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
马赛克医院
Paddle
提交
642fd68c
P
Paddle
项目概览
马赛克医院
/
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
0
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
0
Issue
0
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
642fd68c
编写于
2月 18, 2019
作者:
Y
Yancey1989
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
update by comment test=develop
上级
7cd6de37
变更
7
隐藏空白更改
内联
并排
Showing
7 changed file
with
35 addition
and
32 deletion
+35
-32
paddle/fluid/framework/details/all_reduce_deps_pass.h
paddle/fluid/framework/details/all_reduce_deps_pass.h
+0
-2
paddle/fluid/framework/details/memory_optimize_helper.h
paddle/fluid/framework/details/memory_optimize_helper.h
+0
-2
paddle/fluid/framework/details/multi_devices_graph_pass.cc
paddle/fluid/framework/details/multi_devices_graph_pass.cc
+0
-1
paddle/fluid/framework/details/parallel_ssa_graph_executor.cc
...le/fluid/framework/details/parallel_ssa_graph_executor.cc
+20
-8
paddle/fluid/framework/details/parallel_ssa_graph_executor.h
paddle/fluid/framework/details/parallel_ssa_graph_executor.h
+6
-5
paddle/fluid/framework/ir/graph.h
paddle/fluid/framework/ir/graph.h
+5
-0
paddle/fluid/framework/parallel_executor.cc
paddle/fluid/framework/parallel_executor.cc
+4
-14
未找到文件。
paddle/fluid/framework/details/all_reduce_deps_pass.h
浏览文件 @
642fd68c
...
...
@@ -21,8 +21,6 @@ namespace paddle {
namespace
framework
{
namespace
details
{
constexpr
char
kAllOpDescs
[]
=
"all_op_descs"
;
// TODO(gongwb): overlap allreduce with backward computation.
class
AllReduceDepsPass
:
public
ir
::
Pass
{
protected:
...
...
paddle/fluid/framework/details/memory_optimize_helper.h
浏览文件 @
642fd68c
...
...
@@ -29,8 +29,6 @@ namespace paddle {
namespace
framework
{
namespace
details
{
constexpr
char
kAllOpDescs
[]
=
"all_op_descs"
;
std
::
vector
<
ir
::
Node
*>
SortOpLikeDescOrder
(
const
ir
::
Graph
&
graph
);
// NOTE(dzh): A ordered set for node reuse in memory optimize.
...
...
paddle/fluid/framework/details/multi_devices_graph_pass.cc
浏览文件 @
642fd68c
...
...
@@ -221,7 +221,6 @@ std::unique_ptr<ir::Graph> MultiDevSSAGraphBuilderBase::ApplyImpl(
* Only variables should be the leaves of graph.
*/
AddOutputToLeafOps
(
&
result
);
// result.Erase(kGraphOps);
return
graph
;
}
...
...
paddle/fluid/framework/details/parallel_ssa_graph_executor.cc
浏览文件 @
642fd68c
...
...
@@ -19,12 +19,12 @@ namespace paddle {
namespace
framework
{
namespace
details
{
std
::
vector
<
std
::
unique_ptr
<
ir
::
Graph
>>
SeparateMultiDevicesGraph
(
const
std
::
vector
<
platform
::
Place
>
&
places
,
std
::
unique_ptr
<
ir
::
Graph
>
graph
)
{
std
::
vector
<
std
::
unique_ptr
<
ir
::
Graph
>>
ParallelSSAGraphExecutor
::
SeparateMultiDevicesGraph
(
std
::
unique_ptr
<
ir
::
Graph
>
&&
graph
)
{
std
::
vector
<
std
::
unique_ptr
<
ir
::
Graph
>>
graphs
;
graphs
.
reserve
(
places
.
size
());
for
(
size_t
i
=
0
;
i
<
places
.
size
();
++
i
)
{
graphs
.
reserve
(
places
_
.
size
());
for
(
size_t
i
=
0
;
i
<
places
_
.
size
();
++
i
)
{
ProgramDesc
empty
;
graphs
.
emplace_back
(
std
::
unique_ptr
<
ir
::
Graph
>
(
new
ir
::
Graph
(
empty
)));
auto
&
g
=
graphs
.
back
();
...
...
@@ -60,7 +60,7 @@ std::vector<std::unique_ptr<ir::Graph>> SeparateMultiDevicesGraph(
}
}
for
(
size_t
dev_id
=
0
;
dev_id
<
places
.
size
();
++
dev_id
)
{
for
(
size_t
dev_id
=
0
;
dev_id
<
places
_
.
size
();
++
dev_id
)
{
auto
&
dev_vars
=
graphs
[
dev_id
]
->
Get
<
GraphVars
>
(
kGraphVars
)[
0
];
auto
&
origin_vars
=
graph
->
Get
<
GraphVars
>
(
kGraphVars
)[
dev_id
];
for
(
auto
&
name_pair
:
origin_vars
)
{
...
...
@@ -80,14 +80,26 @@ std::vector<std::unique_ptr<ir::Graph>> SeparateMultiDevicesGraph(
ParallelSSAGraphExecutor
::
ParallelSSAGraphExecutor
(
const
ExecutionStrategy
&
strategy
,
const
std
::
vector
<
Scope
*>
&
local_scopes
,
const
std
::
vector
<
platform
::
Place
>
&
places
,
std
::
vector
<
std
::
unique_ptr
<
ir
::
Graph
>>
&&
graphs
)
const
framework
::
ProgramDesc
&
main_prog
,
std
::
unique_ptr
<
ir
::
Graph
>
&&
graph
)
:
strategy_
(
std
::
move
(
strategy
)),
local_scopes_
(
std
::
move
(
local_scopes
)),
pool_
(
places
.
size
()
>=
2
?
new
::
ThreadPool
(
places
.
size
())
:
nullptr
),
places_
(
std
::
move
(
places
)),
graphs_
(
std
::
move
(
graphs
))
{
main_prog_
(
main_prog
),
// TODO(Yancey1989): copy graphs is not safely since it deleted the attrs.
graphs_
(
SeparateMultiDevicesGraph
(
std
::
move
(
graph
)))
{
PADDLE_ENFORCE_EQ
(
places_
.
size
(),
local_scopes_
.
size
());
auto
seq_allreduce_pass
=
ir
::
PassRegistry
::
Instance
().
