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
OPTHREE
Paddle
提交
43c82376
P
Paddle
项目概览
OPTHREE
/
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
43c82376
编写于
2月 25, 2019
作者:
Q
Qiao Longfei
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
use one graph
上级
10393dd0
变更
3
隐藏空白更改
内联
并排
Showing
3 changed file
with
33 addition
and
46 deletion
+33
-46
paddle/fluid/framework/details/async_ssa_graph_executor.cc
paddle/fluid/framework/details/async_ssa_graph_executor.cc
+3
-4
paddle/fluid/framework/details/async_ssa_graph_executor.h
paddle/fluid/framework/details/async_ssa_graph_executor.h
+3
-3
paddle/fluid/framework/parallel_executor.cc
paddle/fluid/framework/parallel_executor.cc
+27
-39
未找到文件。
paddle/fluid/framework/details/async_ssa_graph_executor.cc
浏览文件 @
43c82376
...
...
@@ -21,15 +21,14 @@ namespace details {
AsyncSSAGraphExecutor
::
AsyncSSAGraphExecutor
(
const
ExecutionStrategy
&
strategy
,
const
std
::
vector
<
Scope
*>
&
local_scopes
,
const
std
::
vector
<
platform
::
Place
>
&
places
,
std
::
vector
<
std
::
unique_ptr
<
ir
::
Graph
>>
&&
graphs
)
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
)),
graph
s_
(
std
::
move
(
graphs
))
{
graph
_
(
std
::
move
(
graph
))
{
VLOG
(
3
)
<<
"build AsyncSSAGraphExecutor"
;
PADDLE_ENFORCE_EQ
(
places_
.
size
(),
local_scopes_
.
size
());
PADDLE_ENFORCE_EQ
(
graphs_
.
size
(),
local_scopes_
.
size
());
// set the correct size of thread pool to each device.
strategy_
.
num_threads_
=
strategy_
.
num_threads_
<
places_
.
size
()
...
...
@@ -39,7 +38,7 @@ AsyncSSAGraphExecutor::AsyncSSAGraphExecutor(
<<
" to run the operators of the graph on each device."
;
for
(
size_t
i
=
0
;
i
<
places
.
size
();
++
i
)
{
executors_
.
emplace_back
(
new
details
::
ThreadedSSAGraphExecutor
(
strategy_
,
{
local_scopes_
[
i
]},
{
places_
[
i
]},
std
::
move
(
graphs_
[
i
]
)));
strategy_
,
{
local_scopes_
[
i
]},
{
places_
[
i
]},
graph_
.
get
(
)));
}
}
...
...
paddle/fluid/framework/details/async_ssa_graph_executor.h
浏览文件 @
43c82376
...
...
@@ -29,9 +29,9 @@ class AsyncSSAGraphExecutor : public SSAGraphExecutor {
AsyncSSAGraphExecutor
(
const
ExecutionStrategy
&
strategy
,
const
std
::
vector
<
Scope
*>
&
local_scopes
,
const
std
::
vector
<
platform
::
Place
>
&
places
,
std
::
vector
<
std
::
unique_ptr
<
ir
::
Graph
>>
&&
graphs
);
std
::
unique_ptr
<
ir
::
Graph
>
&&
graph
);
~
AsyncSSAGraphExecutor
()
final
=
default
;
const
ir
::
Graph
&
Graph
()
const
override
{
return
*
graph
s_
[
0
]
;
}
const
ir
::
Graph
&
Graph
()
const
override
{
return
*
graph
_
;
}
FeedFetchList
Run
(
const
std
::
vector
<
std
::
string
>
&
fetch_tensors
)
override
;
...
...
@@ -40,7 +40,7 @@ class AsyncSSAGraphExecutor : public SSAGraphExecutor {
std
::
vector
<
Scope
*>
local_scopes_
;
std
::
unique_ptr
<::
ThreadPool
>
pool_
{
nullptr
};
std
::
vector
<
platform
::
Place
>
places_
;
std
::
vector
<
std
::
unique_ptr
<
ir
::
Graph
>>
graphs
_
;
std
::
unique_ptr
<
ir
::
Graph
>
graph
_
;
std
::
vector
<
std
::
unique_ptr
<
details
::
ThreadedSSAGraphExecutor
>>
executors_
;
ExceptionHolder
exception_holder_
;
...
