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
ab369976
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 >>
未验证
提交
ab369976
编写于
10月 19, 2022
作者:
C
Chen Weihang
提交者:
GitHub
10月 19, 2022
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
remove fluid symbol depend in sync bn (#47122)
上级
d00b7d83
变更
2
显示空白变更内容
内联
并排
Showing
2 changed file
with
36 addition
and
10 deletion
+36
-10
paddle/phi/kernels/gpu/sync_batch_norm_kernel.cu
paddle/phi/kernels/gpu/sync_batch_norm_kernel.cu
+22
-10
paddle/phi/kernels/sync_batch_norm_kernel.h
paddle/phi/kernels/sync_batch_norm_kernel.h
+14
-0
未找到文件。
paddle/phi/kernels/gpu/sync_batch_norm_kernel.cu
浏览文件 @
ab369976
...
...
@@ -18,6 +18,26 @@
#include "paddle/phi/kernels/gpu/sync_batch_norm_utils.h"
namespace
phi
{
namespace
detail
{
ccl
::
CCLComm
GetCCLComm
(
const
Place
&
place
,
int
global_gid
)
{
#if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL)
ncclComm_t
comm
=
nullptr
;
if
(
paddle
::
distributed
::
ProcessGroupMapFromGid
::
getInstance
()
->
has
(
global_gid
))
{
auto
*
nccl_pg
=
static_cast
<
paddle
::
distributed
::
ProcessGroupNCCL
*>
(
paddle
::
distributed
::
ProcessGroupMapFromGid
::
getInstance
()
->
get
(
global_gid
));
comm
=
nccl_pg
->
NCCLComm
(
place
);
}
return
comm
;
#else
return
nullptr
;
#endif
}
}
// namespace detail
template
<
typename
T
,
typename
Context
>
void
SyncBatchNormKernel
(
const
Context
&
ctx
,
...
...
@@ -102,16 +122,8 @@ void SyncBatchNormKernel(const Context &ctx,
}
#if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL)
int
global_gid
=
0
;
ncclComm_t
comm
=
nullptr
;
if
(
paddle
::
distributed
::
ProcessGroupMapFromGid
::
getInstance
()
->
has
(
global_gid
))
{
auto
*
nccl_pg
=
static_cast
<
paddle
::
distributed
::
ProcessGroupNCCL
*>
(
paddle
::
distributed
::
ProcessGroupMapFromGid
::
getInstance
()
->
get
(
global_gid
));
comm
=
nccl_pg
->
NCCLComm
(
x
.
place
());
}
else
{
ncclComm_t
comm
=
static_cast
<
ncclComm_t
>
(
detail
::
GetCCLComm
(
x
.
place
(),
0
));
if
(
comm
==
nullptr
)
{
comm
=
ctx
.
nccl_comm
();
}
...
...
paddle/phi/kernels/sync_batch_norm_kernel.h
浏览文件 @
ab369976
...
...
@@ -16,9 +16,23 @@
#include <string>
#include "paddle/phi/backends/c_comm_lib.h"
#include "paddle/phi/core/dense_tensor.h"
namespace
phi
{
namespace
detail
{
// FIXME(paddle-dev): Since the singleton of ProcessGroup in fluid is used in
// SyncBN, the fluid symbol will be dependent on external hardware access.
// Here, the part that depends on the fluid symbol is individually encapsulated
// as a temporary function to isolate external symbol dependencies.
// In the future, the dependence on the singleton in fluid in SyncBN needs
// to be removed.
// In principle, the PHI Kernel cannot use the global singleton internally,
// and the required members need to be passed in from the eucalyptus tree.
ccl
::
CCLComm
GetCCLComm
(
const
Place
&
place
,
int
global_gid
=
0
);
}
// namespace detail
template
<
typename
T
,
typename
Context
>
void
SyncBatchNormKernel
(
const
Context
&
dev_ctx
,
...
...
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