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
PaddlePaddle
PaddleDetection
提交
e02859c0
P
PaddleDetection
项目概览
PaddlePaddle
/
PaddleDetection
12 个月 前同步成功
通知
692
Star
11112
Fork
2696
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
184
列表
看板
标记
里程碑
合并请求
40
Wiki
0
Wiki
分析
仓库
DevOps
项目成员
Pages
P
PaddleDetection
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
184
Issue
184
列表
看板
标记
里程碑
合并请求
40
合并请求
40
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
e02859c0
编写于
6月 26, 2017
作者:
Y
Yi Wang
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Replace {cpu,gpu}_allocator.h and {cpu,gpu}_allocator_test.cc by system_allocator{.h,_test.cc}
上级
ce70df86
变更
8
隐藏空白更改
内联
并排
Showing
8 changed file
with
106 addition
and
186 deletion
+106
-186
paddle/memory/CMakeLists.txt
paddle/memory/CMakeLists.txt
+6
-0
paddle/memory/detail/CMakeLists.txt
paddle/memory/detail/CMakeLists.txt
+1
-2
paddle/memory/detail/cpu_allocator.h
paddle/memory/detail/cpu_allocator.h
+0
-71
paddle/memory/detail/cpu_allocator_test.cc
paddle/memory/detail/cpu_allocator_test.cc
+0
-30
paddle/memory/detail/system_allocator.h
paddle/memory/detail/system_allocator.h
+47
-32
paddle/memory/detail/system_allocator_test.cc
paddle/memory/detail/system_allocator_test.cc
+19
-1
paddle/memory/memory.cc
paddle/memory/memory.cc
+30
-37
paddle/memory/memory.h
paddle/memory/memory.h
+3
-13
未找到文件。
paddle/memory/CMakeLists.txt
浏览文件 @
e02859c0
add_subdirectory
(
detail
)
add_subdirectory
(
detail
)
if
(
${
WITH_GPU
}
)
nv_library
(
memory SRCS memory.cc
)
else
(
${
WITH_GPU
}
)
cc_library
(
memory SRCS memroy.cc
)
endif
(
${
WITH_GPU
}
)
paddle/memory/detail/CMakeLists.txt
浏览文件 @
e02859c0
cc_test
(
cpu_allocator_test SRCS cpu_allocator_test.cc
)
cc_test
(
system_allocator_test SRCS system_allocator_test.cc
)
nv_test
(
gpu_allocator_test SRCS gpu_allocator_test.cc
)
paddle/memory/detail/cpu_allocator.h
已删除
100644 → 0
浏览文件 @
ce70df86
/* 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. */
#pragma once
#include <stddef.h> // for size_t
#include <cstdlib> // for malloc and free
#ifndef _WIN32
#include <sys/mman.h> // for mlock and munlock
#endif
namespace
paddle
{
namespace
memory
{
namespace
detail
{
// CPUAllocator<staging=true> calls mlock, which returns
// pinned and locked memory as staging areas for data exchange
// between host and device. Allocates too much would reduce the
// amount of memory available to the system for paging. So, by
// default, we should use CPUAllocator<staging=false>.
template
<
bool
staging
>
class
CPUAllocator
{
public:
void
*
Alloc
(
size_t
size
);
void
Free
(
void
*
p
,
size_t
size
);
};
template
<
>
class
CPUAllocator
<
false
>
{
public:
void
*
Alloc
(
size_t
size
)
{
return
std
::
malloc
(
size
);
}
void
Free
(
void
*
p
,
size_t
size
)
{
std
::
free
(
p
);
}
};
template
<
>
class
CPUAllocator
<
true
>
{
public:
void
*
Alloc
(
size_t
size
)
{
void
*
p
=
std
::
malloc
(
size
);
if
(
p
==
nullptr
)
{
return
p
;
}
#ifndef _WIN32
mlock
(
p
,
size
);
#endif
return
p
;
}
void
Free
(
void
*
p
,
size_t
size
)
{
#ifndef _WIN32
munlock
(
p
,
size
);
#endif
std
::
free
(
p
);
}
};
}
// namespace detail
}
// namespace memory
}
// namespace paddle
paddle/memory/detail/cpu_allocator_test.cc
已删除
100644 → 0
浏览文件 @
ce70df86
/* 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 "paddle/memory/detail/cpu_allocator.h"
#include "gtest/gtest.h"
TEST
(
CPUAllocator
,
NonStaging
)
{
paddle
::
memory
::
detail
::
CPUAllocator
<
false
>
a
;
void
*
p
=
a
.
