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
Paddle
提交
f879ef23
P
Paddle
项目概览
PaddlePaddle
/
Paddle
12 个月 前同步成功
通知
2292
Star
20931
Fork
5422
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
1423
列表
看板
标记
里程碑
合并请求
543
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
1,423
Issue
1,423
列表
看板
标记
里程碑
合并请求
543
合并请求
543
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
f879ef23
编写于
12月 22, 2017
作者:
Y
Yang Yang
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
pass forward backward runtime
上级
f899150e
变更
4
隐藏空白更改
内联
并排
Showing
4 changed file
with
124 addition
and
29 deletion
+124
-29
paddle/framework/lod_tensor.cc
paddle/framework/lod_tensor.cc
+39
-0
paddle/framework/lod_tensor.h
paddle/framework/lod_tensor.h
+3
-0
paddle/operators/elementwise_op.h
paddle/operators/elementwise_op.h
+5
-0
paddle/operators/parallel_do_op.cc
paddle/operators/parallel_do_op.cc
+77
-29
未找到文件。
paddle/framework/lod_tensor.cc
浏览文件 @
f879ef23
...
...
@@ -314,6 +314,45 @@ void DeserializeFromStream(std::istream &is, LoDTensor *tensor) {
}
}
std
::
vector
<
LoDTensor
>
LoDTensor
::
SplitLoDTensor
(
const
std
::
vector
<
platform
::
Place
>
places
)
const
{
check_memory_size
();
// PADDLE_ENFORCE(lod().empty() || (lod().size() == 1 && lod()[0].empty())
// , "Disable parallel lod for now");
PADDLE_ENFORCE
(
lod
().
empty
(),
"Disable parallel lod for now"
);
PADDLE_ENFORCE
(
dims
()[
0
]
%
places
.
size
()
==
0
,
"Batch size should be divided by places size"
);
std
::
vector
<
LoDTensor
>
lods
;
for
(
int
place_idx
=
0
;
place_idx
<
places
.
size
();
++
place_idx
)
{
int
begin
=
place_idx
*
dims
()[
0
]
/
places
.
size
();
int
end
=
(
place_idx
+
1
)
*
dims
()[
0
]
/
places
.
size
();
auto
src
=
Slice
(
begin
,
end
);
LoDTensor
dst
;
dst
.
Resize
(
src
.
dims
());
auto
&
dst_place
=
places
[
place_idx
];
auto
dst_ptr
=
dst
.
mutable_data
(
dst_place
,
src
.
type
());
// TODO(tonyyang-svail):
// change the following to framework::CopyFrom
auto
src_place
=
src
.
place
();
auto
src_ptr
=
src
.
data
<
void
>
();
auto
size
=
src
.
numel
()
*
SizeOfType
(
src
.
type
());
if
(
platform
::
is_cpu_place
(
src_place
)
&&
platform
::
is_cpu_place
(
dst_place
))
{
memory
::
Copy
(
boost
::
get
<
platform
::
CPUPlace
>
(
dst_place
),
dst_ptr
,
boost
::
get
<
platform
::
CPUPlace
>
(
src_place
),
src_ptr
,
size
);
}
else
{
PADDLE_THROW
(
"Not Implemented"
);
}
lods
.
emplace_back
(
dst
);
}
return
lods
;
}
void
LoDTensor
::
MergeLoDTensor
(
const
std
::
vector
<
const
LoDTensor
*>
&
lod_tensors
,
platform
::
Place
place
)
{
PADDLE_ENFORCE
(
platform
::
is_cpu_place
(
place
));
...
...
paddle/framework/lod_tensor.h
浏览文件 @
f879ef23
...
...
@@ -144,6 +144,9 @@ class LoDTensor : public Tensor {
*/
void
ShrinkInLevel
(
size_t
level
,
size_t
elem_begin
,
size_t
elem_end
);
std
::
vector
<
LoDTensor
>
SplitLoDTensor
(
const
std
::
vector
<
platform
::
Place
>
places
)
const
;
void
MergeLoDTensor
(
const
std
::
vector
<
const
LoDTensor
*>&
lod_tensors
,
platform
::
Place
place
);
...
...
paddle/operators/elementwise_op.h
浏览文件 @
f879ef23
...
...
