1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
|
// This file is part of Eigen, a lightweight C++ template library
// for linear algebra.
//
// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
//
// This Source Code Form is subject to the terms of the Mozilla
// Public License v. 2.0. If a copy of the MPL was not distributed
// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
#define EIGEN_TEST_NO_LONGDOUBLE
#define EIGEN_TEST_FUNC cxx11_tensor_cuda
#define EIGEN_USE_GPU
#include "main.h"
#include <unsupported/Eigen/CXX11/Tensor>
using Eigen::Tensor;
template <int Layout>
void test_cuda_simple_argmax()
{
Tensor<double, 3, Layout> in(Eigen::array<DenseIndex, 3>(72,53,97));
Tensor<DenseIndex, 1, Layout> out_max(Eigen::array<DenseIndex, 1>(1));
Tensor<DenseIndex, 1, Layout> out_min(Eigen::array<DenseIndex, 1>(1));
in.setRandom();
in *= in.constant(100.0);
in(0, 0, 0) = -1000.0;
in(71, 52, 96) = 1000.0;
std::size_t in_bytes = in.size() * sizeof(double);
std::size_t out_bytes = out_max.size() * sizeof(DenseIndex);
double* d_in;
DenseIndex* d_out_max;
DenseIndex* d_out_min;
cudaMalloc((void**)(&d_in), in_bytes);
cudaMalloc((void**)(&d_out_max), out_bytes);
cudaMalloc((void**)(&d_out_min), out_bytes);
cudaMemcpy(d_in, in.data(), in_bytes, cudaMemcpyHostToDevice);
Eigen::CudaStreamDevice stream;
Eigen::GpuDevice gpu_device(&stream);
Eigen::TensorMap<Eigen::Tensor<double, 3, Layout>, Aligned > gpu_in(d_in, Eigen::array<DenseIndex, 3>(72,53,97));
Eigen::TensorMap<Eigen::Tensor<DenseIndex, 1, Layout>, Aligned > gpu_out_max(d_out_max, Eigen::array<DenseIndex, 1>(1));
Eigen::TensorMap<Eigen::Tensor<DenseIndex, 1, Layout>, Aligned > gpu_out_min(d_out_min, Eigen::array<DenseIndex, 1>(1));
gpu_out_max.device(gpu_device) = gpu_in.argmax();
gpu_out_min.device(gpu_device) = gpu_in.argmin();
assert(cudaMemcpyAsync(out_max.data(), d_out_max, out_bytes, cudaMemcpyDeviceToHost, gpu_device.stream()) == cudaSuccess);
assert(cudaMemcpyAsync(out_min.data(), d_out_min, out_bytes, cudaMemcpyDeviceToHost, gpu_device.stream()) == cudaSuccess);
assert(cudaStreamSynchronize(gpu_device.stream()) == cudaSuccess);
VERIFY_IS_EQUAL(out_max(Eigen::array<DenseIndex, 1>(0)), 72*53*97 - 1);
VERIFY_IS_EQUAL(out_min(Eigen::array<DenseIndex, 1>(0)), 0);
cudaFree(d_in);
cudaFree(d_out_max);
cudaFree(d_out_min);
}
template <int DataLayout>
void test_cuda_argmax_dim()
{
Tensor<float, 4, DataLayout> tensor(2,3,5,7);
std::vector<int> dims;
dims.push_back(2); dims.push_back(3); dims.push_back(5); dims.push_back(7);
for (int dim = 0; dim < 4; ++dim) {
tensor.setRandom();
tensor = (tensor + tensor.constant(0.5)).log();
array<DenseIndex, 3> out_shape;
