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
机器未来
Paddle
提交
47ebe435
P
Paddle
项目概览
机器未来
/
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
未验证
提交
47ebe435
编写于
2月 01, 2018
作者:
D
dzhwinter
提交者:
GitHub
2月 01, 2018
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Fix/vector (#8045)
* "clean code" * "clean code"
上级
9b83462a
变更
1
隐藏空白更改
内联
并排
Showing
1 changed file
with
29 addition
and
48 deletion
+29
-48
paddle/framework/mixed_vector.h
paddle/framework/mixed_vector.h
+29
-48
未找到文件。
paddle/framework/mixed_vector.h
浏览文件 @
47ebe435
...
...
@@ -34,18 +34,6 @@ namespace framework {
template
<
typename
T
>
class
Vector
:
public
std
::
vector
<
T
>
{
public:
/* NOTE(dzhwinter):
* Data always store and modified on Host.
* If the data is modified when use cuda_data interface,
* You need to call the CopyFromCUDA explicitly to synchronize data.
*
*/
enum
class
kDataPosition
{
kDataOnHost
=
0
,
kDataOnDevice
=
1
,
};
public:
using
std
::
vector
<
T
>::
vector
;
...
...
@@ -55,11 +43,12 @@ class Vector : public std::vector<T> {
virtual
~
Vector
()
{
#ifdef PADDLE_WITH_CUDA
if
(
cuda_ptr_
!=
nullptr
)
{
memory
::
Free
<
platform
::
CUDAPlace
>
(
place_
,
static_cast
<
void
*>
(
cuda_ptr_
)
);
memory
::
Free
<
platform
::
CUDAPlace
>
(
place_
,
cuda_ptr_
);
}
#endif
}
/* Get device vector */
T
*
cuda_data
()
{
CopyToCUDA
();
PADDLE_ENFORCE_NOT_NULL
(
...
...
@@ -67,81 +56,73 @@ class Vector : public std::vector<T> {
return
static_cast
<
T
*>
(
cuda_ptr_
);
}
/* Get host vector */
T
*
data
()
{
return
std
::
vector
<
T
>::
data
();
}
const
T
*
data
()
const
{
return
std
::
vector
<
T
>::
data
();
}
/* Synchronize host vector to device vector */
void
CopyToCUDA
();
/* Synchronize device vector to host vector */
void
CopyFromCUDA
();
/* Switch device vector location */
void
CopyToPeer
(
platform
::
Place
);
private:
void
*
cuda_ptr_
=
nullptr
;
size_t
cuda_size_
=
0
;
/*The DataPosition is unused now,
if we want support random access from cpu and cuda,
we need to overload all the vector method */
kDataPosition
position_
=
kDataPosition
::
kDataOnHost
;
size_t
cuda_size_
=
0
;
// device vector numel
platform
::
CUDAPlace
place_
;
};
template
<
typename
T
>
void
Vector
<
T
>::
CopyToCUDA
()
{
#ifdef PADDLE_WITH_CUDA
if
(
cuda_ptr_
==
nullptr
)
{
if
(
cuda_size_
<
this
->
size
())
{
if
(
cuda_ptr_
!=
nullptr
)
{
memory
::
Free
<
platform
::
CUDAPlace
>
(
place_
,
cuda_ptr_
);
}
cuda_ptr_
=
memory
::
Alloc
<
platform
::
CUDAPlace
>
(
place_
,
this
->
size
()
*
sizeof
(
T
));
}
cuda_size_
=
this
->
size
();
platform
::
DeviceContextPool
&
pool
=
platform
::
DeviceContextPool
::
Instance
();
auto
*
cuda_ctx
=
pool
.
GetByPlace
(
place_
);
memory
::
Copy
(
place_
,
static_cast
<
void
*>
(
cuda_ptr_
),
platform
::
CPUPlace
(),
auto
*
ctx
=
pool
.
GetByPlace
(
place_
);
memory
::
Copy
(
place_
,
cuda_ptr_
,
platform
::
CPUPlace
(),
static_cast
<
const
void
*>
(
this
->
data
()),
this
->
size
()
*
sizeof
(
T
),
cuda_ctx
->
stream
());
cuda_ctx
->
Wait
();
cuda_size_
=
this
->
size
();
this
->
size
()
*
sizeof
(
T
),
ctx
->
stream
());
ctx
->
Wait
();
#endif
}
template
<
typename
T
>
void
Vector
<
T
>::
CopyFromCUDA
()
{
#ifdef PADDLE_WITH_CUDA
platform
::
DeviceContextPool
&
pool
=
platform
::
DeviceContextPool
::
Instance
();
auto
*
cuda_ctx
=
pool
.
GetByPlace
(
place_
);
if
(
cuda_ptr_
==
nullptr
)
{
LOG
(
WARNING
)
<<
"No uncommited cuda data."
;
LOG
(
WARNING
)
<<
"No uncommit
t
ed cuda data."
;
return
;
}
this
->
resize
(
cuda_size_
);
platform
::
DeviceContextPool
&
pool
=
platform
::
DeviceContextPool
::
Instance
();
auto
*
ctx
=
pool
.
GetByPlace
(
place_
);
memory
::
Copy
(
platform
::
CPUPlace
(),
static_cast
<
void
*>
(
this
->
data
()),
place_
,
static_cast
<
const
void
*>
(
cuda_ptr_
),
this
->
size
()
*
sizeof
(
T
),
cuda_ctx
->
stream
());
cuda_ctx
->
Wait
();
ctx
->
stream
());
ctx
->
Wait
();
#endif
}
template
<
typename
T
>
void
Vector
<
T
>::
CopyToPeer
(
platform
::
Place
peer_place
)
{
if
(
platform
::
is_cpu_place
(
peer_place
))
{
return
;
}
#ifdef PADDLE_WITH_CUDA
auto
*
c
uda_c
tx
=
platform
::
DeviceContextPool
::
Instance
().
GetByPlace
(
place_
);
void
*
peer_cuda_ptr
_
=
memory
::
Alloc
<
platform
::
CUDAPlace
>
(
auto
*
ctx
=
platform
::
DeviceContextPool
::
Instance
().
GetByPlace
(
place_
);
void
*
peer_cuda_ptr
=
memory
::
Alloc
<
platform
::
CUDAPlace
>
(
boost
::
get
<
platform
::
CUDAPlace
>
(
peer_place
),
this
->
size
()
*
sizeof
(
T
));
memory
::
Copy
(
boost
::
get
<
platform
::
CUDAPlace
>
(
peer_place
),
static_cast
<
void
*>
(
peer_cuda_ptr_
),
place_
,
static_cast
<
const
void
*>
(
cuda_ptr_
),
this
->
size
()
*
sizeof
(
T
),
cuda_ctx
->
stream
());
cuda_ctx
->
Wait
();
memory
::
Free
<
platform
::
CUDAPlace
>
(
place_
,
static_cast
<
void
*>
(
cuda_ptr_
));
memory
::
Copy
(
boost
::
get
<
platform
::
CUDAPlace
>
(
peer_place
),
peer_cuda_ptr
,
place_
,
cuda_ptr_
,
this
->
size
()
*
sizeof
(
T
),
ctx
->
stream
());
ctx
->
Wait
();
memory
::
Free
<
platform
::
CUDAPlace
>
(
place_
,
cuda_ptr_
);
place_
=
boost
::
get
<
platform
::
CUDAPlace
>
(
peer_place
);
cuda_ptr_
=
peer_cuda_ptr
_
;
cuda_ptr_
=
peer_cuda_ptr
;
#endif
}
...
...
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