Alloc
(
4096
);
EXPECT_NE
(
p
,
nullptr
);
a
.
Free
(
p
,
4096
);
}
TEST
(
CPUAllocator
,
Staging
)
{
paddle
::
memory
::
detail
::
CPUAllocator
<
true
>
a
;
void
*
p
=
a
.
Alloc
(
4096
);
EXPECT_NE
(
p
,
nullptr
);
a
.
Free
(
p
,
4096
);
}
paddle/memory/detail/
gpu
_allocator.h
→
paddle/memory/detail/
system
_allocator.h
浏览文件 @
e02859c0
...
@@ -14,20 +14,58 @@ limitations under the License. */
...
@@ -14,20 +14,58 @@ limitations under the License. */
#pragma once
#pragma once
#include <stddef.h> // for size_t
#include <stddef.h> // for size_t
#include <sys/mman.h> // for mlock and munlock
#include <cstdlib> // for malloc and free
#i
nclude <thrust/system_error.h>
#i
fndef PADDLE_ONLY_CPU
#include <thrust/system/cuda/error.h>
#include <thrust/system/cuda/error.h>
#include <thrust/system_error.h>
#endif // PADDLE_ONLY_CPU
namespace
paddle
{
namespace
paddle
{
namespace
memory
{
namespace
memory
{
namespace
detail
{
namespace
detail
{
class
SystemAllocator
{
public:
virtual
void
*
Alloc
(
size_t
size
)
=
0
;
virtual
void
*
Free
(
void
*
p
)
=
0
;
};
// CPUAllocator<lock_memory=true> calls mlock, which returns pinned
// and locked memory as staging areas for data exchange between host
// and device. Allocates too much would reduce the amount of memory
// available to the system for paging. So, by default, we should use
// CPUAllocator<staging=false>.
template
<
bool
lock_memory
>
class
CPUAllocator
:
public
SystemAllocator
{
public:
virtual
void
*
Alloc
(
size_t
size
)
{
void
*
p
=
std
::
malloc
(
size
);
if
(
p
!=
nullptr
&&
lock_memory
)
{
mlock
(
p
,
size
);
}
return
p
;
}
virtual
void
Free
(
void
*
p
,
size_t
size
)
{
if
(
p
!=
nullptr
&&
lock_memory
)
{
munlock
(
p
,
size
);
}
std
::
free
(
p
);
}
};
#ifndef PADDLE_ONLY_CPU // The following code are for CUDA.
namespace
{
inline
void
throw_on_error
(
cudaError_t
e
,
const
char
*
message
)
{
inline
void
throw_on_error
(
cudaError_t
e
,
const
char
*
message
)
{
if
(
e
)
{
if
(
e
)
{
throw
thrust
::
system_error
(
e
,
thrust
::
cuda_category
(),
message
);
throw
thrust
::
system_error
(
e
,
thrust
::
cuda_category
(),
message
);
}
}
}
}
}
// namespace
// GPUAllocator<staging=true> calls cudaHostMalloc, which returns
// GPUAllocator<staging=true> calls cudaHostMalloc, which returns
// pinned and locked memory as staging areas for data exchange
// pinned and locked memory as staging areas for data exchange
...
@@ -36,17 +74,11 @@ inline void throw_on_error(cudaError_t e, const char* message) {
...
@@ -36,17 +74,11 @@ inline void throw_on_error(cudaError_t e, const char* message) {
// default, we should use GPUAllocator<staging=false>.
// default, we should use GPUAllocator<staging=false>.
template
<
bool
staging
>
template
<
bool
staging
>
class
GPUAllocator
{
class
GPUAllocator
{
public:
public:
void
*
Alloc
(
size_t
size
);
void
Free
(
void
*
p
,
size_t
size
);
};
template
<
>
class
GPUAllocator
<
false
>
{
public:
void
*
Alloc
(
size_t
size
)
{
void
*
Alloc
(
size_t
size
)
{
void
*
p
=
0
;
void
*
p
=
0
;
cudaError_t
result
=
cudaMalloc
(
&
p
,
size
);
cudaError_t
result
=
staging
?
cudaMallocHost
(
&
p
,
size
)
:
cudaMalloc
(
&
p
,
size
);
if
(
result
==
cudaSuccess
)
{
if
(
result
==
cudaSuccess
)
{
return
p
;
return
p
;
}
}
...
