 gpu_float3.device(gpu_device) = gpu_float3.random();
204 gpu_res1_float.device(gpu_device) = gpu_float1.exp().cast<Eigen::half>();
205 gpu_res2_float.device(gpu_device) = gpu_float2.log().cast<Eigen::half>();
206 gpu_res3_float.device(gpu_device) = gpu_float3.log1p().cast<Eigen::half>();
207
208 gpu_res1_half.device(gpu_device) = gpu_float1.cast<Eigen::half>();
209 gpu_res1_half.device(gpu_device) = gpu_res1_half.exp();
210
211 gpu_res2_half.device(gpu_device) = gpu_float2.cast<Eigen::half>();
212 gpu_res2_half.device(gpu_device) = gpu_res2_half.log();
213
214 gpu_res3_half.device(gpu_device) = gpu_float3.cast<Eigen::half>();
215 gpu_res3_half.device(gpu_device) = gpu_res3_half.log1p();
216
217 Tensor<float, 1> input1(num_elem);
218 Tensor<Eigen::half, 1> half_prec1(num_elem);
219 Tensor<Eigen::half, 1> full_prec1(num_elem);
220 Tensor<float, 1> input2(num_elem);
221 Tensor<Eigen::half, 1> half_prec2(num_elem);
222 Tensor<Eigen::half, 1> full_prec2(num_elem);
223 Tensor<float, 1> input3(num_elem);
224 Tensor<Eigen::half, 1> half_prec3(num_elem);
225 Tensor<Eigen::half, 1> full_prec3(num_elem);
226 gpu_device.memcpyDeviceToHost(input1.data(), d_float1, num_elem*sizeof(float));
227 gpu_device.memcpyDeviceToHost(input2.data(), d_float2, num_elem*sizeof(float));
228 gpu_device.memcpyDeviceToHost(input3.data(), d_float3, num_elem*sizeof(float));
229 gpu_device.memcpyDeviceToHost(half_prec1.data(), d_res1_half, num_elem*sizeof(Eigen::half));
230 gpu_device.memcpyDeviceToHost(full_prec1.data(), d_res1_float, num_elem*sizeof(Eigen::half));
231 gpu_device.memcpyDeviceToHost(half_prec2.data(), d_res2_half, num_elem*sizeof(Eigen::half));
232 gpu_device.memcpyDeviceToHost(full_prec2.data(), d_res2_float, num_elem*sizeof(Eigen::half));
233 gpu_device.memcpyDeviceToHost(half_prec3.data(), d_res3_half, num_elem*sizeof(Eigen::half));
234 gpu_device.memcpyDeviceToHost(full_prec3.data(), d_res3_float, num_elem*sizeof(Eigen::half));
235 gpu_device.synchronize();
236
237 for (int i = 0; i < num_elem; ++i) {
238 std::cout << "Checking elemwise exp " << i << " input = " << input1(i) << " full = " << full_prec1(i) << " half = " << half_prec1(i) << std::endl;
239 VERIFY_IS_APPROX(full_prec1(i), half_prec1(i));
240 }
241 for (int i = 0; i < num_elem; ++i) {
242 std::cout << "Checking elemwise log " << i << " input = " << input2(i) << " full = " << full_prec2(i) << " half = " << half_prec2(i) << std::endl;
243 if(std::abs(input2(i)-1.f)<0.05f) // log lacks accurary nearby 1
244 VERIFY_IS_APPROX(full_prec2(i)+Eigen::half(0.1f), half_prec2(i)+Eigen::half(0.1f));
245 else
246 VERIFY_IS_APPROX(full_prec2(i), half_prec2(i));
247 }
248 for (int i = 0; i < num_elem; ++i) {
249 std::cout << "Checking elemwise plog1 " << i << " input = " << input3(i) << " full = " << full_prec3(i) << " half = " << half_prec3(i) << std::endl;
250 VERIFY_IS_APPROX(full_prec3(i), half_prec3(i));
251 }
252 gpu_device.deallocate(d_float1);
253 gpu_device.deallocate(d_float2);
254 gpu_device.deallocate(d_float3);
255 gpu_device.deallocate(d_res1_half);
256 gpu_device.deallocate(d_res1_float);
257 gpu_device.deallocate(d_res2_half);
258 gpu_device.deallocate(d_res2_float);
259 gpu_device.deallocate(d_res3_float);
260 gpu_device.deallocate(d_res3_half);
261 }
262
263 template<typename>
test_cuda_contractions()264 void test_cuda_contractions() {