@@ -34,6 +34,8 @@ class ElementwiseOp : public framework::OperatorWithKernel {
auto
x_dim
=
ctx
->
GetInputDim
(
"X"
);
auto
y_dim
=
ctx
->
GetInputDim
(
"Y"
);
LOG
(
INFO
)
<<
x_dim
;
LOG
(
INFO
)
<<
y_dim
;
PADDLE_ENFORCE_GE
(
x_dim
.
size
(),
y_dim
.
size
(),
"Rank of first input must >= rank of second input."
);
ctx
->
SetOutputDim
(
"Out"
,
x_dim
);
...
...
@@ -118,6 +120,9 @@ class ElementwiseOpGrad : public framework::OperatorWithKernel {
auto
x_dims
=
ctx
->
GetInputDim
(
"X"
);
auto
y_dims
=
ctx
->
GetInputDim
(
"Y"
);
auto
out_dims
=
ctx
->
GetInputDim
(
framework
::
GradVarName
(
"Out"
));
LOG
(
INFO
)
<<
x_dims
;
LOG
(
INFO
)
<<
y_dims
;
LOG
(
INFO
)
<<
out_dims
;
PADDLE_ENFORCE_GE
(
x_dims
.
size
(),
y_dims
.
size
(),
"Rank of first input must >= rank of second input."
);
...
...
paddle/operators/parallel_do_op.cc
浏览文件 @
f879ef23
...
...
@@ -13,11 +13,9 @@
limitations under the License. */
#include <vector>
#include "chunk_eval_op.h"
#include "paddle/framework/executor.h"
#include "paddle/framework/op_registry.h"
#include "paddle/framework/operator.h"
#include "paddle/platform/place.h"
namespace
paddle
{
namespace
operators
{
...
...
@@ -31,10 +29,31 @@ constexpr char kParallelScopes[] = "parallel_scopes";
constexpr
char
kParallelBlock
[]
=
"sub_block"
;
using
ParallelScopeVar
=
std
::
vector
<
framework
::
Scope
*>
;
// using ParallelScopeVar = std::vector<framework::Scope *>;
using
LoDTensor
=
framework
::
LoDTensor
;
using
OperatorBase
=
framework
::
OperatorBase
;
class
ParallelDoOp
:
public
OperatorBase
{
void
SplitTensorAndMoveTensorToScopes
(
const
framework
::
Scope
&
scope
,
const
std
::
vector
<
framework
::
Scope
*>
&
sub_scopes
,
const
std
::
vector
<
platform
::
Place
>
&
places
,
const
std
::
vector
<
std
::
string
>
&
names
)
{
for
(
auto
&
argu
:
names
)
{
auto
*
var
=
scope
.
FindVar
(
argu
);
const
auto
&
tensor
=
var
->
Get
<
LoDTensor
>
();
auto
lod_tensors
=
tensor
.
SplitLoDTensor
(
places
);
for
(
auto
&
lod
:
lod_tensors
)
{
LOG
(
INFO
)
<<
lod
.
dims
();
}
for
(
int
i
=
0
;
i
<
sub_scopes
.
size
();
++
i
)
{
*
sub_scopes
[
i
]
->
Var
(
argu
)
->
GetMutable
<
LoDTensor
>
()
=
lod_tensors
[
i
];
}
}
}
class
ParallelDoOp
:
public
framework
::
OperatorBase
{
public:
ParallelDoOp
(
const
std
::
string
&
type
,
const
framework
::
VariableNameMap
&
inputs
,
...
...
@@ -52,11 +71,18 @@ class ParallelDoOp : public OperatorBase {
places
.
emplace_back
(
platform
::
CPUPlace
());
places
.
emplace_back
(
platform
::
CPUPlace
());
std
::
vector
<
framework
::
Scope
*>
sub_scopes
;
auto
&
sub_scopes
=
*
scope
.
FindVar
(
Output
(
kParallelScopes
))
->
GetMutable
<
std
::
vector
<
framework
::
Scope
*>>
();
// std::vector<framework::Scope *> sub_scopes;
for
(
int
place_idx
=
0
;
place_idx
<
places
.
size
();
++
place_idx
)
{
VLOG
(
3
)
<<
"Run "
<<
place_idx
;
sub_scopes
.
push_back
(
&
scope
.
