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
s920243400
PaddleDetection
提交
d5090c89
P
PaddleDetection
项目概览
s920243400
/
PaddleDetection
与 Fork 源项目一致
Fork自
PaddlePaddle / PaddleDetection
通知
2
Star
0
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
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
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
提交
d5090c89
编写于
2月 19, 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
polish code test=develop
上级
0f8bd73c
变更
3
显示空白变更内容
内联
并排
Showing
3 changed file
with
10 addition
and
11 deletion
+10
-11
paddle/fluid/framework/details/build_strategy.cc
paddle/fluid/framework/details/build_strategy.cc
+1
-1
paddle/fluid/framework/details/multi_devices_graph_pass.cc
paddle/fluid/framework/details/multi_devices_graph_pass.cc
+7
-9
paddle/fluid/framework/details/parallel_ssa_graph_executor.cc
...le/fluid/framework/details/parallel_ssa_graph_executor.cc
+2
-1
未找到文件。
paddle/fluid/framework/details/build_strategy.cc
浏览文件 @
d5090c89
...
...
@@ -34,7 +34,7 @@ namespace details {
static
inline
bool
SeqOnlyAllReduceOps
(
const
BuildStrategy
&
strategy
)
{
// Should fix the allreduce op order if scheduling
// them in multiple threads or processes to avoid hang.
// NOTE: Parallel
Executor
would execute this pass on each graph, so
// NOTE: Parallel
Graph
would execute this pass on each graph, so
// don't need to append it here.
return
(
!
strategy
.
enable_sequential_execution_
&&
strategy
.
num_trainers_
>
1
)
&&
...
...
paddle/fluid/framework/details/multi_devices_graph_pass.cc
浏览文件 @
d5090c89
...
...
@@ -389,8 +389,8 @@ void MultiDevSSAGraphBuilderBase::CreateAllReduceOp(
OpHandleBase
*
op_handle
=
nullptr
;
auto
append_allreduce_op
=
[
&
](
std
::
vector
<
Scope
*>
&
scopes
,
std
::
vector
<
platform
::
Place
>
&
places
)
->
OpHandleBase
*
{
const
std
::
vector
<
Scope
*>
&
scopes
,
const
std
::
vector
<
platform
::
Place
>
&
places
)
->
OpHandleBase
*
{
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
result
->
Get
<
GraphOps
>
(
kGraphOps
).
emplace_back
(
new
AllReduceOpHandle
(
result
->
CreateEmptyNode
(
"allreduce"
,
ir
::
Node
::
Type
::
kOperation
),
...
...
@@ -407,13 +407,11 @@ void MultiDevSSAGraphBuilderBase::CreateAllReduceOp(
op_handle
=
append_allreduce_op
(
local_scopes_
,
places_
);
for
(
size_t
i
=
0
;
i
<
places_
.
size
();
++
i
)
{
auto
p
=
places_
[
i
];
std
::
vector
<
Scope
*>
ss
{
local_scopes_
[
i
]};
std
::
vector
<
platform
::
Place
>
ps
{
p
};
if
(
strategy_
.
enable_parallel_graph_
)
op_handle
=
append_allreduce_op
(
ss
,
ps
);
if
(
strategy_
.
enable_parallel_graph_
)
{
op_handle
=
append_allreduce_op
({
local_scopes_
[
i
]},
{
places_
[
i
]});
}
SetCommunicationContext
(
op_handle
,
p
);
SetCommunicationContext
(
op_handle
,
p
laces_
[
i
]
);
auto
&
vars
=
result
->
Get
<
GraphVars
>
(
kGraphVars
)[
i
][
og
];
PADDLE_ENFORCE
(
!
vars
.
empty
());
auto
&
prev_grad
=
vars
.
back
();
...
...
@@ -421,7 +419,7 @@ void MultiDevSSAGraphBuilderBase::CreateAllReduceOp(
auto
var
=
new
VarHandle
(
result
->
CreateEmptyNode
(
og
,
ir
::
Node
::
Type
::
kVariable
),
vars
.
size
(),
i
,
og
,
p
);
vars
.
size
(),
i
,
og
,
p
laces_
[
i
]
);
vars
.
emplace_back
(
var
);
op_handle
->
AddOutput
(
var
);
}
...
...
paddle/fluid/framework/details/parallel_ssa_graph_executor.cc
浏览文件 @
d5090c89
...
...
@@ -32,8 +32,9 @@ ParallelSSAGraphExecutor::SeparateMultiDevicesGraph(
g
->
Set
(
kGraphDepVars
,
new
GraphDepVars
);
g
->
Set
(
kGraphOps
,
new
GraphOps
);
}
auto
op_handles
=
ir
::
FilterByNodeWrapper
<
OpHandleBase
>
(
*
graph
);
for
(
auto
&
op
:
graph
->
Get
<
GraphOps
>
(
kGraphOps
)
)
{
for
(
auto
&
op
:
op_handles
)
{
auto
&
dev_ctx
=
op
->
DeviceContext
();
auto
&
p
=
dev_ctx
.
begin
()
->
first
;
int
dev_id
=
boost
::
get
<
platform
::
CUDAPlace
>
(
p
).
device
;
...
...
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