265 Eigen::CudaStreamDevice stream;
266 Eigen::GpuDevice gpu_device(&stream);
267 int rows = 23;
268 int cols = 23;
269 int num_elem = rows*cols;
270
271 float* d_float1 = (float*)gpu_device.allocate(num_elem * sizeof(float));
272 float* d_float2 = (float*)gpu_device.allocate(num_elem * sizeof(float));
273 Eigen::half* d_res_half = (Eigen::half*)gpu_device.allocate(num_elem * sizeof(Eigen::half));
274 Eigen::half* d_res_float = (Eigen::half*)gpu_device.allocate(num_elem * sizeof(Eigen::half));
275
276 Eigen::TensorMap<Eigen::Tensor<float, 2>, Eigen::Aligned> gpu_float1(
277 d_float1, rows, cols);
278 Eigen::TensorMap<Eigen::Tensor<float, 2>, Eigen::Aligned> gpu_float2(
279 d_float2, rows, cols);
280 Eigen::TensorMap<Eigen::Tensor<Eigen::half, 2>, Eigen::Aligned> gpu_res_half(
281 d_res_half, rows, cols);
282 Eigen::TensorMap<Eigen::Tensor<Eigen::half, 2>, Eigen::Aligned> gpu_res_float(
283 d_res_float, rows, cols);
284
285 gpu_float1.device(gpu_device) = gpu_float1.random() - gpu_float1.constant(0.5f);
286 gpu_float2.device(gpu_device) = gpu_float2.random() - gpu_float2.constant(0.5f);
287
288 typedef Tensor<float, 2>::DimensionPair DimPair;
289 Eigen::array<DimPair, 1> dims(DimPair(1, 0));
290 gpu_res_float.device(gpu_device) = gpu_float1.contract(gpu_float2, dims).cast<Eigen::half>();
291 gpu_res_half.device(gpu_device) = gpu_float1.cast<Eigen::half>().contract(gpu_float2.cast<Eigen::half>(), dims);
292
293 Tensor<Eigen::half, 2> half_prec(rows, cols);
294 Tensor<Eigen::half, 2> full_prec(rows, cols);
295 gpu_device.memcpyDeviceToHost(half_prec.data(), d_res_half, num_elem*sizeof(Eigen::half));
296 gpu_device.memcpyDeviceToHost(full_prec.data(), d_res_float, num_elem*sizeof(Eigen::half));
297 gpu_device.synchronize();
298
299 for (int i = 0; i < rows; ++i) {
300 for (int j = 0; j < cols; ++j) {
301 std::cout << "Checking contract " << i << " " << j << full_prec(i, j) << " " << half_prec(i, j) << std::endl;
302 if (numext::abs(full_prec(i, j) - half_prec(i, j)) > Eigen::half(1e-2f)) {
303 VERIFY_IS_APPROX(full_prec(i, j), half_prec(i, j));
304 }
305 }
306 }
307
308 gpu_device.deallocate(d_float1);
309 gpu_device.deallocate(d_float2);
310 gpu_device.deallocate(d_res_half);
311 gpu_device.deallocate(d_res_float);
312 }
313
314 template<typename>
test_cuda_reductions(int size1,int size2,int redux)315 void test_cuda_reductions(int size1, int size2, int redux) {
316
317 std::cout << "Reducing " << size1 << " by " << size2
318 << " tensor along dim " << redux << std::endl;
319
320 Eigen::CudaStreamDevice stream;
321 Eigen::GpuDevice gpu_device(&stream);
322 int num_elem = size1*size2;
323 int result_size = (redux == 1 ? size1 : size2);
324
325 float* d_float1 = (float*)gpu_device.allocate(num_elem * sizeof(float));
326 float* d_float2 = (float*)gpu_device.allocate(num_elem * sizeof(float));
327 Eigen::half* d_res_half = (Eigen::half*)gpu_device.allocate(result_size * sizeof(Eigen::half));
328 Eigen::half* d_res_float = (Eigen::half*)gpu_device.allocate(result_size * sizeof(Eigen::half));
329
330 Eigen::TensorMap<Eigen::Tensor<float, 2>, Eigen::Aligned> gpu_float1(
331 d_float1, size1, size2);
332 Eigen::TensorMap<Eigen::Tensor<float, 2>, Eigen::Aligned> gpu_float2(
333 d_float2, size1, size2);
334 Eigen::TensorMap<Eigen::Tensor<Eigen::half, 1>, Eigen::Aligned> gpu_res_half(
