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
兽拳
Paddle
提交
ce08dc87
P
Paddle
项目概览
兽拳
/
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 >>
提交
ce08dc87
编写于
4月 10, 2018
作者:
T
typhoonzero
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
have stream removed error
上级
0bf799a5
变更
6
显示空白变更内容
内联
并排
Showing
6 changed file
with
24 addition
and
31 deletion
+24
-31
paddle/fluid/framework/details/multi_devices_graph_builder.cc
...le/fluid/framework/details/multi_devices_graph_builder.cc
+14
-20
paddle/fluid/framework/details/multi_devices_graph_builder.h
paddle/fluid/framework/details/multi_devices_graph_builder.h
+1
-1
paddle/fluid/framework/details/send_op_handle.cc
paddle/fluid/framework/details/send_op_handle.cc
+4
-6
paddle/fluid/framework/details/send_op_handle.h
paddle/fluid/framework/details/send_op_handle.h
+1
-3
python/paddle/fluid/distribute_transpiler.py
python/paddle/fluid/distribute_transpiler.py
+1
-0
python/paddle/fluid/parallel_executor.py
python/paddle/fluid/parallel_executor.py
+3
-1
未找到文件。
paddle/fluid/framework/details/multi_devices_graph_builder.cc
浏览文件 @
ce08dc87
...
...
@@ -57,8 +57,11 @@ MultiDevSSAGraphBuilder::MultiDevSSAGraphBuilder(
void
MultiDevSSAGraphBuilder
::
CreateOpHandleIOs
(
SSAGraph
*
result
,
OpDesc
*
op
,
const
platform
::
Place
&
p
,
const
size_t
&
i
)
const
{
const
size_t
&
i
,
bool
create_output
)
const
{
auto
*
op_handle
=
result
->
ops_
.
back
().
get
();
op_handle
->
dev_ctxes_
[
p
]
=
const_cast
<
platform
::
DeviceContext
*>
(
platform
::
DeviceContextPool
::
Instance
().
Get
(
p
));
auto
var_names
=
op
->
InputArgumentNames
();
...
...
@@ -66,11 +69,13 @@ void MultiDevSSAGraphBuilder::CreateOpHandleIOs(SSAGraph *result, OpDesc *op,
VarHandle
*
var
=
CreateOrGetLatestVarHandle
(
result
,
each_var_name
,
p
,
i
);
op_handle
->
AddInput
(
var
);
}
if
(
create_output
)
{
var_names
=
op
->
OutputArgumentNames
();
for
(
auto
&
each_var_name
:
var_names
)
{
CreateOpOutput
(
result
,
op_handle
,
each_var_name
,
p
,
i
);
}
}
}
std
::
unique_ptr
<
SSAGraph
>
MultiDevSSAGraphBuilder
::
Build
(
...
...
@@ -100,9 +105,11 @@ std::unique_ptr<SSAGraph> MultiDevSSAGraphBuilder::Build(
if
(
!
is_forwarding
&&
op
->
Type
()
==
"send"
)
{
auto
&
p
=
places_
[
0
];
auto
*
s
=
local_scopes_
[
0
];
size_t
i
=
0
;
result
.
ops_
.
emplace_back
(
new
SendOpHandle
(
*
op
,
s
,
p
));
CreateOpHandleIOs
(
&
result
,
op
,
p
,
i
);
// FIXME(wuyi): send op always copy from GPU 0
result
.
ops_
.
emplace_back
(
new
SendOpHandle
(
*
op
,
s
));
// Create inputs for output on original place and no ssa output
// is created for send op.
CreateOpHandleIOs
(
&
result
,
op
,
p
,
0
,
false
);
continue
;
}
...
...
@@ -112,23 +119,10 @@ std::unique_ptr<SSAGraph> MultiDevSSAGraphBuilder::Build(
result
.
ops_
.
emplace_back
(
new
ComputationOpHandle
(
*
op
,
s
,
p
));
auto
*
op_handle
=
result
.
ops_
.
back
().
get
();
op_handle
->
dev_ctxes_
[
p
]
=
const_cast
<
platform
::
DeviceContext
*>
(
platform
::
DeviceContextPool
::
Instance
().
