Skip to content
体验新版
项目
组织
正在加载...
登录
切换导航
打开侧边栏
兽拳
Paddle
提交
311b8f2f
P
Paddle
项目概览
兽拳
/
Paddle
与 Fork 源项目一致
Fork自
PaddlePaddle / Paddle
通知
1
Star
1
Fork
0
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
1
列表
看板
标记
里程碑
合并请求
0
Wiki
0
Wiki
分析
仓库
DevOps
项目成员
Pages
P
Paddle
项目概览
项目概览
详情
发布
仓库
仓库
文件
提交
分支
标签
贡献者
分支图
比较
Issue
1
Issue
1
列表
看板
标记
里程碑
合并请求
0
合并请求
0
Pages
分析
分析
仓库分析
DevOps
Wiki
0
Wiki
成员
成员
收起侧边栏
关闭侧边栏
动态
分支图
创建新Issue
提交
Issue看板
体验新版 GitCode,发现更多精彩内容 >>
提交
311b8f2f
编写于
9月 30, 2018
作者:
Y
Yu Yang
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Refine Allocator facade
上级
2f16f47e
变更
5
隐藏空白更改
内联
并排
Showing
5 changed file
with
169 addition
and
21 deletion
+169
-21
paddle/fluid/memory/allocation/CMakeLists.txt
paddle/fluid/memory/allocation/CMakeLists.txt
+2
-1
paddle/fluid/memory/allocation/allocator_facade.cc
paddle/fluid/memory/allocation/allocator_facade.cc
+46
-20
paddle/fluid/memory/allocation/allocator_facade.h
paddle/fluid/memory/allocation/allocator_facade.h
+3
-0
paddle/fluid/memory/allocation/auto_increment_allocator.cc
paddle/fluid/memory/allocation/auto_increment_allocator.cc
+39
-0
paddle/fluid/memory/allocation/auto_increment_allocator.h
paddle/fluid/memory/allocation/auto_increment_allocator.h
+79
-0
未找到文件。
paddle/fluid/memory/allocation/CMakeLists.txt
浏览文件 @
311b8f2f
...
@@ -33,7 +33,7 @@ else ()
...
@@ -33,7 +33,7 @@ else ()
endif
()
endif
()
cc_library
(
aligned_allocator SRCS aligned_allocator.cc DEPS allocator
)
cc_library
(
aligned_allocator SRCS aligned_allocator.cc DEPS allocator
)
cc_library
(
auto_increment_allocator SRCS auto_increment_allocator.cc DEPS allocator
)
cc_library
(
allocator_facade SRCS allocator_facade.cc DEPS
cc_library
(
allocator_facade SRCS allocator_facade.cc DEPS
${
AllocatorFacadeDeps
}
${
AllocatorFacadeDeps
}
cpu_allocator
cpu_allocator
...
@@ -41,6 +41,7 @@ cc_library(allocator_facade SRCS allocator_facade.cc DEPS
...
@@ -41,6 +41,7 @@ cc_library(allocator_facade SRCS allocator_facade.cc DEPS
best_fit_allocator
best_fit_allocator
naive_managed_allocator
naive_managed_allocator
aligned_allocator
aligned_allocator
auto_increment_allocator
cuda_device_guard
)
cuda_device_guard
)
nv_test
(
allocation_and_eigen_test SRCS allocation_and_eigen_test.cu DEPS allocator_facade
)
nv_test
(
allocation_and_eigen_test SRCS allocation_and_eigen_test.cu DEPS allocator_facade
)
paddle/fluid/memory/allocation/allocator_facade.cc
浏览文件 @
311b8f2f
...
@@ -17,6 +17,7 @@
...
