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
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// 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 "paddle/fluid/memory/allocation/allocator.h"
#include <gflags/gflags.h>
#include <map>
#include <string>
#include <unordered_map>
#include <vector>
#include "paddle/fluid/memory/allocation/aligned_allocator.h"
#include "paddle/fluid/memory/allocation/allocator_facade.h"
#include "paddle/fluid/memory/allocation/allocator_strategy.h"
#include "paddle/fluid/memory/allocation/auto_increment_allocator.h"
#include "paddle/fluid/memory/allocation/best_fit_allocator.h"
#include "paddle/fluid/memory/allocation/conditional_allocator.h"
#include "paddle/fluid/memory/allocation/cpu_allocator.h"
#include "paddle/fluid/memory/allocation/legacy_allocator.h"
#include "paddle/fluid/memory/allocation/locked_allocator.h"
#include "paddle/fluid/memory/allocation/retry_allocator.h"
#include "paddle/fluid/memory/allocation/zero_size_allocator.h"
#include "paddle/fluid/platform/cpu_info.h"
#include "paddle/fluid/platform/place.h"
#ifdef PADDLE_WITH_CUDA
#include "paddle/fluid/memory/allocation/cuda_allocator.h"
#include "paddle/fluid/memory/allocation/pinned_allocator.h"
#include "paddle/fluid/platform/cuda_device_guard.h"
#include "paddle/fluid/platform/gpu_info.h"
#endif
DEFINE_int64(
gpu_allocator_retry_time, 0,
"The retry time (milliseconds) when allocator fails "
"to allocate memory. No retry if this value is not greater than 0");
namespace paddle {
namespace memory {
namespace allocation {
// TODO(yy): Dirty code here. This class should be configurable in runtime.
class CPUManagedAllocator : public Allocator {
public:
CPUManagedAllocator() : normal_allocator_(new CPUAllocator()) {}
bool IsAllocThreadSafe() const override { return true; }
protected:
Allocation* AllocateImpl(size_t size, Allocator::Attr attr) override {
return normal_allocator_->Allocate(size, attr).release();
}
private:
std::shared_ptr<Allocator> normal_allocator_;
};
// TODO(yy): Dirty code here. This class should be configurable in runtime.
class ChunkedAllocator : public Allocator {
public:
explicit ChunkedAllocator(std::unique_ptr<Allocator> system_allocator,
size_t max_chunk_size, size_t capacity = 1,
int64_t retry_time = -1)
: max_chunk_size_(max_chunk_size), retry_time_(retry_time) {
raw_allocator_ = std::move(system_allocator);
if (max_chunk_size_ == 0) {
default_allocator_ = raw_allocator_;
} else {
if (capacity == 1) {
VLOG(10) << "Create BestFitAllocator with chunk_size "
<< max_chunk_size_;
default_allocator_ = CreateAllocatorWithChunk();
} else {
VLOG(10) << "Create AutoIncrementAllocator with chunk_size "
<< max_chunk_size_ << " and capacity " << capacity;
default_allocator_ = std::make_shared<AutoIncrementAllocator>(
[this] { return std::move(CreateAllocatorWithChunk()); }, capacity);
}
}
auto* cond_allocator = new ConditionalAllocator();
cond_allocator
->AddAllocator(
[this](size_t size, Attr attr) { return size < max_chunk_size_; },
default_allocator_)
.AddAllocator(
[](size_t size, Attr attr) {
return true; // default case
},
raw_allocator_);
default_allocator_.reset(cond_allocator);
}
~ChunkedAllocator() override {
// Specify destruct order.
