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
8c38aca9
P
Paddle
项目概览
PaddlePaddle
/
Paddle
1 年多 前同步成功
通知
2302
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
提交
8c38aca9
编写于
3月 05, 2019
作者:
Q
Qiao Longfei
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
tmp commit
上级
fab1b54d
变更
3
隐藏空白更改
内联
并排
Showing
3 changed file
with
69 addition
and
7 deletion
+69
-7
paddle/fluid/framework/details/CMakeLists.txt
paddle/fluid/framework/details/CMakeLists.txt
+1
-1
paddle/fluid/framework/details/async_ssa_graph_executor.cc
paddle/fluid/framework/details/async_ssa_graph_executor.cc
+38
-0
paddle/fluid/operators/distributed/communicator.h
paddle/fluid/operators/distributed/communicator.h
+30
-6
未找到文件。
paddle/fluid/framework/details/CMakeLists.txt
浏览文件 @
8c38aca9
...
...
@@ -82,7 +82,7 @@ cc_library(threaded_ssa_graph_executor SRCS threaded_ssa_graph_executor.cc DEPS
cc_library
(
parallel_ssa_graph_executor SRCS parallel_ssa_graph_executor.cc DEPS threaded_ssa_graph_executor
)
cc_library
(
async_ssa_graph_executor SRCS async_ssa_graph_executor.cc DEPS threaded_ssa_graph_executor
)
cc_library
(
async_ssa_graph_executor SRCS async_ssa_graph_executor.cc DEPS threaded_ssa_graph_executor
communicator
)
cc_test
(
broadcast_op_test SRCS broadcast_op_handle_test.cc DEPS var_handle op_handle_base scope ddim memory
device_context broadcast_op_handle
)
...
...
paddle/fluid/framework/details/async_ssa_graph_executor.cc
浏览文件 @
8c38aca9
...
...
@@ -15,6 +15,7 @@
#include "paddle/fluid/framework/details/async_ssa_graph_executor.h"
#include "paddle/fluid/framework/variable_helper.h"
#include "paddle/fluid/operators/distributed/communicator.h"
namespace
paddle
{
namespace
framework
{
...
...
@@ -39,6 +40,43 @@ inline void NewTempScopeAndInitVars(const std::vector<VarInfo> &var_infos,
}
}
// get RpcContext and remote send and recv op
void
ProcessGraph
(
std
::
vector
<
ir
::
Graph
*>
graphs
,
Scope
*
scope
)
{
using
RpcCtxMap
=
operators
::
distributed
::
RpcCtxMap
;
RpcCtxMap
send_varname_to_ctx
;
RpcCtxMap
recv_varname_to_ctx
;
for
(
auto
i
=
0
;
i
<
graphs
.
size
();
++
i
)
{
for
(
auto
&
node
:
graphs
[
i
]
->
Nodes
())
{
if
(
node
->
IsOp
())
{
if
(
node
->
Op
()
->
Type
()
==
"send"
)
{
auto
send_var_name
=
node
->
Op
()
->
Input
(
"X"
)[
0
];
auto
send_varnames
=
boost
::
get
<
std
::
vector
<
std
::
string
>>
(
node
->
Op
()
->
GetNullableAttr
(
"send_varnames"
));
auto
epmap
=
boost
::
get
<
std
::
vector
<
std
::
string
>>
(
node
->
Op
()
->
GetNullableAttr
(
"epmap"
));
auto
height_section
=
boost
::
get
<
std
::
vector
<
int64_t
>>
(
node
->
Op
()
->
GetNullableAttr
(
"sections"
));
send_varname_to_ctx
[
send_var_name
]
=
operators
::
distributed
::
RpcContext
(
send_var_name
,
send_varnames
,
epmap
,
height_section
);
}
else
if
(
node
->
Op
()
->
Type
()
==
"recv"
)
{
auto
recv_var_name
=
node
->
Op
()
->
Input
(
"X"
)[
0
];
auto
recv_varnames
=
boost
::
get
<
std
::
vector
<
std
::
string
>>
(
node
->
Op
()
->
GetNullableAttr
(
"recv_varnames"
));
auto
epmap
=
boost
::
get
<
std
::
vector
<
std
::
string
>>
(
node
->
Op
()
->
GetNullableAttr
(
"epmap"
));
recv_varname_to_ctx
[
recv_var_name
]
=
operators
::
distributed
::
RpcContext
(
recv_var_name
,
recv_varnames
,
epmap
,
{});
}
}
}
}
// init communicator here
operators
::
distributed
::
Communicator
::
Init
(
send_varname_to_ctx
,
recv_varname_to_ctx
,
scope
);
}
AsyncSSAGraphExecutor
::
AsyncSSAGraphExecutor
(
const
ExecutionStrategy
&
strategy
,
const
std
::
vector
<
Scope
*>
&
local_scopes
,
const
std
::
vector
<
platform
::
Place
>
&
places
,
std
::
vector
<
ir
::
Graph
*>
graphs
)
...