@@ -17,6 +17,7 @@
#include <vector>
#include <vector>
#include "paddle/fluid/memory/allocation/aligned_allocator.h"
#include "paddle/fluid/memory/allocation/aligned_allocator.h"
#include "paddle/fluid/memory/allocation/allocator_facade.h"
#include "paddle/fluid/memory/allocation/allocator_facade.h"
#include "paddle/fluid/memory/allocation/auto_increment_allocator.h"
#include "paddle/fluid/memory/allocation/best_fit_allocator.h"
#include "paddle/fluid/memory/allocation/best_fit_allocator.h"
#include "paddle/fluid/memory/allocation/cpu_allocator.h"
#include "paddle/fluid/memory/allocation/cpu_allocator.h"
#include "paddle/fluid/memory/allocation/locked_allocator.h"
#include "paddle/fluid/memory/allocation/locked_allocator.h"
...
@@ -33,6 +34,7 @@ namespace paddle {
...
@@ -33,6 +34,7 @@ namespace paddle {
namespace
memory
{
namespace
memory
{
namespace
allocation
{
namespace
allocation
{
// TODO(yy): Dirty code here. This class should be configurable in runtime.
class
CPUManagedAllocator
:
public
ManagedAllocator
{
class
CPUManagedAllocator
:
public
ManagedAllocator
{
public:
public:
CPUManagedAllocator
()
CPUManagedAllocator
()
...
@@ -56,24 +58,59 @@ class CPUManagedAllocator : public ManagedAllocator {
...
@@ -56,24 +58,59 @@ class CPUManagedAllocator : public ManagedAllocator {
return
normal_allocator_
->
AllocateShared
(
size
,
attr
);
return
normal_allocator_
->
AllocateShared
(
size
,
attr
);
}
}
}
}
bool
IsAllocThreadSafe
()
const
override
{
return
true
;
}
private:
private:
std
::
shared_ptr
<
ManagedAllocator
>
normal_allocator_
;
std
::
shared_ptr
<
ManagedAllocator
>
normal_allocator_
;
std
::
shared_ptr
<
ManagedAllocator
>
communication_allocator_
;
std
::
shared_ptr
<
ManagedAllocator
>
communication_allocator_
;
};
};
class
AllocatorFacadePrivate
{
// TODO(yy): Dirty code here. This class should be configurable in runtime.
class
CUDAManagedAllocator
:
public
ManagedAllocator
{
public:
public:
std
::
map
<
platform
::
Place
,
std
::
shared_ptr
<
ManagedAllocator
>>
allocators_
;
explicit
CUDAManagedAllocator
(
int
dev_id
)
{
std
::
vector
<
std
::
unique_ptr
<
Allocation
>>
pre_allocations_
;
platform
::
CUDADeviceGuard
guard
(
dev_id
);
std
::
vector
<
std
::
shared_ptr
<
Allocator
>>
holding_allocators_
;
max_chunk_size_
=
platform
::
GpuMaxChunkSize
();
raw_allocator_
=
NaiveManagedAllocator
::
Create
(
std
::
unique_ptr
<
Allocator
>
(
new
CUDAAllocator
(
platform
::
CUDAPlace
(
dev_id
))));
default_allocator_
=
std
::
make_shared
<
AutoIncrementAllocator
>
(
[
this
]
{
return
std
::
move
(
BestFitAllocatorCreator
());
});
}
~
AllocatorFacadePrivate
()
{
~
CUDAManagedAllocator
()
{
// Specify destruct order.
// Specify destruct order.
pre_allocations_
.
clear
();
default_allocator_
.
reset
();
allocators_
.
clear
();
chunks_
.
clear
();
holding_allocators_
.
clear
();
raw_allocator_
.
reset
();
}
std
::
unique_ptr
<
Allocation
>
Allocate
(
size_t
size
,
Attr
attr
)
override
{
return
default_allocator_
->
Allocate
(
size
,
attr
);
}
std
::
shared_ptr
<
Allocation
>
AllocateShared
(
size_t
size
,
Attr
attr
)
override
{
return
default_allocator_
->
AllocateShared
(
size
,
attr
);
}
std
::
shared_ptr
<
ManagedAllocator
>
BestFitAllocatorCreator
()
{
chunks_
.