for (int d = 0; d < 3; ++d) out_shape[d] = (d < dim) ? dims[d] : dims[d+1];
Tensor<DenseIndex, 3, DataLayout> tensor_arg(out_shape);
array<DenseIndex, 4> ix;
for (int i = 0; i < 2; ++i) {
for (int j = 0; j < 3; ++j) {
for (int k = 0; k < 5; ++k) {
for (int l = 0; l < 7; ++l) {
ix[0] = i; ix[1] = j; ix[2] = k; ix[3] = l;
if (ix[dim] != 0) continue;
// suppose dim == 1, then for all i, k, l, set tensor(i, 0, k, l) = 10.0
tensor(ix) = 10.0;
}
}
}
}
std::size_t in_bytes = tensor.size() * sizeof(float);
std::size_t out_bytes = tensor_arg.size() * sizeof(DenseIndex);
float* d_in;
DenseIndex* d_out;
cudaMalloc((void**)(&d_in), in_bytes);
cudaMalloc((void**)(&d_out), out_bytes);
cudaMemcpy(d_in, tensor.data(), in_bytes, cudaMemcpyHostToDevice);
Eigen::CudaStreamDevice stream;
Eigen::GpuDevice gpu_device(&stream);
Eigen::TensorMap<Eigen::Tensor<float, 4, DataLayout>, Aligned > gpu_in(d_in, Eigen::array<DenseIndex, 4>(2, 3, 5, 7));
Eigen::TensorMap<Eigen::Tensor<DenseIndex, 3, DataLayout>, Aligned > gpu_out(d_out, out_shape);
gpu_out.device(gpu_device) = gpu_in.argmax(dim);
assert(cudaMemcpyAsync(tensor_arg.data(), d_out, out_bytes, cudaMemcpyDeviceToHost, gpu_device.stream()) == cudaSuccess);
assert(cudaStreamSynchronize(gpu_device.stream()) == cudaSuccess);
VERIFY_IS_EQUAL(tensor_arg.size(),
size_t(2*3*5*7 / tensor.dimension(dim)));
for (DenseIndex n = 0; n < tensor_arg.size(); ++n) {
// Expect max to be in the first index of the reduced dimension
VERIFY_IS_EQUAL(tensor_arg.data()[n], 0);
}
for (int i = 0; i < 2; ++i) {
for (int j = 0; j < 3; ++j) {
for (int k = 0; k < 5; ++k) {
for (int l = 0; l < 7; ++l) {
ix[0] = i; ix[1] = j; ix[2] = k; ix[3] = l;
if (ix[dim] != tensor.dimension(dim) - 1) continue;
// suppose dim == 1, then for all i, k, l, set tensor(i, 2, k, l) = 20.0
tensor(ix) = 20.0;
}
}
}
}
cudaMemcpy(d_in, tensor.data(), in_bytes, cudaMemcpyHostToDevice);
gpu_out.device(gpu_device) = gpu_in.argmax(dim);
assert(cudaMemcpyAsync(tensor_arg.data(), d_out, out_bytes, cudaMemcpyDeviceToHost, gpu_device.stream()) == cudaSuccess);
assert(cudaStreamSynchronize(gpu_device.stream()) == cudaSuccess);
for (DenseIndex n = 0; n < tensor_arg.size(); ++n) {
// Expect max to be in the last index of the reduced dimension
VERIFY_IS_EQUAL(tensor_arg.data()[n], tensor.dimension(dim) - 1);
}
cudaFree(d_in);
cudaFree(d_out);
}
}
template <int DataLayout>
void test_cuda_argmin_dim()
{
Tensor<float, 4, DataLayout> tensor(2,3,5,7);
std::vector<int> dims;
dims.push_back(2); dims.push_back(3); dims.push_back(5); dims.push_back(7);
for (int dim = 0; dim < 4; ++dim) {
tensor.setRandom();
tensor = (tensor + tensor.constant(0.5)).log();
array<DenseIndex, 3> out_shape;
