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
e67325cd
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
e67325cd
编写于
3月 14, 2018
作者:
Y
Yang Yang
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
update readme
上级
0621c327
变更
1
隐藏空白更改
内联
并排
Showing
1 changed file
with
32 addition
and
10 deletion
+32
-10
doc/design/parallel_executor.md
doc/design/parallel_executor.md
+32
-10
未找到文件。
doc/design/parallel_executor.md
浏览文件 @
e67325cd
...
...
@@ -30,23 +30,45 @@ operator run on each GPU, it will automatically sync with different streams when
// if op's input is params' grad:
// sync with allreduce stream
// e.g. sgd should wait for allreduce to be finished
SyncMultipleStreams
(
op
);
CallBack
->
BeforeOp
(
op
);
op
->
Run
(
*
local_scope
,
place_
);
// if op's output is params' grad:
// sync with computation stream
// e.g. allreduce shoudl wait for fc_grad to be finished.
SyncMultipleStreams
(
op
);
CallBack
->
AfterOp
(
op
);
```
And the
`Callback`
object can be implemented as the following
## API
```
c++
struct
AllReduceCallBack
{
void
BeforeOp
(
framework
::
OperatorBase
*
op
);
void
AfterOp
(
framework
::
OperatorBase
*
op
);
std
::
unordered_set
<
std
::
string
>
reduced_param_grad_names
;
std
::
unordered_set
<
std
::
string
>
param_grad_names_
;
platform
::
DeviceContext
*
computation_dev_ctx
;
// computation device context
platform
::
DeviceContext
*
communication_dev_ctx
;
// communication device context
The
`ParallelExecutor.run`
has similar interface as
`Executor.run`
. Besides
1.
Scope: we don't expose
`scope`
in
`ParallelExecutor.run`
since
`ParallelExecutor`
has its
own scope to maintain NCCL.
1.
Feed: we don't expose
`feed`
in the API either, because the whole point of implementing
parallel_executor is the speed. The input for NN should be implemented in an reader OP.
1.
Fetch: we return the fetched value on all GPUs as a list. (e.g.
`exe.run(..., fetch=loss)`
with return
`[loss_on_gpu0, loss_on_gpu1]`
)
framework
::
Scope
*
scope
;
platform
::
NCCL
::
Communicator
*
nccl_com
;
};
AllReduceCallBack
::
BeforeOp
(
framework
::
OperatorBase
*
op
)
{
if
(
op
->
Input
()
in
reduced_param_grad_names
)
{
communication_dev_ctx
->
Wait
();
reduced_param_grad_names
.
erase
(
op
->
Input
())
}
}
AllReduceCallBack
::
AfterOp
(
framework
::
OperatorBase
*
op
)
{
if
(
op
->
Output
()
in
param_grad_names
)
{
computation_dev_ctx
->
Wait
();
reduced_param_grad_names
.
insert
(
op
->
Output
());
ncclAllreduce
(
scope
,
op
->
Output
(),
communication_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