335 d_res_half, result_size);
336 Eigen::TensorMap<Eigen::Tensor<Eigen::half, 1>, Eigen::Aligned> gpu_res_float(
337 d_res_float, result_size);
338
339 gpu_float1.device(gpu_device) = gpu_float1.random() * 2.0f;
340 gpu_float2.device(gpu_device) = gpu_float2.random() * 2.0f;
341
342 Eigen::array<int, 1> redux_dim = {{redux}};
343 gpu_res_float.device(gpu_device) = gpu_float1.sum(redux_dim).cast<Eigen::half>();
344 gpu_res_half.device(gpu_device) = gpu_float1.cast<Eigen::half>().sum(redux_dim);
345
346 Tensor<Eigen::half, 1> half_prec(result_size);
347 Tensor<Eigen::half, 1> full_prec(result_size);
348 gpu_device.memcpyDeviceToHost(half_prec.data(), d_res_half, result_size*sizeof(Eigen::half));
349 gpu_device.memcpyDeviceToHost(full_prec.data(), d_res_float, result_size*sizeof(Eigen::half));
350 gpu_device.synchronize();
351
352 for (int i = 0; i < result_size; ++i) {
353 std::cout << "EXPECTED " << full_prec(i) << " GOT " << half_prec(i) << std::endl;
354 VERIFY_IS_APPROX(full_prec(i), half_prec(i));
355 }
356
357 gpu_device.deallocate(d_float1);
358 gpu_device.deallocate(d_float2);
359 gpu_device.deallocate(d_res_half);
360 gpu_device.deallocate(d_res_float);
361 }
362
363 template<typename>
test_cuda_reductions()364 void test_cuda_reductions() {
365 test_cuda_reductions<void>(13, 13, 0);
366 test_cuda_reductions<void>(13, 13, 1);
367
368 test_cuda_reductions<void>(35, 36, 0);
369 test_cuda_reductions<void>(35, 36, 1);
370
371 test_cuda_reductions<void>(36, 35, 0);
372 test_cuda_reductions<void>(36, 35, 1);
373 }
374
375 template<typename>
test_cuda_full_reductions()376 void test_cuda_full_reductions() {
377 Eigen::CudaStreamDevice stream;
378 Eigen::GpuDevice gpu_device(&stream);
379 int size = 13;
380 int num_elem = size*size;
381
382 float* d_float1 = (float*)gpu_device.allocate(num_elem * sizeof(float));
383 float* d_float2 = (float*)gpu_device.allocate(num_elem * sizeof(float));
384 Eigen::half* d_res_half = (Eigen::half*)gpu_device.allocate(1 * sizeof(Eigen::half));
385 Eigen::half* d_res_float = (Eigen::half*)gpu_device.allocate(1 * sizeof(Eigen::half));
386
387 Eigen::TensorMap<Eigen::Tensor<float, 2>, Eigen::Aligned> gpu_float1(
388 d_float1, size, size);
389 Eigen::TensorMap<Eigen::Tensor<float, 2>, Eigen::Aligned> gpu_float2(
390 d_float2, size, size);
391 Eigen::TensorMap<Eigen::Tensor<Eigen::half, 0>, Eigen::Aligned> gpu_res_half(
392 d_res_half);
393 Eigen::TensorMap<Eigen::Tensor<Eigen::half, 0>, Eigen::Aligned> gpu_res_float(
394 d_res_float);
395
396 gpu_float1.device(gpu_device) = gpu_float1.random();
397 gpu_float2.device(gpu_device) = gpu_float2.random();
398
399 gpu_res_float.device(gpu_device) = gpu_float1.sum().cast<Eigen::half>();
400 gpu_res_half.device(gpu_device) = gpu_float1.cast<Eigen::half>().sum();
401
402 Tensor<Eigen::half, 0> half_prec;
403 Tensor<Eigen::half, 0> full_prec;
404 gpu_device.memcpyDeviceToHost(half_prec.data(), d_res_half, sizeof(Eigen::half));
405 gpu_device.memcpyDeviceToHost(full_prec.data(), d_res_float, sizeof(Eigen::half));
406 gpu_device.synchronize();
407
408 VERIFY_IS_APPROX(full_prec(), half_prec());
409
410 gpu_res_float.device(gpu_device) = gpu_float1.maximum().cast<Eigen::half>();
411 gpu_res_half.device(gpu_device) = gpu_float1.cast<Eigen::half>().maximum();
