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
wmsofts
Paddle
提交
81d93514
P
Paddle
项目概览
wmsofts
/
Paddle
与 Fork 源项目一致
Fork自
PaddlePaddle / Paddle
通知
1
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
81d93514
编写于
3月 30, 2018
作者:
C
chengduo
提交者:
GitHub
3月 30, 2018
浏览文件
操作
浏览文件
下载
差异文件
Merge pull request #9522 from chengduoZH/feature/refine_parallel_exe
Refine parallel
上级
23bab34c
60d0a059
变更
1
隐藏空白更改
内联
并排
Showing
1 changed file
with
29 addition
and
15 deletion
+29
-15
paddle/fluid/framework/parallel_executor.cc
paddle/fluid/framework/parallel_executor.cc
+29
-15
未找到文件。
paddle/fluid/framework/parallel_executor.cc
浏览文件 @
81d93514
...
...
@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/framework/parallel_executor.h"
#include <string>
#include "ThreadPool.h"
...
...
@@ -102,30 +103,43 @@ void ParallelExecutor::BCastParamsToGPUs(
auto
*
main_scope
=
member_
->
local_scopes_
[
0
];
for
(
auto
*
var_desc
:
startup_program
.
Block
(
0
).
AllVars
())
{
size_t
idx
=
var_desc
->
Name
().
find
(
"@GRAD"
);
if
(
idx
!=
std
::
string
::
npos
)
continue
;
if
(
var_desc
->
GetType
()
==
proto
::
VarType
::
LOD_TENSOR
)
{
auto
&
main_tensor
=
main_scope
->
FindVar
(
var_desc
->
Name
())
->
Get
<
LoDTensor
>
();
ncclDataType_t
data_type
=
platform
::
ToNCCLDataType
(
main_tensor
.
type
());
auto
&
dims
=
main_tensor
.
dims
();
size_t
numel
=
main_tensor
.
numel
();
platform
::
NCCLGroupGuard
guard
;
auto
&
dims
=
main_tensor
.
dims
()
;
for
(
size_t
i
=
0
;
i
<
member_
->
places_
.
size
();
++
i
)
{
auto
place
=
member_
->
places_
[
i
];
void
*
buffer
;
if
(
i
==
0
)
{
buffer
=
const_cast
<
void
*>
(
main_tensor
.
data
<
void
>
());
}
else
{
if
(
paddle
::
platform
::
is_gpu_place
(
main_tensor
.
place
()))
{
size_t
numel
=
main_tensor
.
numel
();
ncclDataType_t
data_type
=
platform
::
ToNCCLDataType
(
main_tensor
.
type
());
platform
::
NCCLGroupGuard
guard
;
for
(
size_t
i
=
0
;
i
<
member_
->
places_
.
size
();
++
i
)
{
auto
place
=
member_
->
places_
[
i
];
void
*
buffer
;
if
(
i
==
0
)
{
buffer
=
const_cast
<
void
*>
(
main_tensor
.
data
<
void
>
());
}
else
{
auto
local_scope
=
member_
->
local_scopes_
[
i
];
auto
*
t
=
local_scope
->
Var
(
var_desc
->
Name
())
->
GetMutable
<
LoDTensor
>
();
t
->
Resize
(
dims
);
buffer
=
t
->
mutable_data
(
place
,
main_tensor
.
type
());
}
auto
&
nccl_ctx
=
member_
->
nccl_ctxs_
->
at
(
place
);
platform
::
dynload
::
ncclBcast
(
buffer
,
numel
,
data_type
,
0
,
nccl_ctx
.
comm_
,
nccl_ctx
.
stream
());
}
}
else
{
platform
::
CPUPlace
cpu
;
for
(
size_t
i
=
1
;
i
<
member_
->
places_
.
size
();
++
i
)
{
auto
local_scope
=
member_
->
local_scopes_
[
i
];
auto
*
t
=
local_scope
->
Var
(
var_desc
->
Name
())
->
GetMutable
<
LoDTensor
>
();
t
->
Resize
(
dims
);
buffer
=
t
->
mutable_data
(
place
,
main_tensor
.
type
());
t
->
mutable_data
(
cpu
,
main_tensor
.
type
());
paddle
::
framework
::
TensorCopy
(
main_tensor
,
cpu
,
t
);
}
auto
&
nccl_ctx
=
member_
->
nccl_ctxs_
->
at
(
place
);
platform
::
dynload
::
ncclBcast
(
buffer
,
numel
,
data_type
,
0
,
nccl_ctx
.
comm_
,
nccl_ctx
.
stream
());
}
}
member_
->
nccl_ctxs_
->
WaitAll
();
...
...
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