forked from PaddlePaddle/Paddle
-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathParameterUpdater.cpp
More file actions
152 lines (128 loc) · 4.94 KB
/
ParameterUpdater.cpp
File metadata and controls
152 lines (128 loc) · 4.94 KB
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
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "ParameterUpdater.h"
#include "paddle/utils/Logging.h"
#include "paddle/utils/Thread.h"
namespace paddle {
static const hl_stream_t kDeviceToHostStream = HPPL_STREAM_1;
static const hl_stream_t kHostToDeviceStream = HPPL_STREAM_2;
SgdUpdaterWithCpuAverager::SgdUpdaterWithCpuAverager(
const OptimizationConfig& optConfig)
: SgdLocalUpdater(optConfig, false /*with averager*/) {
CHECK(FLAGS_use_gpu && optConfig.do_average_in_cpu());
averager_.reset(AverageOptimizer::create(optConfig,
new DummyOptimizer(optConfig),
false /*sparse*/,
true /*apply*/));
updateWorker_.addJob([]() { hl_set_device(FLAGS_gpu_id); });
}
void SgdUpdaterWithCpuAverager::init(
const std::vector<ParameterPtr>& parameters) {
SgdLocalUpdater::init(parameters);
averager_->init(parameters_.size(), nullptr);
copyEvents_.resize(parameters_.size());
for (auto& parameter : parameters) {
SetDevice device(parameter->getDeviceId());
cpuParameters_.emplace_back(new Parameter(parameter->getConfig(),
/* useGpu= */ false,
/* doInit= */ false));
if (parameter->useGpu()) {
cpuParameters_.back()->enableType(PARAMETER_APPLY);
} else {
cpuParameters_.back()->enableSharedType(
PARAMETER_APPLY, parameter->getBuf(PARAMETER_VALUE));
}
for (ParameterType type : averager_->getParameterTypes()) {
cpuParameters_.back()->enableType(type);
}
hl_create_event(©Events_[nonStaticParaIDMap_[parameter->getID()]]);
}
}
SgdUpdaterWithCpuAverager::~SgdUpdaterWithCpuAverager() {
for (auto& event : copyEvents_) {
hl_destroy_event(event);
}
}
void SgdUpdaterWithCpuAverager::updateImpl(Parameter* para) {
SgdLocalUpdater::updateImpl(para);
if (para->useGpu()) {
size_t pid = nonStaticParaIDMap_[para->getID()];
Parameter* cpuPara = cpuParameters_[pid].get();
cpuPara->getBuf(PARAMETER_VALUE)
->copyFrom(*para->getBuf(PARAMETER_VALUE), kDeviceToHostStream);
hl_stream_record_event(kDeviceToHostStream, copyEvents_[pid]);
}
updateWorker_.addJob(
std::bind(&SgdUpdaterWithCpuAverager::updateFunc, this, para));
}
void SgdUpdaterWithCpuAverager::updateFunc(Parameter* para) {
SetDevice setDevice(para->getDeviceId());
size_t pid = nonStaticParaIDMap_[para->getID()];
Parameter* cpuPara = cpuParameters_[pid].get();
if (para->useGpu()) {
hl_event_synchronize(copyEvents_[pid]);
}
averager_->update(cpuPara->getBufs(), cpuPara->getConfig(), -1LU);
}
void SgdUpdaterWithCpuAverager::finishBatch(real cost) {
SgdLocalUpdater::finishBatch(cost);
updateWorker_.wait();
for (auto para : cpuParameters_) {
if (auto callback = averager_->needSpecialTraversal(para->getConfig())) {
callback(para->getBufs(), para->getConfig(), -1LU);
}
}
averager_->finishBatch();
}
void SgdUpdaterWithCpuAverager::apply() {
// backup gpu value
for (auto& para : parameters_) {
SetDevice setDevice(para->getDeviceId());
para->getBuf(PARAMETER_GRADIENT)
->copyFrom(*para->getBuf(PARAMETER_VALUE), kHostToDeviceStream);
}
// apply on cpu parameter
if (auto callback = averager_->apply()) {
for (auto para : cpuParameters_) {
callback(para->getBufs(), para->getConfig(), -1LU);
}
}
// copy to gpu value
for (auto& para : parameters_) {
SetDevice setDevice(para->getDeviceId());
size_t pid = nonStaticParaIDMap_[para->getID()];
Parameter* cpuPara = cpuParameters_[pid].get();
if (parameters_[pid]->useGpu()) {
para->getBuf(PARAMETER_VALUE)
->copyFrom(*cpuPara->getBuf(PARAMETER_APPLY), kHostToDeviceStream);
}
}
hl_stream_synchronize(kHostToDeviceStream);
for (auto& para : parameters_) {
para->setValueUpdated();
}
}
void SgdUpdaterWithCpuAverager::restore() {
// restore on cpu parameter
if (auto callback = averager_->restore()) {
for (auto para : cpuParameters_) {
callback(para->getBufs(), para->getConfig(), -1LU);
}
}
// restore gpu value
for (auto& para : parameters_) {
SetDevice device(para->getDeviceId());
para->getBuf(PARAMETER_VALUE)->copyFrom(*para->getBuf(PARAMETER_GRADIENT));
para->getBuf(PARAMETER_GRADIENT)->zeroMem();
para->setValueUpdated();
}
}
} // namespace paddle