...
paddle/fluid/framework/parallel_executor.cc
浏览文件 @
43c82376
...
...
@@ -264,71 +264,59 @@ ParallelExecutor::ParallelExecutor(
// Step 2. Convert main_program to SSA form and dependency graph. Also, insert
// ncclOp
std
::
vector
<
std
::
unique_ptr
<
ir
::
Graph
>>
graphs
;
std
::
unique_ptr
<
ir
::
Graph
>
graph
;
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
if
(
build_strategy
.
async_mode_
&&
!
build_strategy
.
is_distribution_
)
{
VLOG
(
3
)
<<
"use local async mode"
;
for
(
size_t
i
=
0
;
i
<
member_
->
places_
.
size
();
++
i
)
{
std
::
unique_ptr
<
ir
::
Graph
>
graph
=
build_strategy
.
Apply
(
main_program
,
{
member_
->
places_
[
i
]},
loss_var_name
,
{
member_
->
local_scopes_
[
i
]},
member_
->
nranks_
,
member_
->
use_cuda_
,
member_
->
nccl_ctxs_
.
get
());
graphs
.
push_back
(
std
::
move
(
graph
));
}
graph
=
build_strategy
.
Apply
(
main_program
,
{
member_
->
places_
[
0
]},
loss_var_name
,
{
member_
->
local_scopes_
[
0
]},
member_
->
nranks_
,
member_
->
use_cuda_
,
member_
->
nccl_ctxs_
.
get
());
}
else
{
std
::
unique_ptr
<
ir
::
Graph
>
graph
=
build_strategy
.
Apply
(
main_program
,
member_
->
places_
,
loss_var_name
,
member_
->
local_scopes_
,
member_
->
nranks_
,
member_
->
use_cuda_
,
member_
->
nccl_ctxs_
.
get
());
graphs
.
push_back
(
std
::
move
(
graph
));
graph
=
build_strategy
.
Apply
(
main_program
,
member_
->
places_
,
loss_var_name
,
member_
->
local_scopes_
,
member_
->
nranks_
,
member_
->
use_cuda_
,
member_
->
nccl_ctxs_
.
get
());
}
#else
if
(
build_strategy
.
async_mode_
&&
!
build_strategy
.
is_distribution_
)
{
VLOG
(
3
)
<<
"use local async mode"
;
for
(
size_t
i
=
0
;
i
<
member_
->
places_
.
size
();
++
i
)
{
std
::
unique_ptr
<
ir
::
Graph
>
graph
=
build_strategy
.
Apply
(
main_program
,
{
member_
->
places_
[
i
]},
loss_var_name
,
{
member_
->
local_scopes_
[
i
]},
member_
->
nranks_
,
member_
->
use_cuda_
);
graphs
.
push_back
(
std
::
move
(
graph
));
}
graph
=
build_strategy
.
Apply
(
main_program
,
{
member_
->
places_
[
0
]},
loss_var_name
,
{
member_
->
local_scopes_
[
0
]},
member_
->
nranks_
,
member_
->
use_cuda_
);
}
else
{
std
::
unique_ptr
<
ir
::
Graph
>
graph
=
build_strategy
.
Apply
(
main_program
,
member_
->
places_
,
loss_var_name
,
member_
->
local_scopes_
,
member_
->
nranks_
,
member_
->
use_cuda_
);
graphs
.
push_back
(
std
::
move
(
graph
));
graph
=
build_strategy
.