for (int d = 0; d < 3; ++d) out_shape[d] = (d < dim) ? dims[d] : dims[d+1];
Tensor<DenseIndex, 3, DataLayout> tensor_arg(out_shape);
array<DenseIndex, 4> ix;
for (int i = 0; i < 2; ++i) {
for (int j = 0; j < 3; ++j) {
for (int k = 0; k < 5; ++k) {
for (int l = 0; l < 7; ++l) {
ix[0] = i; ix[1] = j; ix[2] = k; ix[3] = l;
if (ix[dim] != 0) continue;
// suppose dim == 1, then for all i, k, l, set tensor(i, 0, k, l) = 10.0
tensor(ix) = -10.0;
}
}
}
}
std::size_t in_bytes = tensor.size() * sizeof(float);
std::size_t out_bytes = tensor_arg.size() * sizeof(DenseIndex);
float* d_in;
DenseIndex* d_out;
cudaMalloc((void**)(&d_in), in_bytes);
cudaMalloc((void**)(&d_out), out_bytes);
cudaMemcpy(d_in, tensor.data(), in_bytes, cudaMemcpyHostToDevice);
Eigen::CudaStreamDevice stream;
Eigen::GpuDevice gpu_device(&stream);
Eigen::TensorMap<Eigen::Tensor<float, 4, DataLayout>, Aligned > gpu_in(d_in, Eigen::array<DenseIndex, 4>(2, 3, 5, 7));
Eigen::TensorMap<Eigen::Tensor<DenseIndex, 3, DataLayout>, Aligned > gpu_out(d_out, out_shape);
gpu_out.device(gpu_device) = gpu_in.argmin(dim);
assert(cudaMemcpyAsync(tensor_arg.data(), d_out, out_bytes, cudaMemcpyDeviceToHost, gpu_device.stream()) == cudaSuccess);
assert(cudaStreamSynchronize(gpu_device.stream()) == cudaSuccess);
VERIFY_IS_EQUAL(tensor_arg.size(),
2*3*5*7 / tensor.dimension(dim));
for (DenseIndex n = 0; n < tensor_arg.size(); ++n) {
// Expect min to be in the first index of the reduced dimension
VERIFY_IS_EQUAL(tensor_arg.data()[n], 0);
}
for (int i = 0; i < 2; ++i) {
for (int j = 0; j < 3; ++j) {
for (int k = 0; k < 5; ++k) {
for (int l = 0; l < 7; ++l) {
ix[0] = i; ix[1] = j; ix[2] = k; ix[3] = l;
if (ix[dim] != tensor.dimension(dim) - 1) continue;
// suppose dim == 1, then for all i, k, l, set tensor(i, 2, k, l) = 20.0
tensor(ix) = -20.0;
}
}
}
}
cudaMemcpy(d_in, tensor.data(), in_bytes, cudaMemcpyHostToDevice);
gpu_out.device(gpu_device) = gpu_in.argmin(dim);
assert(cudaMemcpyAsync(tensor_arg.data(), d_out, out_bytes, cudaMemcpyDeviceToHost, gpu_device.stream()) == cudaSuccess);
assert(cudaStreamSynchronize(gpu_device.stream()) == cudaSuccess);
for (DenseIndex n = 0; n < tensor_arg.size(); ++n) {
// Expect max to be in the last index of the reduced dimension
VERIFY_IS_EQUAL(tensor_arg.data()[n], tensor.dimension(dim) - 1);
}
cudaFree(d_in);
cudaFree(d_out);
}
}
void test_cxx11_tensor_cuda()
{
CALL_SUBTEST_1(test_cuda_simple_argmax<RowMajor>());
CALL_SUBTEST_1(test_cuda_simple_argmax<ColMajor>());
CALL_SUBTEST_2(test_cuda_argmax_dim<RowMajor>());
CALL_SUBTEST_2(test_cuda_argmax_dim<ColMajor>());
CALL_SUBTEST_3(test_cuda_argmin_dim<RowMajor>());
CALL_SUBTEST_3(test_cuda_argmin_dim<ColMajor>());
}
|