412 gpu_device.memcpyDeviceToHost(half_prec.data(), d_res_half, sizeof(Eigen::half));
413 gpu_device.memcpyDeviceToHost(full_prec.data(), d_res_float, sizeof(Eigen::half));
414 gpu_device.synchronize();
415
416 VERIFY_IS_APPROX(full_prec(), half_prec());
417
418 gpu_device.deallocate(d_float1);
419 gpu_device.deallocate(d_float2);
420 gpu_device.deallocate(d_res_half);
421 gpu_device.deallocate(d_res_float);
422 }
423
424 template<typename>
test_cuda_forced_evals()425 void test_cuda_forced_evals() {
426
427 Eigen::CudaStreamDevice stream;
428 Eigen::GpuDevice gpu_device(&stream);
429 int num_elem = 101;
430
431 float* d_float = (float*)gpu_device.allocate(num_elem * sizeof(float));
432 float* d_res_half1 = (float*)gpu_device.allocate(num_elem * sizeof(float));
433 float* d_res_half2 = (float*)gpu_device.allocate(num_elem * sizeof(float));
434 float* d_res_float = (float*)gpu_device.allocate(num_elem * sizeof(float));
435
436 Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_float(
437 d_float, num_elem);
438 Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_res_half1(
439 d_res_half1, num_elem);
440 Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Unaligned> gpu_res_half2(
441 d_res_half2, num_elem);
442 Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_res_float(
443 d_res_float, num_elem);
444
445 Eigen::array<int, 1> no_bcast;
446 no_bcast[0] = 1;
447
448 gpu_float.device(gpu_device) = gpu_float.random() - gpu_float.constant(0.5f);
449 gpu_res_float.device(gpu_device) = gpu_float.abs();
450 gpu_res_half1.device(gpu_device) = gpu_float.cast<Eigen::half>().abs().eval().cast<float>();
451 gpu_res_half2.device(gpu_device) = gpu_float.cast<Eigen::half>().abs().broadcast(no_bcast).eval().cast<float>();
452
453 Tensor<float, 1> half_prec1(num_elem);
454 Tensor<float, 1> half_prec2(num_elem);
455 Tensor<float, 1> full_prec(num_elem);
456 gpu_device.memcpyDeviceToHost(half_prec1.data(), d_res_half1, num_elem*sizeof(float));
457 gpu_device.memcpyDeviceToHost(half_prec2.data(), d_res_half1, num_elem*sizeof(float));
458 gpu_device.memcpyDeviceToHost(full_prec.data(), d_res_float, num_elem*sizeof(float));
459 gpu_device.synchronize();
460
461 for (int i = 0; i < num_elem; ++i) {
462 std::cout << "Checking forced eval " << i << full_prec(i) << " vs " << half_prec1(i) << " vs " << half_prec2(i) << std::endl;
463 VERIFY_IS_APPROX(full_prec(i), half_prec1(i));
464 VERIFY_IS_APPROX(full_prec(i), half_prec2(i));
465 }
466
467 gpu_device.deallocate(d_float);
468 gpu_device.deallocate(d_res_half1);
469 gpu_device.deallocate(d_res_half2);
470 gpu_device.deallocate(d_res_float);
471 }
472 #endif
473
474
test_cxx11_tensor_of_float16_cuda()475 void test_cxx11_tensor_of_float16_cuda()
476 {
477 CALL_SUBTEST_1(test_cuda_numext<void>());
478
479 #ifdef EIGEN_HAS_CUDA_FP16
480 CALL_SUBTEST_1(test_cuda_conversion<void>());
481 CALL_SUBTEST_1(test_cuda_unary<void>());
482 CALL_SUBTEST_1(test_cuda_elementwise<void>());
483 CALL_SUBTEST_1(test_cuda_trancendental<void>());
484 CALL_SUBTEST_2(test_cuda_contractions<void>());
485 CALL_SUBTEST_3(test_cuda_reductions<void>());
486 CALL_SUBTEST_4(test_cuda_full_reductions<void>());
487 CALL_SUBTEST_5(test_cuda_forced_evals<void>());
488 #else
489 std::cout << "Half floats are not supported by this version of cuda: skipping the test" << std::endl;
490 #endif
491 }
492