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
wmsofts
Paddle
提交
35ebe2ec
P
Paddle
项目概览
wmsofts
/
Paddle
与 Fork 源项目一致
Fork自
PaddlePaddle / Paddle
通知
1
Star
0
Fork
0
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
0
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
0
Issue
0
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
35ebe2ec
编写于
4月 18, 2018
作者:
Y
Yu Yang
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Clean MultiDevicesGraphBuilder
上级
c3c7b7bd
变更
2
显示空白变更内容
内联
并排
Showing
2 changed file
with
122 addition
and
94 deletion
+122
-94
paddle/fluid/framework/details/multi_devices_graph_builder.cc
...le/fluid/framework/details/multi_devices_graph_builder.cc
+108
-94
paddle/fluid/framework/details/multi_devices_graph_builder.h
paddle/fluid/framework/details/multi_devices_graph_builder.h
+14
-0
未找到文件。
paddle/fluid/framework/details/multi_devices_graph_builder.cc
浏览文件 @
35ebe2ec
...
...
@@ -89,88 +89,60 @@ std::unique_ptr<SSAGraph> MultiDevSSAGraphBuilder::Build(
bool
is_forwarding
=
true
;
for
(
auto
*
op
:
program
.
Block
(
0
).
AllOps
())
{
bool
change_forward
=
false
;
if
(
op
->
Type
()
==
"send"
)
{
// append send op if program is distributed trainer main program.
// always use the first device
CreateSendOp
(
&
result
,
*
op
);
}
else
if
(
IsScaleLossOp
(
*
op
))
{
CreateScaleLossGradOp
(
&
result
);
is_forwarding
=
false
;
}
else
{
CreateComputationalOps
(
&
result
,
*
op
);
if
(
!
is_forwarding
)
{
// FIXME(yy): Do not hard code like this
if
(
op
->
OutputArgumentNames
().
size
()
==
1
&&
op
->
OutputArgumentNames
()[
0
]
==
GradVarName
(
loss_var_name_
))
{
continue
;
// Drop fill 1. for backward coeff;
// Currently, we assume that once gradient is generated, it can be
// broadcast, and each gradient is only broadcast once. But there are no
// other cases, for example, we need to adjust the gradient according to
// the input when we get the gradient, which is not considered at
// present.
for
(
auto
&
og
:
op
->
OutputArgumentNames
())
{
if
(
IsParameterGradientOnce
(
og
,
&
og_has_been_broadcast
))
{
InsertNCCLAllReduceOp
(
&
result
,
og
);
}
}
}
}
// append send op if program is distributed trainer main program.
// always use the first device
if
(
!
is_forwarding
&&
op
->
Type
()
==
"send"
)
{
auto
&
p
=
places_
[
0
];
auto
*
s
=
local_scopes_
[
0
];
// FIXME(wuyi): send op always copy from GPU 0
result
.
ops_
.
emplace_back
(
new
SendOpHandle
(
*
op
,
s
,
p
));
// Create inputs for output on original place and no ssa output
// is created for send op.
CreateOpHandleIOs
(
&
result
,
*
op
,
p
,
0
);
continue
;
}
for
(
size_t
i
=
0
;
i
<
places_
.
size
();
++
i
)
{
auto
&
p
=
places_
[
i
];
auto
*
s
=
local_scopes_
[
i
];
result
.
ops_
.
emplace_back
(
new
ComputationOpHandle
(
*
op
,
s
,
p
));
auto
*
op_handle
=
result
.
ops_
.
back
().
get
();
CreateOpHandleIOs
(
&
result
,
*
op
,
p
,
i
);
auto
var_names
=
op
->
OutputArgumentNames
();
if
(
is_forwarding
)
{
if
(
var_names
.
size
()
==
1
&&
var_names
[
0
]
==
loss_var_name_
)
{
// Insert ScaleCost OpHandle
#ifdef PADDLE_WITH_CUDA
auto
*
communication_dev_ctx
=
nccl_ctxs_
->
DevCtx
(
p
);
#else
auto
*
communication_dev_ctx
=
platform
::
DeviceContextPool
::
Instance
().
