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
be1373dc
P
Paddle
项目概览
PaddlePaddle
/
Paddle
大约 1 年 前同步成功
通知
2298
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
be1373dc
编写于
4月 01, 2018
作者:
X
Xin Pan
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Polish
上级
46f3a39e
变更
3
隐藏空白更改
内联
并排
Showing
3 changed file
with
11 addition
and
7 deletion
+11
-7
paddle/fluid/framework/details/nccl_all_reduce_op_handle.h
paddle/fluid/framework/details/nccl_all_reduce_op_handle.h
+1
-1
paddle/fluid/framework/details/op_handle_base.h
paddle/fluid/framework/details/op_handle_base.h
+3
-1
paddle/fluid/framework/details/threaded_ssa_graph_executor.cc
...le/fluid/framework/details/threaded_ssa_graph_executor.cc
+7
-5
未找到文件。
paddle/fluid/framework/details/nccl_all_reduce_op_handle.h
浏览文件 @
be1373dc
...
...
@@ -39,7 +39,7 @@ struct NCCLAllReduceOpHandle : public OpHandleBase {
// Delay and buffer nccl_all_reduce together can significantly increase
// performance. Disable this feature by returning false.
bool
Is
DelayedOp
()
override
{
return
true
;
};
bool
Is
MultiDeviceTransfer
()
override
{
return
true
;
};
protected:
void
RunImpl
()
override
;
...
...
paddle/fluid/framework/details/op_handle_base.h
浏览文件 @
be1373dc
...
...
@@ -55,7 +55,9 @@ class OpHandleBase {
void
AddOutput
(
VarHandleBase
*
out
);
virtual
bool
IsDelayedOp
()
{
return
false
;
}
// If the Op involves data transfer of multiple devices that
// will likely block other computations.
virtual
bool
IsMultiDeviceTransfer
()
{
return
false
;
}
protected:
virtual
void
RunImpl
()
=
0
;
...
...
paddle/fluid/framework/details/threaded_ssa_graph_executor.cc
浏览文件 @
be1373dc
...
...
@@ -50,7 +50,7 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
// together since we currently cannot overlap computation and memcpy streams.
// Should revisit it if overlapping is available.
std
::
unordered_set
<
OpHandleBase
*>
delayed_ops
;
std
::
unordered_set
<
OpHandleBase
*>
after
_delayed_ops
;
std
::
unordered_set
<
OpHandleBase
*>
blocked_by
_delayed_ops
;
std
::
unordered_set
<
VarHandleBase
*>
delayed_vars
;
auto
InsertPendingVar
=
[
&
pending_vars
,
&
ready_vars
](
VarHandleBase
&
var
)
{
...
...
@@ -119,7 +119,7 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
auto
run_all_ready_ops
=
[
&
]
{
for
(
auto
*
op
:
ready_ops
)
{
if
(
op
->
Is
DelayedOp
())
{
if
(
op
->
Is
MultiDeviceTransfer
())
{
delayed_ops
.
insert
(
op
);
delayed_vars
.
insert
(
op
->
outputs_
.
begin
(),
op
->
outputs_
.
end
());
ready_vars
.
Extend
(
op
->
outputs_
);
...
...
@@ -162,20 +162,22 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
--
deps
;
if
(
deps
==
0
)
{
if
(
delayed_vars
.
find
(
ready_var
)
!=
delayed_vars
.
end
())
{
after
_delayed_ops
.
insert
(
op
);
blocked_by
_delayed_ops
.
insert
(
op
);
}
else
{
ready_ops
.
insert
(
op
);
}
}
}
}
// When there are no other ops to schedule, schedule buffered delayed
// ops and unblock other ops.
if
(
ready_ops
.
empty
()
&&
!
delayed_ops
.
empty
()
&&
running_ops_
==
0
)
{
RunDelayedOps
(
delayed_ops
);
delayed_ops
.
clear
();
for
(
auto
*
op
:
after
_delayed_ops
)
{
for
(
auto
*
op
:
blocked_by
_delayed_ops
)
{
ready_ops
.
insert
(
op
);
}
after
_delayed_ops
.
clear
();
blocked_by
_delayed_ops
.
clear
();
}
// Keep loop until all vars are ready.
}
...
...
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