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
0621c327
P
Paddle
项目概览
PaddlePaddle
/
Paddle
12 个月 前同步成功
通知
2292
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
0621c327
编写于
3月 13, 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
init commit
上级
3621d9a3
变更
6
隐藏空白更改
内联
并排
Showing
6 changed file
with
148 addition
and
0 deletion
+148
-0
doc/design/parallel_executor.md
doc/design/parallel_executor.md
+52
-0
paddle/fluid/framework/CMakeLists.txt
paddle/fluid/framework/CMakeLists.txt
+2
-0
paddle/fluid/framework/executor.cc
paddle/fluid/framework/executor.cc
+13
-0
paddle/fluid/framework/executor.h
paddle/fluid/framework/executor.h
+1
-0
paddle/fluid/framework/parallel_executor.cc
paddle/fluid/framework/parallel_executor.cc
+19
-0
paddle/fluid/framework/parallel_executor.h
paddle/fluid/framework/parallel_executor.h
+61
-0
未找到文件。
doc/design/parallel_executor.md
0 → 100644
浏览文件 @
0621c327
# ParallelExecutor Design Doc
## Introduction
We introduce
`ParallelExecutor`
to run multi-GPU training in PaddlePaddle Fluid. It supports
1.
keeping a copy of the parameters on each GPU
1.
allreduce on a separate stream allowing computation and communication overlap
An example of switching single GPU training to multiple GPUs:
```
python
cost
=
your_neural_network
()
opt
=
fluid
.
optimizer
.
SGDOptimizer
()
opt
.
minimize
(
avg_cost
)
# change Executor -> ParallelExecutor
exe
=
fluid
.
ParallelExecutor
(
gpu_list
=
[
0
,
1
])
for
iter
in
xranges
(
iter_num
):
exe
.
run
()
```
## Design
In the constructor, a list of parameter, whose gradients need to be allreduced, is given.
During the runtime,
`ParallelExecutor`
starts
`#gpu`
threads to run each
`Executor`
. For every
operator run on each GPU, it will automatically sync with different streams when necessary.
```
c++
// if op's input is params' grad:
// sync with allreduce stream
// e.g. sgd should wait for allreduce to be finished
SyncMultipleStreams
(
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
);
```
## API
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]`
)
paddle/fluid/framework/CMakeLists.txt
浏览文件 @
0621c327
...
...
@@ -86,6 +86,8 @@ cc_library(feed_fetch_method SRCS feed_fetch_method.cc DEPS lod_tensor scope glo
cc_library
(
executor SRCS executor.cc DEPS op_registry device_context scope
framework_proto backward glog lod_rank_table feed_fetch_method
)
cc_library
(
parallel_executor SRCS parallel_executor.cc DEPS op_registry device_context scope
framework_proto backward glog lod_rank_table feed_fetch_method executor
)
cc_library
(
prune SRCS prune.cc DEPS framework_proto
)
cc_test
(
prune_test SRCS prune_test.cc DEPS op_info prune recurrent_op device_context
)
...
...
paddle/fluid/framework/executor.cc
浏览文件 @
0621c327
...
...
@@ -305,10 +305,23 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
}
// if (create_vars)
for
(
auto
&
op
:
ctx
->
ops_
)
{
// TODO(ty):
// e.g. sgd should wait for allreduce to be finished
// if op's input is params' grad:
// sync with allreduce stream
// SyncMultipleStreams(op);
VLOG
(
4
)
<<
place_
<<
" "
<<
op
->
DebugStringEx
(
local_scope
);
op
->
Run
(
*
local_scope
,
place_
);
VLOG
(
3
)
<<
place_
<<
" "
<<
op
->
DebugStringEx
(
local_scope
);
// TODO(ty):
// e.g. allreduce shoudl wait for fc_grad to be finished.
// if op's output is params' grad:
// sync with computation stream
// apply allreduce on allreduce stream
// SyncMultipleStreams(op);
if
(
FLAGS_benchmark
)
{
VLOG
(
2
)
<<
"Memory used after operator "
+
op
->
Type
()
+
" running: "
<<
memory
::
memory_usage
(
place_
);
...
...
paddle/fluid/framework/executor.h
浏览文件 @
0621c327
...
...
@@ -47,6 +47,7 @@ class Executor {
const
std
::
string
&
feed_holder_name
=
"feed"
,
const
std
::
string
&
fetch_holder_name
=
"fetch"
);
private:
static
ExecutorPrepareContext
*
Prepare
(
const
ProgramDesc
&
program
,
int
block_id
);
...
...
paddle/fluid/framework/parallel_executor.cc
0 → 100644
浏览文件 @
0621c327
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
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/parallel_executor.h"
namespace
paddle
{
namespace
framework
{}
// namespace framework
}
// namespace paddle
paddle/fluid/framework/parallel_executor.h
0 → 100644
浏览文件 @
0621c327
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
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. */
#pragma once
#include <unordered_set>
#include "paddle/fluid/framework/executor.h"
#include "paddle/fluid/framework/op_info.h"
#include "paddle/fluid/framework/program_desc.h"
#include "paddle/fluid/framework/scope.h"
#include "paddle/fluid/framework/tensor.h"
#include "paddle/fluid/operators/nccl/nccl_gpu_common.h"
#include "paddle/fluid/platform/device_context.h"
namespace
paddle
{
namespace
framework
{
struct
AllReduceCallBack
{
void
operator
()(
framework
::
OperatorBase
*
op
);
std
::
unordered_set
<
std
::
string
>
param_grad_names_
;
platform
::
DeviceContext
dev_ctx
;
};
class
ParallelExecutor
{
explicit
ParallelExecutor
(
const
std
::
vector
<
platform
::
Place
>&
places
,
const
std
::
unordered_set
&
params
);
/* @Brief
* Runtime evaluation of the given ProgramDesc under certain Scope
*
* @param
* ProgramDesc
* Scope
*/
void
Run
(
const
ProgramDesc
&
prog
,
Scope
*
scope
,
int
block_id
,
bool
create_local_scope
=
true
,
bool
create_vars
=
true
);
private:
std
::
vector
<
framework
::
Executor
>
exes_
;
std
::
vector
<
framework
::
Scope
*>
scopes_
;
AllReduceCallBack
all_reduce_callbacks_
;
std
::
unordered_set
<
std
::
string
>
params_
;
// where to initilize it?
platform
::
Communicator
nccl_com_
;
};
}
// namespace framework
}
// namespace paddle
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