Get
(
platform
::
CPUPlace
());
#endif
op_handle
=
new
ScaleLossGradOpHandle
(
local_scopes_
.
size
(),
s
,
p
,
communication_dev_ctx
);
result
.
ops_
.
emplace_back
(
op_handle
);
/*
Dependency graph has been constructed. However, there are still data
harzaeds need to be handled.
*/
PolishGraphToSupportDataHazards
(
&
result
);
// FIXME: Currently ScaleLossGradOp only use device_count as scale
// factor. So it does not depend on any other operators.
// VarHandle *loss = GetVarHandle(loss_var_name, place);
// loss->pending_ops_.emplace_back(op_handle);
// op_handle->inputs_.emplace_back(loss);
/*
* Only variables should be the leaves of graph.
*/
AddOutputToLeafOps
(
&
result
);
CreateOpOutput
(
&
result
,
op_handle
,
GradVarName
(
loss_var_name_
),
p
,
i
);
change_forward
=
true
;
}
}
if
(
VLOG_IS_ON
(
10
))
{
std
::
ostringstream
sout
;
PrintGraphviz
(
*
graph
,
sout
);
VLOG
(
10
)
<<
sout
.
str
();
}
if
(
change_forward
)
{
is_forwarding
=
false
;
}
return
std
::
unique_ptr
<
SSAGraph
>
(
graph
);
}
if
(
!
is_forwarding
)
{
auto
var_names
=
op
->
OutputArgumentNames
();
// Currently, we assume that once gradient is generated, it can be
// broadcast, and each gradient is only broadcast once. But there are no
// other cases, for example, we need to adjust the gradient according to
// the input when we get the gradient, which is not considered at present.
for
(
auto
&
og
:
var_names
)
{
if
(
grad_names_
.
count
(
og
)
!=
0
&&
og_has_been_broadcast
.
count
(
og
)
==
0
)
{
// is param grad
// Insert NCCL AllReduce Op
og_has_been_broadcast
.
insert
(
og
);
void
MultiDevSSAGraphBuilder
::
InsertNCCLAllReduceOp
(
SSAGraph
*
result
,
const
std
::
string
&
og
)
const
{
#ifdef PADDLE_WITH_CUDA
result
.
ops_
.
emplace_back
(
result
->
ops_
.
emplace_back
(
new
NCCLAllReduceOpHandle
(
local_scopes_
,
places_
,
*
nccl_ctxs_
));
auto
*
op_handle
=
result
.
ops_
.
back
().
get
();
auto
*
op_handle
=
result
->
ops_
.
back
().
get
();
for
(
size_t
i
=
0
;
i
<
places_
.
size
();
++
i
)
{
auto
&
p
=
places_
[
i
];
auto
&
vars
=
result
.
vars_
[
i
][
og
];
auto
&
vars
=
result
->
vars_
[
i
][
og
];
if
(
vars
.
empty
())
{
// This device has no data. continue.
continue
;
}
...
...
@@ -184,30 +156,72 @@ std::unique_ptr<SSAGraph> MultiDevSSAGraphBuilder::Build(
#else
PADDLE_ENFORCE
(
"Not implemented"
);
#endif
}
bool
MultiDevSSAGraphBuilder
::
IsParameterGradientOnce
(
const
std
::
string
&
og
,
std
::
unordered_set
<
std
::
string
>
*
og_has_been_broadcast
)
const
{
bool
is_pg_once
=
grad_names_
.
count
(
og
)
!=
0
&&
og_has_been_broadcast
->
count
(
og
)
==
0
;
if
(
is_pg_once
)
{
// Insert NCCL AllReduce Op
og_has_been_broadcast
->
insert
(
og
);
}
}
}
}
return
is_pg_once
;
}
/*
Dependency graph has been constructed. However, there are still data
harzaeds need to be handled.
*/
PolishGraphToSupportDataHazards
(
&
result
);
void
MultiDevSSAGraphBuilder
::
CreateScaleLossGradOp
(
SSAGraph
*
result
)
const
{
for
(
size_t
i
=
0
;
i
<
places_
.
size
();
++
i
)
{
// Insert ScaleCost OpHandle
#ifdef PADDLE_WITH_CUDA
auto
*
communication_dev_ctx
=
nccl_ctxs_
->
DevCtx
(
places_
[
i
]);
#else
auto
*
communication_dev_ctx
=
platform
::
DeviceContextPool
::
Instance
().
