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
机器未来
Paddle
提交
fb8c1cf0
P
Paddle
项目概览
机器未来
/
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
1
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
1
Issue
1
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 >>
未验证
提交
fb8c1cf0
编写于
4月 04, 2018
作者:
W
Wu Yi
提交者:
GitHub
4月 04, 2018
浏览文件
操作
浏览文件
下载
差异文件
Merge pull request #9377 from typhoonzero/prepare_pserver_executor
prepare pserver executor
上级
66e0aed7
1f6e0448
变更
3
隐藏空白更改
内联
并排
Showing
3 changed file
with
45 addition
and
15 deletion
+45
-15
paddle/fluid/framework/executor.cc
paddle/fluid/framework/executor.cc
+15
-0
paddle/fluid/framework/executor.h
paddle/fluid/framework/executor.h
+3
-0
paddle/fluid/operators/listen_and_serv_op.cc
paddle/fluid/operators/listen_and_serv_op.cc
+27
-15
未找到文件。
paddle/fluid/framework/executor.cc
浏览文件 @
fb8c1cf0
...
...
@@ -279,6 +279,21 @@ std::unique_ptr<ExecutorPrepareContext> Executor::Prepare(
return
std
::
unique_ptr
<
ExecutorPrepareContext
>
(
ctx
);
}
std
::
vector
<
std
::
shared_ptr
<
ExecutorPrepareContext
>>
Executor
::
Prepare
(
const
ProgramDesc
&
program
,
const
std
::
vector
<
int
>&
block_ids
)
{
std
::
vector
<
std
::
shared_ptr
<
ExecutorPrepareContext
>>
result
;
for
(
auto
&
bid
:
block_ids
)
{
auto
*
ctx
=
new
ExecutorPrepareContext
(
program
,
bid
);
PADDLE_ENFORCE_LT
(
static_cast
<
size_t
>
(
bid
),
program
.
Size
());
auto
&
block
=
program
.
Block
(
bid
);
for
(
auto
&
op_desc
:
block
.
AllOps
())
{
ctx
->
ops_
.
push_back
(
OpRegistry
::
CreateOp
(
*
op_desc
));
}
result
.
push_back
(
std
::
shared_ptr
<
ExecutorPrepareContext
>
(
ctx
));
}
return
result
;
}
void
Executor
::
RunPreparedContext
(
ExecutorPrepareContext
*
ctx
,
Scope
*
scope
,
bool
create_local_scope
,
bool
create_vars
)
{
auto
&
block
=
ctx
->
prog_
.
Block
(
ctx
->
block_id_
);
...
...
paddle/fluid/framework/executor.h
浏览文件 @
fb8c1cf0
...
...
@@ -61,6 +61,9 @@ class Executor {
static
std
::
unique_ptr
<
ExecutorPrepareContext
>
Prepare
(
const
ProgramDesc
&
program
,
int
block_id
);
static
std
::
vector
<
std
::
shared_ptr
<
ExecutorPrepareContext
>>
Prepare
(
const
ProgramDesc
&
program
,
const
std
::
vector
<
int
>&
block_ids
);
void
RunPreparedContext
(
ExecutorPrepareContext
*
ctx
,
Scope
*
scope
,
bool
create_local_scope
=
true
,
bool
create_vars
=
true
);
...
...
paddle/fluid/operators/listen_and_serv_op.cc
浏览文件 @
fb8c1cf0
...
...