Get
(
p
));
CreateOpHandleIOs
(
&
result
,
op
,
p
,
i
);
// auto var_names = op->InputArgumentNames();
// for (auto &each_var_name : var_names) {
// VarHandle *var =
// CreateOrGetLatestVarHandle(&result, each_var_name, p, i);
// op_handle->AddInput(var);
// }
auto
var_names
=
op
->
OutputArgumentNames
();
// for (auto &each_var_name : var_names) {
// CreateOpOutput(&result, op_handle, each_var_name, p, i);
// }
if
(
is_forwarding
)
{
if
(
var_names
.
size
()
==
1
&&
var_names
[
0
]
==
loss_var_name_
)
{
// Insert ScaleCost OpHandle
...
...
paddle/fluid/framework/details/multi_devices_graph_builder.h
浏览文件 @
ce08dc87
...
...
@@ -46,7 +46,7 @@ class MultiDevSSAGraphBuilder : public SSAGraphBuilder {
private:
void
CreateOpHandleIOs
(
SSAGraph
*
result
,
OpDesc
*
op
,
const
platform
::
Place
&
p
,
const
size_t
&
i
)
const
;
const
size_t
&
i
,
bool
create_output
=
true
)
const
;
private:
std
::
string
loss_var_name_
;
...
...
paddle/fluid/framework/details/send_op_handle.cc
浏览文件 @
ce08dc87
...
...
@@ -19,11 +19,9 @@ namespace framework {
namespace
details
{
SendOpHandle
::
SendOpHandle
(
const
framework
::
OpDesc
&
op_desc
,
const
Scope
*
local_scope
,
const
platform
::
Place
&
place
)
const
Scope
*
local_scope
)
:
op_
(
framework
::
OpRegistry
::
CreateOp
(
op_desc
)),
local_scope_
(
local_scope
),
place_
(
place
)
{}
local_scope_
(
local_scope
)
{}
void
SendOpHandle
::
RunImpl
()
{
// Wait input done
...
...
@@ -31,8 +29,8 @@ void SendOpHandle::RunImpl() {
auto
&
p
=
static_cast
<
VarHandle
*>
(
in
)
->
place_
;
in
->
generated_op_
->
Wait
(
dev_ctxes_
[
p
]);
}
op_
->
Run
(
*
local_scope_
,
place_
);
platform
::
CPUPlace
cpu
;
op_
->
Run
(
*
local_scope_
,
cpu
);
}
std
::
string
SendOpHandle
::
Name
()
const
{
return
"send"
;
}
...
...
paddle/fluid/framework/details/send_op_handle.h
浏览文件 @
ce08dc87
...
...
@@ -31,10 +31,8 @@ namespace details {
struct
SendOpHandle
:
public
OpHandleBase
{
std
::
unique_ptr
<
OperatorBase
>
op_
;
const
Scope
*
local_scope_
;
const
platform
::
Place
&
place_
;
SendOpHandle
(
const
framework
::
OpDesc
&
op_desc
,
const
Scope
*
local_scope
,
const
platform
::
Place
&
place
);
SendOpHandle
(
const
framework
::
OpDesc
&
op_desc
,
const
Scope
*
local_scope
);
std
::
string
Name
()
const
override
;
...
...
python/paddle/fluid/distribute_transpiler.py
浏览文件 @
ce08dc87
...
...
@@ -255,6 +255,7 @@ class DistributeTranspiler:
def
get_trainer_program
(
self
):
# remove optimize ops and add a send op to main_program
self
.
program
.
global_block
().
delete_ops
(
self
.
optimize_ops
)
self
.
program
.
sync_with_cpp
()
# FIXME(typhoonzero): serialize once will fix error occurs when clone.
self
.
program
.
__str__
()
return
self
.
program
...
...
python/paddle/fluid/parallel_executor.py
浏览文件 @
ce08dc87
...
...
@@ -101,7 +101,9 @@ class ParallelExecutor(object):
self
.
persistable_vars
=
[
v
.
name
for
v
in
filter
(
lambda
var
:
var
.
persistable
,
main
.
list_vars
())
for
v
in
filter
(
lambda
var
:
\
var
.
persistable
and
var
.
type
!=
core
.
VarDesc
.
VarType
.
RAW
,
main
.
list_vars
())
]
self
.
executor
=
core
.
ParallelExecutor
(
...
...
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