@@ -60,32 +92,15 @@ public:
...
@@ -60,32 +92,15 @@ public:
// that is returned if you ever call cudaFree after the
// that is returned if you ever call cudaFree after the
// driver has already shutdown. This happens only if the
// driver has already shutdown. This happens only if the
// process is terminating, in which case we don't care if
// process is terminating, in which case we don't care if
// cudaFree succeeds.
// cudaFree succeeds.
auto
err
=
cudaFree
(
p
);
auto
err
=
staging
?
cudaFreeHost
(
p
)
:
cudaFree
(
p
);
if
(
err
!=
cudaErrorCudartUnloading
)
{
if
(
err
!=
cudaErrorCudartUnloading
)
{
throw_on_error
(
err
,
"cudaFree failed"
);
throw_on_error
(
err
,
"cudaFree failed"
);
}
}
}
}
};
};
template
<
>
#endif // PADDLE_ONLY_CPU
class
GPUAllocator
<
true
>
{
public:
void
*
Alloc
(
size_t
size
)
{
void
*
p
=
0
;
cudaError_t
result
=
cudaMallocHost
(
&
p
,
size
);
if
(
result
==
cudaSuccess
)
{
return
p
;
}
// clear last error
cudaGetLastError
();
return
nullptr
;
}
void
Free
(
void
*
p
,
size_t
size
)
{
throw_on_error
(
cudaFreeHost
(
p
),
"cudaFreeHost failed"
);
}
};
}
// namespace detail
}
// namespace detail
}
// namespace memory
}
// namespace memory
...
...
paddle/memory/detail/
gpu
_allocator_test.cc
→
paddle/memory/detail/
system
_allocator_test.cc
浏览文件 @
e02859c0
...
@@ -12,9 +12,25 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
...
@@ -12,9 +12,25 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
See the License for the specific language governing permissions and
limitations under the License. */
limitations under the License. */
#include "paddle/memory/detail/
gpu
_allocator.h"
#include "paddle/memory/detail/
system
_allocator.h"
#include "gtest/gtest.h"
#include "gtest/gtest.h"
TEST
(
CPUAllocator
,
NoLockMem
)
{
paddle
::
memory
::
detail
::
CPUAllocator
<
false
>
a
;
void
*
p
=
a
.
Alloc
(
4096
);
EXPECT_NE
(
p
,
nullptr
);
a
.
Free
(
p
,
4096
);
}
TEST
(
CPUAllocator
,
LockMem
)
{
paddle
::
memory
::
detail
::
CPUAllocator
<
true
>
a
;
void
*
p
=
a
.
Alloc
(
4096
);
EXPECT_NE
(
p
,
nullptr
);
a
.
Free
(
p
,
4096
);
}
#ifndef PADDLE_ONLY_CPU
TEST
(
GPUAllocator
,
NonStaging
)
{
TEST
(
GPUAllocator
,
NonStaging
)
{
paddle
::
memory
::
detail
::
GPUAllocator
<
false
>
a
;
paddle
::
memory
::
detail
::
GPUAllocator
<
false
>
a
;
void
*
p
=
a
.
Alloc
(
4096
);
void
*
p
=
a
.
Alloc
(
4096
);
...
@@ -28,3 +44,5 @@ TEST(GPUAllocator, Staging) {
...
@@ -28,3 +44,5 @@ TEST(GPUAllocator, Staging) {
EXPECT_NE
(
p
,
nullptr
);
EXPECT_NE
(
p
,
nullptr
);
a
.
Free
(
p
,
4096
);
a
.
Free
(
p
,
4096
);
}
}
#endif // PADDLE_ONLY_CPU
paddle/memory/memory.cc
浏览文件 @
e02859c0
...
@@ -14,48 +14,41 @@ limitations under the License. */
...
