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
|
// Copyright (C) 2024 Advanced Micro Devices, Inc. All rights reserved.
//
// Permission is hereby granted, free of charge, to any person obtaining a copy
// of this software and associated documentation files (the "Software"), to deal
// in the Software without restriction, including without limitation the rights
// to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
// copies of the Software, and to permit persons to whom the Software is
// furnished to do so, subject to the following conditions:
//
// The above copyright notice and this permission notice shall be included in
// all copies or substantial portions of the Software.
//
// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
// THE SOFTWARE.
#include "../../shared/accuracy_test.h"
#include "../../shared/params_gen.h"
#include "../hipfft_params.h"
#include <algorithm>
#include <gtest/gtest.h>
#include <hip/hip_runtime_api.h>
#include <optional>
extern fft_params::fft_mp_lib mp_lib;
extern int mp_ranks;
static const std::vector<std::vector<size_t>> multi_gpu_sizes = {
{128, 256},
{64, 128, 256},
};
enum SplitType
{
// split both input and output on slow FFT dimension
SLOW_INOUT,
// split only input on slow FFT dimension, output is not split
SLOW_IN,
// split only output on slow FFT dimension, input is not split
SLOW_OUT,
// split input on slow FFT dimension, and output on fast FFT dimension
SLOW_IN_FAST_OUT,
// 3D pencil decomposition - one dimension is contiguous on input
// and another dimension contiguous on output, remaining dims are
// both split
PENCIL_3D,
};
std::vector<fft_params> param_generator_multi_gpu(const std::optional<SplitType> type)
{
int localDeviceCount = 0;
(void)hipGetDeviceCount(&localDeviceCount);
// if we have an explicit split of data on the user side, we need
// to use the multiprocessing API
if(type)
{
if(mp_lib == fft_params::fft_mp_lib_none)
return {};
}
// data is not explicitly split up, that means the library is
// asked to do the split. We need multiple GPUs to do this.
else if(localDeviceCount < 2)
return {};
static const std::vector<std::vector<size_t>> stride_range = {{1}};
auto params_complex = param_generator_complex(test_prob,
multi_gpu_sizes,
precision_range_sp_dp,
{1, 10},
stride_generator(stride_range),
stride_generator(stride_range),
{{0, 0}},
{{0, 0}},
{fft_placement_inplace, fft_placement_notinplace},
false);
auto params_real = param_generator_real(test_prob,
multi_gpu_sizes,
precision_range_sp_dp,
{1, 10},
stride_generator(stride_range),
stride_generator(stride_range),
{{0, 0}},
{{0, 0}},
{fft_placement_notinplace},
false);
std::vector<fft_params> all_params;
auto distribute_params = [=, &all_params](const std::vector<fft_params>& params) {
for(auto& p : params)
{
// test library splitting
if(!type)
{
auto param_multi = p;
// for single-batch, cuFFT only allows in-place
if(p.nbatch == 1 && p.placement == fft_placement_notinplace)
continue;
param_multi.multiGPU = std::min(static_cast<int>(p.nbatch), localDeviceCount);
all_params.emplace_back(std::move(param_multi));
}
else
{
// the API only allows for batch-1 multi-process FFTs
if(p.nbatch > 1)
continue;
// user-specified split
int brickCount = mp_ranks;
// start with all-ones in grids
std::vector<unsigned int> input_grid(p.length.size() + 1, 1);
std::vector<unsigned int> output_grid(p.length.size() + 1, 1);
auto p_dist = p;
switch(*type)
{
case SLOW_INOUT:
input_grid[1] = brickCount;
output_grid[1] = brickCount;
break;
case SLOW_IN:
// this type only specifies input field and no output
// field, but multi-process transforms require both
// fields.
if(mp_lib != fft_params::fft_mp_lib_none)
continue;
input_grid[1] = brickCount;
break;
case SLOW_OUT:
// this type only specifies output field and no input
// field, but multi-process transforms require both
// fields.
if(mp_lib != fft_params::fft_mp_lib_none)
continue;
output_grid[1] = brickCount;
break;
case SLOW_IN_FAST_OUT:
// requires at least rank-2 FFT
if(p.length.size() < 2)
continue;
input_grid[1] = brickCount;
output_grid.back() = brickCount;
break;
case PENCIL_3D:
// need at least 2 bricks per split dimension, or 4 devices.
// also needs to be a 3D problem.
if(brickCount < 4 || p.length.size() != 3)
continue;
// make fast dimension contiguous on input
input_grid[1] = static_cast<unsigned int>(sqrt(brickCount));
input_grid[2] = brickCount / input_grid[1];
// make middle dimension contiguous on output
output_grid[1] = input_grid[1];
output_grid[3] = input_grid[2];
break;
}
p_dist.mp_lib = mp_lib;
p_dist.distribute_input(localDeviceCount, input_grid);
p_dist.distribute_output(localDeviceCount, output_grid);
// "placement" flag is meaningless if exactly one of
// input+output is a field. So just add those cases if
// the flag is "out-of-place", since "in-place" is
// exactly the same test case.
if(p_dist.placement == fft_placement_inplace
&& p_dist.ifields.empty() != p_dist.ofields.empty())
continue;
// in-place transforms require identical input/output layouts
if(p.placement == fft_placement_inplace && input_grid != output_grid)
continue;
all_params.push_back(std::move(p_dist));
}
}
};
distribute_params(params_complex);
distribute_params(params_real);
return all_params;
}
// split both input and output on slowest FFT dim
INSTANTIATE_TEST_SUITE_P(multi_gpu_slowest_dim,
accuracy_test,
::testing::ValuesIn(param_generator_multi_gpu(SLOW_INOUT)),
accuracy_test::TestName);
// split slowest FFT dim only on input, or only on output
INSTANTIATE_TEST_SUITE_P(multi_gpu_slowest_input_dim,
accuracy_test,
::testing::ValuesIn(param_generator_multi_gpu(SLOW_IN)),
accuracy_test::TestName);
INSTANTIATE_TEST_SUITE_P(multi_gpu_slowest_output_dim,
accuracy_test,
::testing::ValuesIn(param_generator_multi_gpu(SLOW_OUT)),
accuracy_test::TestName);
// split input on slowest FFT and output on fastest, to minimize data
// movement (only makes sense for rank-2 and higher FFTs)
INSTANTIATE_TEST_SUITE_P(multi_gpu_slowin_fastout,
accuracy_test,
::testing::ValuesIn(param_generator_multi_gpu(SLOW_IN_FAST_OUT)),
accuracy_test::TestName);
// 3D pencil decompositions
INSTANTIATE_TEST_SUITE_P(multi_gpu_3d_pencils,
accuracy_test,
::testing::ValuesIn(param_generator_multi_gpu(PENCIL_3D)),
accuracy_test::TestName);
// library-decided splits
INSTANTIATE_TEST_SUITE_P(multi_gpu,
accuracy_test,
::testing::ValuesIn(param_generator_multi_gpu({})),
accuracy_test::TestName);
|
