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
wk1234123
Paddle
提交
da556ed6
P
Paddle
项目概览
wk1234123
/
Paddle
与 Fork 源项目一致
Fork自
PaddlePaddle / Paddle
通知
4
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
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 >>
未验证
提交
da556ed6
编写于
6月 22, 2018
作者:
C
chengduo
提交者:
GitHub
6月 22, 2018
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
enhance ParallelExecutor stable (#11637)
上级
073af623
变更
5
显示空白变更内容
内联
并排
Showing
5 changed file
with
25 addition
and
84 deletion
+25
-84
paddle/fluid/framework/details/broadcast_op_handle.cc
paddle/fluid/framework/details/broadcast_op_handle.cc
+15
-42
paddle/fluid/framework/details/op_handle_base.cc
paddle/fluid/framework/details/op_handle_base.cc
+9
-27
paddle/fluid/framework/details/op_handle_base.h
paddle/fluid/framework/details/op_handle_base.h
+0
-4
paddle/fluid/framework/details/reduce_op_handle.cc
paddle/fluid/framework/details/reduce_op_handle.cc
+1
-3
paddle/fluid/platform/device_context.h
paddle/fluid/platform/device_context.h
+0
-8
未找到文件。
paddle/fluid/framework/details/broadcast_op_handle.cc
浏览文件 @
da556ed6
...
...
@@ -103,13 +103,6 @@ void BroadcastOpHandle::RunImpl() {
});
}
// FIXME(zcd): a temporary fix for some language model that has sparse
// parameter.
bool
use_mutex
=
true
;
if
(
in_var
->
IsType
<
paddle
::
framework
::
SelectedRows
>
())
{
use_mutex
=
false
;
}
if
(
use_mutex
)
{
this
->
RunAndRecordEvent
([
&
]
{
{
platform
::
NCCLGroupGuard
guard
;
...
...
@@ -127,26 +120,6 @@ void BroadcastOpHandle::RunImpl() {
&
VariableVisitor
::
GetMutableTensor
(
out_var
));
}
});
}
else
{
this
->
RunAndRecordEventNoMutex
([
&
]
{
{
platform
::
NCCLGroupGuard
guard
;
for
(
auto
&
call
:
broadcast_calls
)
{
call
();
}
}
if
(
!
out_handle
->
IsTheSameVar
(
*
in_var_handle
))
{
auto
out_var
=
var_scopes
.
at
(
in_var_handle
->
scope_idx_
)
->
FindVar
(
out_var_handles
[
0
]
->
name_
);
paddle
::
framework
::
TensorCopy
(
in_tensor
,
in_var_handle
->
place_
,
*
(
dev_ctxes_
.
at
(
in_var_handle
->
place_
)),
&
VariableVisitor
::
GetMutableTensor
(
out_var
));
}
});
}
#else
PADDLE_THROW
(
"CUDA is not enabled."
);
#endif
...
...
paddle/fluid/framework/details/op_handle_base.cc
浏览文件 @
da556ed6
...
...
@@ -11,8 +11,8 @@
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "paddle/fluid/framework/details/op_handle_base.h"
#include <map>
namespace
paddle
{
namespace
framework
{
...
...
@@ -122,34 +122,16 @@ void OpHandleBase::RunAndRecordEvent(const std::function<void()> &callback) {
#ifdef PADDLE_WITH_CUDA
if
(
!
events_
.
empty
())
{
// Use event
std
::
function
<
void
()
>
method
=
callback
;
// NOTE(zcd): device context must be ordered here because RecordEvent
// will use a mutex to ensure the safe of multi-threads.
std
::
map
<
platform
::
DeviceContext
*
,
platform
::
Place
>
ordered_ctxes
;
for
(
auto
&
p
:
dev_ctxes_
)
{
method
=
[
method
,
p
,
this
]()
{
static_cast
<
platform
::
CUDADeviceContext
*>
(
p
.
second
)
->
RecordEvent
(
events_
.
at
(
boost
::
get
<
platform
::
CUDAPlace
>
(
p
.
first
).
device
),
method
);
};
}
method
();
}
else
{
#endif
callback
();
#ifdef PADDLE_WITH_CUDA
ordered_ctxes
.
emplace
(
p
.
second
,
p
.
first
);
}
#endif
}
void
OpHandleBase
::
RunAndRecordEventNoMutex
(
const
std
::
function
<
void
()
>
&
callback
)
{
#ifdef PADDLE_WITH_CUDA
if
(
!
events_
.
empty
())
{
// Use event
std
::
function
<
void
()
>
method
=
callback
;
for
(
auto
&
p
:
dev_ctxes_
)
{
for
(
auto
&
p
:
ordered_ctxes
)
{
method
=
[
method
,
p
,
this
]()
{
static_cast
<
platform
::
CUDADeviceContext
*>
(
p
.
second
)
->
RecordEventNoMutex
(
events_
.
at
(
boost
::
get
<
platform
::
CUDAPlace
>
(
p
.
first
).
device
),
static_cast
<
platform
::
CUDADeviceContext
*>
(
p
.
first
)
->
RecordEvent
(
events_
.
at
(
boost
::
get
<
platform
::
CUDAPlace
>
(
p
.
second
).
device
),
method
);
};
}
...
...
paddle/fluid/framework/details/op_handle_base.h
浏览文件 @
da556ed6
...
...
@@ -85,10 +85,6 @@ class OpHandleBase {
protected:
void
RunAndRecordEvent
(
const
std
::
function
<
void
()
>
&
callback
);
// FIXME(zcd): A temporary fix for some language model that has sparse
// parameter.
void
RunAndRecordEventNoMutex
(
const
std
::
function
<
void
()
>
&
callback
);
void
RunAndRecordEvent
(
platform
::
Place
p
,
const
std
::
function
<
void
()
>
&
callback
);
...
...
paddle/fluid/framework/details/reduce_op_handle.cc
浏览文件 @
da556ed6
...
...
@@ -80,9 +80,7 @@ void ReduceOpHandle::RunImpl() {
}
if
(
pre_in_var
->
IsType
<
framework
::
SelectedRows
>
())
{
// FIXME(zcd): A temporary fix for some language model that has sparse
// parameter.
this
->
RunAndRecordEventNoMutex
([
&
]
{
this
->
RunAndRecordEvent
([
&
]
{
std
::
vector
<
const
SelectedRows
*>
in_selected_rows
=
GetInputValues
<
SelectedRows
>
(
in_var_handles
,
var_scopes
);
GatherSelectedRows
(
in_selected_rows
,
in_places
,
dev_ctxes_
,
t_out_p
,
...
...
paddle/fluid/platform/device_context.h
浏览文件 @
da556ed6
...
...
@@ -106,14 +106,6 @@ class CUDADeviceContext : public DeviceContext {
PADDLE_ENFORCE
(
cudaEventRecord
(
ev
,
stream_
));
}
// FIXME(zcd): A temporary fix for some language model that has sparse
// parameter.
template
<
typename
Callback
>
void
RecordEventNoMutex
(
cudaEvent_t
ev
,
Callback
callback
)
{
callback
();
PADDLE_ENFORCE
(
cudaEventRecord
(
ev
,
stream_
));
}
private:
CUDAPlace
place_
;
...
...
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