@@ -14,48 +14,41 @@ limitations under the License. */
#include "paddle/memory/memory.h"
#include "paddle/memory/memory.h"
#include "paddle/memory/detail/cpu_allocator.h"
#include "paddle/memory/detail/gpu_allocator.h"
namespace
paddle
{
namespace
paddle
{
namespace
memory
{
namespace
memory
{
template
<
>
void
Alloc
(
paddle
::
platform
::
Place
pl
,
size_t
size
)
{
void
*
Alloc
<
CPUPlace
>
(
CPUPlace
,
size_t
size
)
{
#ifndef PADDLE_ONLY_CPU
return
GetCPUBuddyAllocator
(
false
/*non-staging*/
)
->
Alloc
(
size
);
if
(
paddle
::
platform
::
is_gpu_place
(
pl
))
{
}
return
GetGPUBuddyAllocator
(
pl
.
device
)
->
Alloc
(
size
);
}
void
*
AllocStaging
(
CPUPlace
,
size_t
size
)
{
#endif // PADDLE_ONLY_CPU
return
GetCPUBuddyAllocator
(
true
/*staging*/
)
->
Alloc
(
size
);
PADDLE_ASSERT
(
paddle
::
platform
::
is_cpu_place
(
pl
));
}
return
GetCPUBuddyAllocator
()
->
Alloc
(
size
);
}
template
<
>
void
*
Alloc
<
GPUPlace
>
(
GPUPlace
pl
,
size_t
size
)
{
void
Free
(
paddle
::
platform
::
Place
pl
,
void
*
p
)
{
return
GetGPUBuddyAllocator
(
pl
.
device
)
->
Alloc
(
size
);
#ifndef PADDLE_ONLY_CPU
}
if
(
paddle
::
platform
::
is_gpu_place
(
pl
))
{
GetGPUBuddyAllocator
(
pl
.
device
)
->
Free
(
p
);
template
<
>
}
void
Free
<
CPUPlace
>
(
CPUPlace
,
void
*
p
)
{
#endif // PADDLE_ONLY_CPU
return
GetCPUBuddyAllocator
(
false
/*non-staging*/
)
->
Free
(
p
);
PADDLE_ASSERT
(
paddle
::
platform
::
is_cpu_place
(
pl
));
}
GetCPUBuddyAllocator
()
->
Free
(
p
);
}
void
FreeStaging
(
CPUPlace
,
void
*
p
)
{
return
GetCPUBuddyAllocator
(
false
/*non-staging*/
)
->
Free
(
p
);
size_t
Used
(
paddle
::
platform
::
Place
pl
)
{
}
#ifndef PADDLE_ONLY_CPU
if
(
paddle
::
platform
::
is_gpu_place
(
pl
))
{
#ifdef PADDLE_WITH_GPU
return
GetGPUBuddyAllocator
(
pl
.
device
)
->
Used
();
template
<
>
}
void
*
Alloc
<
GPUPlace
>
(
GPUPlace
pl
,
void
*
p
)
{
#endif // PADDLE_ONLY_CPU
return
GetGPUBuddyAllocator
(
pl
.
device
)
->
Free
(
p
);
PADDLE_ASSERT
(
paddle
::
platform
::
is_cpu_place
(
pl
));
}
template
<
>
size_t
Used
<
CPUPlace
>
(
CPUPlace
)
{
return
GetCPUBuddyAllocator
()
->
Used
();
return
GetCPUBuddyAllocator
()
->
Used
();
}
}
template
<
>
size_t
Alloc
<
GPUPlace
>
(
GPUPlace
pl
)
{
return
GetGPUBuddyAllocator
(
pl
.
device
)
->
Used
();
}
#endif // PADDLE_WITH_GPU
}
// namespace memory
}
// namespace memory
}
// namespace paddle
}
// namespace paddle
paddle/memory/memory.h
浏览文件 @
e02859c0
...
@@ -19,19 +19,9 @@ limitations under the License. */
...
@@ -19,19 +19,9 @@ limitations under the License. */
namespace
paddle
{
namespace
paddle
{
namespace
memory
{
namespace
memory
{
template
<
typename
paddle
::
framework
::
Place
>
void
*
Alloc
(
paddle
::
framework
::
Place
,
size_t
);
void
*
Alloc
(
Place
,
size_t
);
void
Free
(
paddle
::
framework
::
Place
,
void
*
);
template
<
typename
paddle
::
framework
::
Place
>
size_t
Used
(
paddle
::
framework
::
Place
);
void
Free
(
Place
,
void
*
);
template
<
typename
paddle
::
framework
::
Place
>
size_t
Used
(
Place
);
// Staging memory means "pinned" host memory that can be mapped into
// the CUDA memory space and accessed by the device rapidly. Don't
// allocate too much staging memory; otherwise system performance will
// degrade because the OS cannot find enough swap memory space.
void
*
AllocStaging
(
CPUPlace
,
size_t
);
void
*
FreeStaging
(
CPUPlace
,
size_t
);
}
// namespace memory
}
// namespace memory
}
// namespace paddle
}
// 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