emplace_back
(
raw_allocator_
->
Allocate
(
max_chunk_size_
));
auto
*
allocation
=
chunks_
.
back
().
get
();
return
NaiveManagedAllocator
::
Create
(
std
::
unique_ptr
<
Allocator
>
(
new
BestFitAllocator
(
allocation
)));
}
}
bool
IsAllocThreadSafe
()
const
override
{
return
true
;
}
private:
size_t
max_chunk_size_
;
std
::
vector
<
std
::
unique_ptr
<
Allocation
>>
chunks_
;
std
::
shared_ptr
<
ManagedAllocator
>
raw_allocator_
;
std
::
shared_ptr
<
ManagedAllocator
>
default_allocator_
;
};
class
AllocatorFacadePrivate
{
public:
std
::
map
<
platform
::
Place
,
std
::
shared_ptr
<
ManagedAllocator
>>
allocators_
;
~
AllocatorFacadePrivate
()
{}
AllocatorFacadePrivate
()
{
AllocatorFacadePrivate
()
{
InitCPUAllocator
();
InitCPUAllocator
();
...
@@ -88,19 +125,8 @@ class AllocatorFacadePrivate {
...
@@ -88,19 +125,8 @@ class AllocatorFacadePrivate {
void
InitCUDAAllocator
()
{
void
InitCUDAAllocator
()
{
#ifdef PADDLE_WITH_CUDA
#ifdef PADDLE_WITH_CUDA
for
(
int
dev_id
=
0
;
dev_id
<
platform
::
GetCUDADeviceCount
();
++
dev_id
)
{
for
(
int
dev_id
=
0
;
dev_id
<
platform
::
GetCUDADeviceCount
();
++
dev_id
)
{
platform
::
CUDADeviceGuard
guard
(
dev_id
);
auto
cuda_allocator
=
NaiveManagedAllocator
::
Create
(
std
::
unique_ptr
<
Allocator
>
(
new
CUDAAllocator
(
platform
::
CUDAPlace
(
dev_id
))));
auto
allocation
=
cuda_allocator
->
Allocate
(
platform
::
GpuMaxChunkSize
());
auto
allocator
=
NaiveManagedAllocator
::
Create
(
std
::
unique_ptr
<
Allocator
>
(
new
LockedAllocator
(
std
::
unique_ptr
<
Allocator
>
(
new
BestFitAllocator
(
allocation
.
get
())))));
pre_allocations_
.
emplace_back
(
std
::
move
(
allocation
));
holding_allocators_
.
emplace_back
(
cuda_allocator
);
allocators_
[
platform
::
CUDAPlace
(
dev_id
)]
=
allocators_
[
platform
::
CUDAPlace
(
dev_id
)]
=
std
::
make_shared
<
AlignedAllocator
<
64
>>
(
std
::
move
(
allocator
)
);
std
::
make_shared
<
CUDAManagedAllocator
>
(
dev_id
);
}
}
#endif
#endif
}
}
...
...
paddle/fluid/memory/allocation/allocator_facade.h
浏览文件 @
311b8f2f
...
@@ -21,6 +21,9 @@ namespace paddle {
...
@@ -21,6 +21,9 @@ namespace paddle {
namespace
memory
{
namespace
memory
{
namespace
allocation
{
namespace
allocation
{
// Allocator Facade is the interface exposed to other modules.
// All the configuration or dirty code under development should
// be hidden behind this facade.
class
AllocatorFacadePrivate
;
class
AllocatorFacadePrivate
;
class
AllocatorFacade
{
class
AllocatorFacade
{
public:
public:
...