NewScope
());
}
SplitTensorAndMoveTensorToScopes
(
scope
,
sub_scopes
,
places
,
Inputs
(
kInputs
));
for
(
int
place_idx
=
0
;
place_idx
<
places
.
size
();
++
place_idx
)
{
VLOG
(
3
)
<<
"Run "
<<
place_idx
;
auto
&
place
=
places
[
place_idx
];
auto
*
cur_scope
=
sub_scopes
[
place_idx
];
...
...
@@ -66,26 +92,6 @@ class ParallelDoOp : public OperatorBase {
PADDLE_THROW
(
"Not Implemented"
);
}
// feed input
for
(
auto
&
argu
:
Inputs
(
kInputs
))
{
auto
*
var
=
scope
.
FindVar
(
argu
);
const
auto
&
tensor
=
var
->
Get
<
LoDTensor
>
();
if
(
!
tensor
.
lod
().
empty
())
{
PADDLE_THROW
(
"Disable parallel lod for now"
);
}
else
{
PADDLE_ENFORCE
(
tensor
.
dims
()[
0
]
%
places
.
size
()
==
0
,
"Batch size should be divided by places size"
);
int
begin
=
place_idx
*
tensor
.
dims
()[
0
]
/
places
.
size
();
int
end
=
(
place_idx
+
1
)
*
tensor
.
dims
()[
0
]
/
places
.
size
();
auto
feed_tensor
=
tensor
.
Slice
(
begin
,
end
);
feed_tensor
.
switch_place
(
place
);
auto
*
cur_var
=
cur_scope
->
Var
(
argu
);
auto
*
cur_tensor
=
cur_var
->
GetMutable
<
Tensor
>
();
*
cur_tensor
=
feed_tensor
;
}
}
// execute
auto
executor
=
framework
::
Executor
(
place
);
executor
.
Run
(
*
program
,
cur_scope
,
block
->
ID
(),
...
...
@@ -132,7 +138,49 @@ class ParallelDoGradOp : public OperatorBase {
:
OperatorBase
(
type
,
inputs
,
outputs
,
attrs
)
{}
void
Run
(
const
framework
::
Scope
&
scope
,
const
platform
::
DeviceContext
&
dev_ctx
)
const
override
{}
const
platform
::
DeviceContext
&
dev_ctx
)
const
override
{
auto
*
block
=
Attr
<
framework
::
BlockDescBind
*>
(
kParallelBlock
);
auto
*
program
=
block
->
Program
();
auto
&
sub_scopes
=
scope
.
FindVar
(
Input
(
kParallelScopes
))
->
Get
<
std
::
vector
<
framework
::
Scope
*>>
();
// TODO(tonyyang-svail): get places from input
std
::
vector
<
platform
::
Place
>
places
;
places
.
emplace_back
(
platform
::
CPUPlace
());
places
.
emplace_back
(
platform
::
CPUPlace
());
// feed output@grad
SplitTensorAndMoveTensorToScopes
(
scope
,
sub_scopes
,
places
,
Inputs
(
framework
::
GradVarName
(
kOutputs
)));
for
(
auto
&
s
:
Inputs
(
framework
::
GradVarName
(
kOutputs
)))
{
LOG
(
INFO
)
<<
s
;
LOG
(
INFO
)
<<
scope
.
FindVar
(
s
)
->
Get
<
LoDTensor
>
().
dims
();
for
(
auto
*
sub_scope
:
sub_scopes
)
{
LOG
(
INFO
)
<<
sub_scope
->
FindVar
(
s
)
->
Get
<
LoDTensor
>
().
dims
();
}
}
// exe run
for
(
int
place_idx
=
0
;
place_idx
<
places
.
size
();
++
place_idx
)
{
VLOG
(
3
)
<<
"Run "
<<
place_idx
;
auto
&
place
=
places
[
place_idx
];
auto
*
cur_scope
=
sub_scopes
[
place_idx
];
// copy parameter
if
(
dev_ctx
.
GetPlace
()
!=
place
)
{
PADDLE_THROW
(
"Not Implemented"
);
}
// execute
auto
executor
=
framework
::
Executor
(
place
);
executor
.
Run
(
*
program
,
cur_scope
,
block
->
ID
(),
false
/*create_local_scope*/
);
}
// merge grad
}
};
class
ParallelDoGradOpDescMaker
:
public
framework
::
SingleGradOpDescMaker
{
...
...
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