Apply
(
main_program
,
member_
->
places_
,
loss_var_name
,
member_
->
local_scopes_
,
member_
->
nranks_
,
member_
->
use_cuda_
);
}
#endif
auto
max_memory_size
=
GetEagerDeletionThreshold
();
VLOG
(
10
)
<<
"Eager Deletion Threshold "
<<
static_cast
<
float
>
(
max_memory_size
)
/
(
1
<<
30
);
if
(
max_memory_size
>=
0
)
{
for
(
size_t
i
=
0
;
i
<
graphs
.
size
();
++
i
)
{
graphs
[
i
]
=
member_
->
PrepareGCAndRefCnts
(
std
::
move
(
graphs
[
i
]),
static_cast
<
size_t
>
(
max_memory_size
));
}
graph
=
member_
->
PrepareGCAndRefCnts
(
std
::
move
(
graph
),
static_cast
<
size_t
>
(
max_memory_size
));
}
// Step 3. Create vars in each scope. Passes may also create new vars.
// skip control vars and empty vars
std
::
vector
<
details
::
VariableInfo
>
var_infos
;
for
(
auto
&
graph
:
graphs
)
{
for
(
auto
&
node
:
graph
->
Nodes
())
{
if
(
node
->
IsVar
()
&&
!
node
->
IsCtrlVar
()
&&
node
->
Var
())
{
var_infos
.
emplace_back
();
var_infos
.
back
().
name_
=
node
->
Var
()
->
Name
();
var_infos
.
back
().
type_
=
node
->
Var
()
->
GetType
();
var_infos
.
back
().
persistable_
=
node
->
Var
()
->
Persistable
();
}
for
(
auto
&
node
:
graph
->
Nodes
())
{
if
(
node
->
IsVar
()
&&
!
node
->
IsCtrlVar
()
&&
node
->
Var
())
{
var_infos
.
emplace_back
();
var_infos
.
back
().
name_
=
node
->
Var
()
->
Name
();
var_infos
.
back
().
type_
=
node
->
Var
()
->
GetType
();
var_infos
.
back
().
persistable_
=
node
->
Var
()
->
Persistable
();
}
}
// If the loss_var_name is given, the number of graph should be only one.
if
(
loss_var_name
.
size
())
{
size_t
graph_num
=
ir
::
GraphNum
(
*
graph
s
[
0
]
);
size_t
graph_num
=
ir
::
GraphNum
(
*
graph
);
if
(
graph_num
>
1
)
{
LOG
(
WARNING
)
<<
"The number of graph should be only one, "
"but the current graph has "
<<
ir
::
GraphNum
(
*
graph
s
[
0
]
)
<<
ir
::
GraphNum
(
*
graph
)
<<
" sub_graphs. If you want to see the nodes of the "
"sub_graphs, you should use 'FLAGS_print_sub_graph_dir' "
"to specify the output dir. NOTES: if you not do training, "
...
...
@@ -340,7 +328,7 @@ ParallelExecutor::ParallelExecutor(
VLOG
(
3
)
<<
"use AsyncSSAGraphExecutor"
;
member_
->
executor_
.
reset
(
new
details
::
AsyncSSAGraphExecutor
(
exec_strategy
,
member_
->
local_scopes_
,
member_
->
places_
,
std
::
move
(
graph
s
)));
std
::
move
(
graph
)));
}
else
if
(
build_strategy
.
enable_parallel_graph_
)
{
VLOG
(
3
)
<<
"use ParallelSSAGraphExecutor"
;
#ifdef PADDLE_WITH_CUDA
...
...
@@ -358,12 +346,12 @@ ParallelExecutor::ParallelExecutor(
VLOG
(
3
)
<<
"use ThreadedSSAGraphExecutor"
;
member_
->
executor_
.
reset
(
new
details
::
ThreadedSSAGraphExecutor
(
exec_strategy
,
member_
->
local_scopes_
,
member_
->
places_
,
std
::
move
(
graph
s
[
0
]
)));
std
::
move
(
graph
)));
}
else
{
VLOG
(
3
)
<<
"use FastThreadedSSAGraphExecutor"
;
member_
->
executor_
.
reset
(
new
details
::
FastThreadedSSAGraphExecutor
(
exec_strategy
,
member_
->
local_scopes_
,
member_
->
places_
,
std
::
move
(
graph
s
[
0
]
)));
std
::
move
(
graph
)));
}
}
...
...
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