...
paddle/fluid/operators/distributed/communicator.h
浏览文件 @
8c38aca9
...
...
@@ -87,12 +87,12 @@ class BlockingQueue {
std
::
condition_variable
send_cv_
;
};
using
RpcCtxMap
=
std
::
unordered_map
<
std
::
string
,
RpcContext
>
;
class
Communicator
{
public:
Communicator
(
const
std
::
unordered_map
<
std
::
string
,
RpcContext
>&
send_varname_to_ctx
,
const
std
::
unordered_map
<
std
::
string
,
RpcContext
>&
recv_varname_to_ctx
,
Scope
*
recv_scope
)
Communicator
(
const
RpcCtxMap
&
send_varname_to_ctx
,
const
RpcCtxMap
&
recv_varname_to_ctx
,
Scope
*
recv_scope
)
:
send_varname_to_ctx_
(
send_varname_to_ctx
),
recv_varname_to_ctx_
(
recv_varname_to_ctx
),
recv_scope_
(
recv_scope
)
{
...
...
@@ -128,14 +128,38 @@ class Communicator {
std
::
unordered_map
<
std
::
string
,
std
::
shared_ptr
<
BlockingQueue
<
std
::
shared_ptr
<
Variable
>>>>
send_varname_to_queue_
;
std
::
unordered_map
<
std
::
string
,
RpcContext
>
send_varname_to_ctx_
;
std
::
unordered_map
<
std
::
string
,
RpcContext
>
recv_varname_to_ctx_
;
RpcCtxMap
send_varname_to_ctx_
;
RpcCtxMap
recv_varname_to_ctx_
;
std
::
unique_ptr
<
std
::
thread
>
send_thread_
;
std
::
unique_ptr
<
std
::
thread
>
recv_thread_
;
Scope
*
recv_scope_
;
// should be global scope
std
::
unique_ptr
<
Scope
>
send_scope_
;
// an independent scope
std
::
unique_ptr
<::
ThreadPool
>
send_threadpool_
{
nullptr
};
std
::
unique_ptr
<::
ThreadPool
>
recv_threadpool_
{
nullptr
};
// the following code is for initialize the commnunicator
public:
static
void
Init
(
const
RpcCtxMap
&
send_varname_to_ctx
,
const
RpcCtxMap
&
recv_varname_to_ctx
,
Scope
*
recv_scope
)
{
InitImpl
(
send_varname_to_ctx
,
recv_varname_to_ctx
,
recv_scope
);
}
static
Communicator
*
GetInstance
()
{
return
communicator_
.
get
();
}
private:
// Init is called by GetInstance.
static
void
InitImpl
(
const
RpcCtxMap
&
send_varname_to_ctx
,
const
RpcCtxMap
&
recv_varname_to_ctx
,
Scope
*
recv_scope
)
{
if
(
communicator_
==
nullptr
)
{
communicator_
.
reset
(
new
Communicator
(
send_varname_to_ctx
,
recv_varname_to_ctx
,
recv_scope
));
}
}
private:
static
std
::
once_flag
init_flag_
;
static
std
::
unique_ptr
<
Communicator
>
communicator_
;
};
}
// namespace distributed
...
...
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