@@ -45,20 +45,23 @@ static void CreateTensorFromMessageType(framework::Variable *var,
}
}
static
void
ParallelExecuteBlocks
(
const
std
::
vector
<
size_t
>
&
parallel_blkids
,
framework
::
Executor
*
executor
,
framework
::
ProgramDesc
*
program
,
framework
::
Scope
*
scope
)
{
static
void
ParallelExecuteBlocks
(
const
std
::
vector
<
size_t
>
&
parallel_blkids
,
framework
::
Executor
*
executor
,
const
std
::
vector
<
std
::
shared_ptr
<
framework
::
ExecutorPrepareContext
>>
&
prepared
,
framework
::
ProgramDesc
*
program
,
framework
::
Scope
*
scope
)
{
std
::
vector
<
std
::
future
<
void
>>
fs
;
for
(
size_t
idx
:
parallel_blkids
)
{
fs
.
push_back
(
framework
::
Async
([
&
executor
,
&
program
,
&
scope
,
idx
]()
{
int
run_block
=
idx
;
// thread local
try
{
executor
->
Run
(
*
program
,
scope
,
run_block
,
false
,
false
);
}
catch
(
std
::
exception
&
e
)
{
LOG
(
ERROR
)
<<
"run sub program error "
<<
e
.
what
();
}
}));
fs
.
push_back
(
framework
::
Async
([
&
executor
,
&
prepared
,
&
program
,
&
scope
,
idx
]()
{
int
run_block
=
idx
;
// thread local
try
{
executor
->
RunPreparedContext
(
prepared
[
run_block
].
get
(),
scope
,
false
,
false
);
}
catch
(
std
::
exception
&
e
)
{
LOG
(
ERROR
)
<<
"run sub program error "
<<
e
.
what
();
}
}));
}
for
(
size_t
i
=
0
;
i
<
fs
.
size
();
++
i
)
fs
[
i
].
wait
();
}
...
...
@@ -101,6 +104,13 @@ class ListenAndServOp : public framework::OperatorBase {
"server program should have at least 2 blocks"
);
framework
::
Executor
executor
(
dev_place
);
std
::
vector
<
int
>
block_list
;
for
(
size_t
blkid
=
1
;
blkid
<
num_blocks
;
++
blkid
)
block_list
.
push_back
(
blkid
);
auto
prepared
=
executor
.
Prepare
(
*
program
,
block_list
);
prepared
.
insert
(
prepared
.
begin
(),
std
::
shared_ptr
<
framework
::
ExecutorPrepareContext
>
(
nullptr
));
// TODO(qiao) set proper fields for table lookup and update
rpc_service_
->
SetExecutor
(
&
executor
);
...
...
@@ -160,14 +170,15 @@ class ListenAndServOp : public framework::OperatorBase {
for
(
size_t
blkid
=
2
;
blkid
<
num_blocks
;
++
blkid
)
{
if
(
program
->
Block
(
blkid
).
Parent
()
!=
last_parent_blkid
)
{
for
(
size_t
idx
:
parallel_blkids
)
VLOG
(
3
)
<<
idx
;
ParallelExecuteBlocks
(
parallel_blkids
,
&
executor
,
program
,
ParallelExecuteBlocks
(
parallel_blkids
,
&
executor
,
pr
epared
,
pr
ogram
,
&
recv_scope
);
parallel_blkids
.
clear
();
last_parent_blkid
=
program
->
Block
(
blkid
).
Parent
();
}
parallel_blkids
.
push_back
(
blkid
);
}
ParallelExecuteBlocks
(
parallel_blkids
,
&
executor
,
program
,
&
recv_scope
);
ParallelExecuteBlocks
(
parallel_blkids
,
&
executor
,
prepared
,
program
,
&
recv_scope
);
VLOG
(
3
)
<<
"run all blocks spent "
<<
detail
::
GetTimestamp
()
-
ts
<<
"(ms)"
;
...
...
@@ -181,7 +192,8 @@ class ListenAndServOp : public framework::OperatorBase {
var
->
GetMutable
<
framework
::
SelectedRows
>
()
->
mutable_rows
()
->
clear
();
}
rpc_service_
->
SetCond
(
1
);
// FIXME(typhoonzero): use another condition to sync wait clients get.
// NOTE: does not consider barrier request retry in here, we may use
// global barrier id to resolve this.
rpc_service_
->
WaitClientGet
(
fan_in
);
sparse_vars
.
clear
();
}
// while(true)
...
...
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