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
马赛克医院
Paddle
提交
52200523
P
Paddle
项目概览
马赛克医院
/
Paddle
与 Fork 源项目一致
Fork自
PaddlePaddle / Paddle
通知
1
Star
1
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 >>
提交
52200523
编写于
10月 26, 2017
作者:
D
Dong Zhihong
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
"polish code based on comment"
上级
6cce5268
变更
2
隐藏空白更改
内联
并排
Showing
2 changed file
with
26 addition
and
3 deletion
+26
-3
paddle/operators/nccl_op.cc
paddle/operators/nccl_op.cc
+8
-0
paddle/operators/nccl_op.cu
paddle/operators/nccl_op.cu
+18
-3
未找到文件。
paddle/operators/nccl_op.cc
浏览文件 @
52200523
...
...
@@ -94,6 +94,11 @@ class NCCLReduceOp : public framework::OperatorWithKernel {
PADDLE_ENFORCE
(
ctx
->
HasOutput
(
"Out"
),
" Input(X) of Reduce op input should not be NULL"
);
std
::
string
reduction
=
ctx
->
Attrs
().
Get
<
std
::
string
>
(
"reduction"
);
PADDLE_ENFORCE
((
reduction
==
"ncclSum"
||
reduction
==
"ncclProd"
||
reduction
==
"ncclMin"
||
reduction
==
"ncclMax"
),
"invalid reduction."
);
auto
x_dims
=
ctx
->
GetInputsDim
(
"X"
);
ctx
->
SetOutputsDim
(
"Out"
,
x_dims
);
ctx
->
ShareLoD
(
"X"
,
/*->*/
"Out"
);
...
...
@@ -150,6 +155,9 @@ class NCCLReduceOpMaker : public framework::OpProtoAndCheckerMaker {
AddInput
(
"X"
,
"The input of Reduce op"
);
AddInput
(
"Communicator"
,
"Communicator for communicating between gpus"
);
AddOutput
(
"Out"
,
"The output of Reduce op"
);
AddAttr
<
std
::
string
>
(
"reduction"
,
"{'ncclMin', 'ncclMax', 'ncclProd', 'ncclSum'}."
)
.
SetDefault
(
"ncclSum"
);
AddAttr
<
int
>
(
"root"
,
"root gpu of the parameter. if not "
"set(platform::kInvalidGPUId). hashed by name."
)
...
...
paddle/operators/nccl_op.cu
浏览文件 @
52200523
...
...
@@ -49,7 +49,6 @@ class NCCLAllReduceKernel : public framework::OpKernel<T> {
auto
outs
=
ctx
.
MultiOutput
<
LoDTensor
>
(
"Out"
);
std
::
string
reduction
=
ctx
.
Attr
<
std
::
string
>
(
"reduction"
);
ncclRedOp_t
reduction_op_
=
ncclSum
;
if
(
reduction
==
"ncclMin"
)
{
...
...
@@ -101,8 +100,23 @@ class NCCLReduceKernel : public framework::OpKernel<T> {
auto
ins
=
ctx
.
MultiInput
<
LoDTensor
>
(
"X"
);
// x0, x1, x2
auto
outs
=
ctx
.
MultiOutput
<
LoDTensor
>
(
"Out"
);
int
root
=
ctx
.
Attr
<
int
>
(
"root"
);
std
::
string
reduction
=
ctx
.
Attr
<
std
::
string
>
(
"reduction"
);
ncclRedOp_t
reduction_op_
=
ncclSum
;
if
(
reduction
==
"ncclMin"
)
{
reduction_op_
=
ncclMin
;
}
else
if
(
reduction
==
"ncclMax"
)
{
reduction_op_
=
ncclMax
;
}
else
if
(
reduction
==
"ncclSum"
)
{
reduction_op_
=
ncclSum
;
}
else
if
(
reduction
==
"ncclProd"
)
{
reduction_op_
=
ncclProd
;
}
else
{
PADDLE_ENFORCE
(
false
,
"Invalid reduction. default ncclSum."
);
}
int
root
=
ctx
.
Attr
<
int
>
(
"root"
);
auto
*
comm
=
ctx
.
Input
<
Communicator
>
(
"Communicator"
);
auto
stream
=
reinterpret_cast
<
const
platform
::
CUDADeviceContext
&>
(
...
...
@@ -128,7 +142,8 @@ class NCCLReduceKernel : public framework::OpKernel<T> {
PADDLE_ENFORCE
(
platform
::
dynload
::
ncclReduce
(
ins
[
i
]
->
data
<
T
>
(),
recvbuffer
,
ins
[
i
]
->
numel
(),
NCCLTypeWrapper
<
T
>::
type
,
ncclSum
,
root
,
comm
->
comms_
[
idx
],
stream
));
NCCLTypeWrapper
<
T
>::
type
,
reduction_op_
,
root
,
comm
->
comms_
[
idx
],
stream
));
PADDLE_ENFORCE
(
cudaStreamSynchronize
(
stream
));
VLOG
(
1
)
<<
"gpu : "
<<
gpu_id
<<
" finished reduce. send "
...
...
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