default_allocator_.reset();
chunks_.clear();
raw_allocator_.reset();
}
std::shared_ptr<Allocator> CreateAllocatorWithChunk() {
chunks_.emplace_back(raw_allocator_->Allocate(max_chunk_size_));
auto* allocation = chunks_.back().get();
std::unique_ptr<Allocator> allocator(new LockedAllocator(
std::unique_ptr<Allocator>(new BestFitAllocator(allocation))));
if (retry_time_ > 0) {
auto* retry_allocator =
new RetryAllocator(std::move(allocator), retry_time_);
allocator.reset(retry_allocator);
}
return std::make_shared<AlignedAllocator<64u>>(std::move(allocator));
}
bool IsAllocThreadSafe() const override { return true; }
protected:
Allocation* AllocateImpl(size_t size, Allocator::Attr attr) override {
return default_allocator_->Allocate(size, attr).release();
}
protected:
size_t max_chunk_size_;
int64_t retry_time_;
std::vector<AllocationPtr> chunks_;
std::shared_ptr<Allocator> raw_allocator_;
std::shared_ptr<Allocator> default_allocator_;
};
#ifdef PADDLE_WITH_CUDA
class CUDAChunkedAllocator : public ChunkedAllocator {
public:
explicit CUDAChunkedAllocator(int dev_id)
: ChunkedAllocator(std::unique_ptr<Allocator>(
new CUDAAllocator(platform::CUDAPlace(dev_id))),
GetMaxChunkSize(dev_id), GetCapcity(dev_id),
GetRetryTime()) {}
private:
static size_t GetMaxChunkSize(int dev_id) {
platform::CUDADeviceGuard guard(dev_id);
return platform::GpuMaxChunkSize();
}
static size_t GetCapcity(int dev_id) {
platform::CUDADeviceGuard guard(dev_id);
size_t available, total;
platform::GpuMemoryUsage(&available, &total);
size_t max_chunk_size = platform::GpuMaxChunkSize();
return max_chunk_size == 0 ? 0 : available / max_chunk_size;
}
static int64_t GetRetryTime() { return FLAGS_gpu_allocator_retry_time; }
};
class CUDAPinnedChunkedAllocator : public ChunkedAllocator {
public:
CUDAPinnedChunkedAllocator()
: ChunkedAllocator(std::unique_ptr<Allocator>(new CPUPinnedAllocator()),
platform::CUDAPinnedMaxChunkSize(), GetCapacity(),
-1) {} // never retry
private:
static size_t GetCapacity() {
size_t total = platform::CpuTotalPhysicalMemory();
size_t max_chunk_size = platform::CUDAPinnedMaxChunkSize();
return max_chunk_size == 0 ? 0 : total / max_chunk_size;
}
};
#endif
class AllocatorFacadePrivate {
public:
std::map<platform::Place, std::shared_ptr<Allocator>> allocators_;
~AllocatorFacadePrivate() = default;
AllocatorFacadePrivate() {
if (GetAllocatorStrategy() == AllocatorStrategy::kLegacy) {
InitLegacyAllocator();
} else {
InitCPUAllocator();
InitCUDAAllocator();
InitCUDAPinnedAllocator();
WrapZeroSizeAllocator();
}
}
private:
void InitLegacyAllocator() {
std::vector<platform::Place> places{platform::CPUPlace()};
#ifdef PADDLE_WITH_CUDA
for (int dev_id = 0; dev_id < platform::GetCUDADeviceCount(); ++dev_id) {
places.emplace_back(platform::CUDAPlace(dev_id));
}
places.emplace_back(platform::CUDAPinnedPlace());
#endif
for (auto& p : places) {
allocators_[p] = std::make_shared<LegacyAllocator>(p);
}
}
void InitCPUAllocator() {
allocators_[platform::CPUPlace()] = std::make_shared<CPUManagedAllocator>();
}
void InitCUDAAllocator() {
#ifdef PADDLE_WITH_CUDA
int device_count = platform::GetCUDADeviceCount();
for (int dev_id = 0; dev_id < device_count; ++dev_id) {
allocators_[platform::CUDAPlace(dev_id)] =
std::make_shared<CUDAChunkedAllocator>(dev_id);
}
#endif
}
void InitCUDAPinnedAllocator() {
#ifdef PADDLE_WITH_CUDA
allocators_[platform::CUDAPinnedPlace()] =
std::make_shared<CUDAPinnedChunkedAllocator>();
#endif
}
void WrapZeroSizeAllocator() {
for (auto& pair : allocators_) {
pair.second =
std::make_shared<ZeroSizeAllocator>(pair.second, pair.first);
}
}
};
// Pimpl. Make interface clean.
AllocatorFacade::AllocatorFacade() : m_(new AllocatorFacadePrivate()) {}
AllocatorFacade::~AllocatorFacade() { delete m_; }
AllocatorFacade& AllocatorFacade::Instance() {
static AllocatorFacade instance;
return instance;
}
std::shared_ptr<Allocation> AllocatorFacade::AllocShared(
const platform::Place& place, size_t size, Allocator::Attr attr) {
return std::shared_ptr<Allocation>(Alloc(place, size, attr).release(),
AllocationDeleter());
}
AllocationPtr AllocatorFacade::Alloc(const platform::Place& place, size_t size,
Allocator::Attr attr) {
auto it = m_->allocators_.find(place);
if (it == m_->allocators_.end()) {
throw BadAlloc(
string::Sprintf("No such allocator for the place, %s", place));
}
return m_->allocators_.at(place)->Allocate(size, attr);
}
} // namespace allocation
} // namespace memory
} // namespace paddle