Get
(
"all_reduce_deps_pass"
);
seq_allreduce_pass
->
Erase
(
details
::
kAllOpDescs
);
seq_allreduce_pass
->
Set
<
const
std
::
vector
<
OpDesc
*>>
(
details
::
kAllOpDescs
,
new
std
::
vector
<
OpDesc
*>
(
main_prog_
.
Block
(
0
).
AllOps
()));
for
(
size_t
i
=
0
;
i
<
graphs_
.
size
();
++
i
)
{
graphs_
[
i
]
=
seq_allreduce_pass
->
Apply
(
std
::
move
(
graphs_
[
i
]));
}
// set the correct size of thread pool to each device.
strategy_
.
num_threads_
=
strategy_
.
num_threads_
<
places_
.
size
()
?
1UL
...
...
paddle/fluid/framework/details/parallel_ssa_graph_executor.h
浏览文件 @
642fd68c
...
...
@@ -28,16 +28,13 @@ namespace paddle {
namespace
framework
{
namespace
details
{
std
::
vector
<
std
::
unique_ptr
<
ir
::
Graph
>>
SeparateMultiDevicesGraph
(
const
std
::
vector
<
platform
::
Place
>
&
places
,
std
::
unique_ptr
<
ir
::
Graph
>
graph
);
class
ParallelSSAGraphExecutor
:
public
SSAGraphExecutor
{
public:
ParallelSSAGraphExecutor
(
const
ExecutionStrategy
&
strategy
,
const
std
::
vector
<
Scope
*>
&
local_scopes
,
const
std
::
vector
<
platform
::
Place
>
&
places
,
std
::
vector
<
std
::
unique_ptr
<
ir
::
Graph
>>
&&
graphs
);
const
framework
::
ProgramDesc
&
main_prog
,
std
::
unique_ptr
<
ir
::
Graph
>
&&
graph
);
~
ParallelSSAGraphExecutor
()
final
=
default
;
const
ir
::
Graph
&
Graph
()
const
override
{
return
*
graphs_
[
0
];
}
...
...
@@ -45,10 +42,14 @@ class ParallelSSAGraphExecutor : public SSAGraphExecutor {
FeedFetchList
Run
(
const
std
::
vector
<
std
::
string
>
&
fetch_tensors
)
override
;
private:
std
::
vector
<
std
::
unique_ptr
<
ir
::
Graph
>>
SeparateMultiDevicesGraph
(
std
::
unique_ptr
<
ir
::
Graph
>
&&
graph
);
ExecutionStrategy
strategy_
;
std
::
vector
<
Scope
*>
local_scopes_
;
std
::
unique_ptr
<::
ThreadPool
>
pool_
{
nullptr
};
std
::
vector
<
platform
::
Place
>
places_
;
framework
::
ProgramDesc
main_prog_
;
std
::
vector
<
std
::
unique_ptr
<
ir
::
Graph
>>
graphs_
;
std
::
vector
<
std
::
unique_ptr
<
details
::
ThreadedSSAGraphExecutor
>>
executors_
;
...
...
paddle/fluid/framework/ir/graph.h
浏览文件 @
642fd68c
...
...
@@ -26,6 +26,11 @@ limitations under the License. */
namespace
paddle
{
namespace
framework
{
namespace
details
{
constexpr
char
kAllOpDescs
[]
=
"all_op_descs"
;
}
// namespace details
namespace
ir
{
/*
...
...
paddle/fluid/framework/parallel_executor.cc
浏览文件 @
642fd68c
...
...
@@ -305,21 +305,11 @@ ParallelExecutor::ParallelExecutor(
if
(
build_strategy
.
enable_parallel_graph_
)
{
#ifdef PADDLE_WITH_CUDA
auto
parallel_graph
=
details
::
SeparateMultiDevicesGraph
(
member_
->
places_
,
std
::
move
(
graph
));
auto
seq_allreduce_pass
=
ir
::
PassRegistry
::
Instance
().
Get
(
"all_reduce_deps_pass"
);
seq_allreduce_pass
->
Erase
(
details
::
kAllOpDescs
);
seq_allreduce_pass
->
Set
<
const
std
::
vector
<
OpDesc
*>>
(
details
::
kAllOpDescs
,
new
std
::
vector
<
OpDesc
*>
(
main_program
.
Block
(
0
).
AllOps
()));
for
(
size_t
i
=
0
;
i
<
parallel_graph
.
size
();
++
i
)
{
parallel_graph
[
i
]
=
seq_allreduce_pass
->
Apply
(
std
::
move
(
parallel_graph
[
i
]));
}
// TODO(Yancey1989): Remove passing in the main_program when
// allreduce_seq_pass doesn't need it as the attr.
member_
->
executor_
.
reset
(
new
details
::
ParallelSSAGraphExecutor
(
exec_strategy
,
member_
->
local_scopes_
,
member_
->
places_
,
std
::
move
(
parallel_
graph
)));
exec_strategy
,
member_
->
local_scopes_
,
member_
->
places_
,
main_program
,
std
::
move
(
graph
)));
#else
PADDLE_THROW
(
"Paddle should be compiled with CUDA for ParallelGraph Execution."
);
...
...
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