...
paddle/fluid/memory/allocation/auto_increment_allocator.cc
0 → 100644
浏览文件 @
311b8f2f
// 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/auto_increment_allocator.h"
namespace
paddle
{
namespace
memory
{
namespace
allocation
{
std
::
unique_ptr
<
Allocation
>
AutoIncrementAllocator
::
Allocate
(
size_t
size
,
Allocator
::
Attr
attr
)
{
return
InvokeOrCreateUnderlyingAllocator
([
&
](
ManagedAllocator
&
allocator
)
{
return
allocator
.
Allocate
(
size
,
attr
);
});
}
std
::
shared_ptr
<
Allocation
>
AutoIncrementAllocator
::
AllocateShared
(
size_t
size
,
Allocator
::
Attr
attr
)
{
return
InvokeOrCreateUnderlyingAllocator
([
&
](
ManagedAllocator
&
allocator
)
{
return
allocator
.
AllocateShared
(
size
,
attr
);
});
}
bool
AutoIncrementAllocator
::
IsAllocThreadSafe
()
const
{
return
true
;
}
}
// namespace allocation
}
// namespace memory
}
// namespace paddle
paddle/fluid/memory/allocation/auto_increment_allocator.h
0 → 100644
浏览文件 @
311b8f2f
// 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.
#pragma once
#include <functional>
#include <memory>
#include <thread> // NOLINT
#include <vector>
#include "paddle/fluid/memory/allocation/allocator.h"
namespace
paddle
{
namespace
memory
{
namespace
allocation
{
class
AutoIncrementAllocator
:
public
ManagedAllocator
{
public:
using
AllocatorCreator
=
std
::
function
<
std
::
shared_ptr
<
ManagedAllocator
>
()
>
;
template
<
typename
Creator
>
explicit
AutoIncrementAllocator
(
Creator
&&
creator
)
:
creator_
(
std
::
move
(
creator
)),
prev_success_allocator_
{
0
}
{}
std
::
unique_ptr
<
Allocation
>
Allocate
(
size_t
size
,
Attr
attr
)
override
;
std
::
shared_ptr
<
Allocation
>
AllocateShared
(
size_t
size
,
Attr
attr
)
override
;
bool
IsAllocThreadSafe
()
const
override
;
private:
// NOTE: here use template Callback, it can be inlined when -O3
template
<
typename
Callback
>
inline
typename
std
::
result_of
<
Callback
(
ManagedAllocator
&
)
>::
type
InvokeOrCreateUnderlyingAllocator
(
Callback
callback
)
{
size_t
retry_count
=
underlying_allocators_
.
size
();
auto
cur
=
prev_success_allocator_
;
while
(
retry_count
--
>
0
)
{
// until there retry count is zero
try
{
auto
res
=
callback
(
*
underlying_allocators_
[
cur
]);
{
std
::
lock_guard
<
std
::
mutex
>
guard
(
mtx_
);
prev_success_allocator_
=
cur
;
}
return
std
::
move
(
res
);
}
catch
(
BadAlloc
&
)
{
++
cur
;
if
(
cur
>=
underlying_allocators_
.
size
())
{
cur
=
0
;
}
}
catch
(...)
{
// if there is another type of allocation, just rethrow it.
throw
;
}
}
// No suitable allocator
{
std
::
lock_guard
<
std
::
mutex
>
guard
(
mtx_
);
underlying_allocators_
.
emplace_back
(
creator_
());
prev_success_allocator_
=
underlying_allocators_
.
size
()
-
1
;
return
callback
(
*
underlying_allocators_
[
prev_success_allocator_
]);
}
}
AllocatorCreator
creator_
;
std
::
vector
<
AllocatorCreator
::
result_type
>
underlying_allocators_
;
size_t
prev_success_allocator_
{
0
};
std
::
mutex
mtx_
;
// NOLINT
};
}
// namespace allocation
}
// namespace memory
}
// namespace paddle
编辑
预览
Markdown
is supported
0%
请重试
或
添加新附件
.
添加附件
取消
You are about to add
0
people
to the discussion. Proceed with caution.
先完成此消息的编辑!
取消
想要评论请
注册
或
登录