Get
(
platform
::
CPUPlace
());
#endif
/*
* Only variables should be the leaves of graph.
*/
AddOutputToLeafOps
(
&
result
);
auto
*
op_handle
=
new
ScaleLossGradOpHandle
(
local_scopes_
.
size
(),
local_scopes_
[
i
],
places_
[
i
],
communication_dev_ctx
);
result
->
ops_
.
emplace_back
(
op_handle
);
if
(
VLOG_IS_ON
(
10
))
{
std
::
ostringstream
sout
;
PrintGraphviz
(
*
graph
,
sout
);
VLOG
(
10
)
<<
sout
.
str
();
// FIXME: Currently ScaleLossGradOp only use device_count as scale
// factor. So it does not depend on any other operators.
// VarHandle *loss = GetVarHandle(loss_var_name, place);
// loss->pending_ops_.emplace_back(op_handle);
// op_handle->inputs_.emplace_back(loss);
CreateOpOutput
(
result
,
op_handle
,
GradVarName
(
loss_var_name_
),
places_
[
i
],
i
);
}
}
return
std
::
unique_ptr
<
SSAGraph
>
(
graph
);
}
// namespace details
void
MultiDevSSAGraphBuilder
::
CreateComputationalOps
(
SSAGraph
*
result
,
const
OpDesc
&
op
)
const
{
for
(
size_t
scope_idx
=
0
;
scope_idx
<
places_
.
size
();
++
scope_idx
)
{
auto
p
=
places_
[
scope_idx
];
auto
s
=
local_scopes_
[
scope_idx
];
result
->
ops_
.
emplace_back
(
new
ComputationOpHandle
(
op
,
s
,
p
));
CreateOpHandleIOs
(
result
,
op
,
p
,
scope_idx
);
}
}
void
MultiDevSSAGraphBuilder
::
CreateSendOp
(
SSAGraph
*
result
,
const
OpDesc
&
op
)
const
{
auto
&
p
=
places_
[
0
];
auto
*
s
=
local_scopes_
[
0
];
// FIXME(wuyi): send op always copy from GPU 0
result
->
ops_
.
emplace_back
(
new
SendOpHandle
(
op
,
s
,
p
));
// Create inputs for output on original place and no ssa output
// is created for send op.
CreateOpHandleIOs
(
result
,
op
,
p
,
0
);
}
bool
MultiDevSSAGraphBuilder
::
IsScaleLossOp
(
const
OpDesc
&
op
)
const
{
// FIXME(yy): Do not hard code like this
return
op
.
OutputArgumentNames
().
size
()
==
1
&&
op
.
OutputArgumentNames
()[
0
]
==
GradVarName
(
loss_var_name_
);
}
}
// namespace details
}
// namespace framework
}
// namespace paddle
paddle/fluid/framework/details/multi_devices_graph_builder.h
浏览文件 @
35ebe2ec
...
...
@@ -57,6 +57,20 @@ class MultiDevSSAGraphBuilder : public SSAGraphBuilder {
#ifdef PADDLE_WITH_CUDA
platform
::
NCCLContextMap
*
nccl_ctxs_
;
#endif
bool
IsScaleLossOp
(
const
OpDesc
&
op
)
const
;
void
CreateSendOp
(
SSAGraph
*
result
,
const
OpDesc
&
op
)
const
;
void
CreateComputationalOps
(
SSAGraph
*
result
,
const
OpDesc
&
op
)
const
;
void
CreateScaleLossGradOp
(
SSAGraph
*
result
)
const
;
bool
IsParameterGradientOnce
(
const
std
::
string
&
og
,
std
::
unordered_set
<
std
::
string
>
*
og_has_been_broadcast
)
const
;
void
InsertNCCLAllReduceOp
(
SSAGraph
*
result
,
const
std
::
string
&
og
)
const
;
};
}
// namespace details
}
// namespace framework
...
...
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