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
TonyTonyFun
Paddle
提交
b75a69ba
P
Paddle
项目概览
TonyTonyFun
/
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 >>
未验证
提交
b75a69ba
编写于
4月 02, 2019
作者:
C
chengduo
提交者:
GitHub
4月 02, 2019
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Add Stream for fetch op handle (#16600)
* expose fuse broadcast ops
上级
1342e2ea
变更
4
隐藏空白更改
内联
并排
Showing
4 changed file
with
23 addition
and
5 deletion
+23
-5
paddle/fluid/framework/details/build_strategy.h
paddle/fluid/framework/details/build_strategy.h
+5
-1
paddle/fluid/framework/details/fetch_op_handle.cc
paddle/fluid/framework/details/fetch_op_handle.cc
+2
-1
paddle/fluid/framework/details/multi_devices_graph_pass.cc
paddle/fluid/framework/details/multi_devices_graph_pass.cc
+2
-2
paddle/fluid/pybind/pybind.cc
paddle/fluid/pybind/pybind.cc
+14
-1
未找到文件。
paddle/fluid/framework/details/build_strategy.h
浏览文件 @
b75a69ba
...
@@ -91,7 +91,11 @@ struct BuildStrategy {
...
@@ -91,7 +91,11 @@ struct BuildStrategy {
bool
enable_sequential_execution_
{
false
};
bool
enable_sequential_execution_
{
false
};
bool
fuse_broadcast_op_
{
false
};
// NOTE(zcd): In reduce mode, fusing broadcast ops may make the program
// faster. Because fusing broadcast OP equals delaying the execution of all
// broadcast Ops, in this case, all nccl streams are used only for reduce
// operations for a period of time.
bool
fuse_broadcast_ops_
{
false
};
// FIXME(zcd): is_distribution_ is a temporary field, because in pserver mode,
// FIXME(zcd): is_distribution_ is a temporary field, because in pserver mode,
// num_trainers is 1, so the current fields of build_strategy doesn't tell if
// num_trainers is 1, so the current fields of build_strategy doesn't tell if
...
...
paddle/fluid/framework/details/fetch_op_handle.cc
浏览文件 @
b75a69ba
...
@@ -63,7 +63,8 @@ void FetchOpHandle::RunImpl() {
...
@@ -63,7 +63,8 @@ void FetchOpHandle::RunImpl() {
auto
&
t
=
var
->
Get
<
framework
::
LoDTensor
>
();
auto
&
t
=
var
->
Get
<
framework
::
LoDTensor
>
();
if
(
platform
::
is_gpu_place
(
t
.
place
()))
{
if
(
platform
::
is_gpu_place
(
t
.
place
()))
{
#ifdef PADDLE_WITH_CUDA
#ifdef PADDLE_WITH_CUDA
TensorCopySync
(
t
,
cpu
,
&
tensors_
[
i
]);
TensorCopy
(
t
,
cpu
,
*
dev_ctxes_
.
at
(
t
.
place
()),
&
tensors_
[
i
]);
dev_ctxes_
.
at
(
t
.
place
())
->
Wait
();
#endif
#endif
}
else
{
}
else
{
tensors_
[
i
].
ShareDataWith
(
t
);
tensors_
[
i
].
ShareDataWith
(
t
);
...
...
paddle/fluid/framework/details/multi_devices_graph_pass.cc
浏览文件 @
b75a69ba
...
@@ -658,7 +658,7 @@ bool ReduceSSAGraphBuilder::DealWithSpecialOp(ir::Graph *result,
...
@@ -658,7 +658,7 @@ bool ReduceSSAGraphBuilder::DealWithSpecialOp(ir::Graph *result,
void
ReduceSSAGraphBuilder
::
InsertPostprocessOps
(
ir
::
Graph
*
result
)
const
{
void
ReduceSSAGraphBuilder
::
InsertPostprocessOps
(
ir
::
Graph
*
result
)
const
{
if
(
UseGPU
())
{
if
(
UseGPU
())
{
if
(
strategy_
.
fuse_broadcast_op_
)
{
if
(
strategy_
.
fuse_broadcast_op
s
_
)
{
CreateFusedBroadcastOp
(
result
,
bcast_var_name_set_
);
CreateFusedBroadcastOp
(
result
,
bcast_var_name_set_
);
}
else
{
}
else
{
for
(
size_t
dev_id
=
0
;
dev_id
<
bcast_var_name_set_
.
size
();
++
dev_id
)
{
for
(
size_t
dev_id
=
0
;
dev_id
<
bcast_var_name_set_
.
size
();
++
dev_id
)
{
...
@@ -1021,7 +1021,7 @@ void DistSSAGraphBuilder::InsertPostprocessOps(ir::Graph *result) const {
...
@@ -1021,7 +1021,7 @@ void DistSSAGraphBuilder::InsertPostprocessOps(ir::Graph *result) const {
strategy_
.
reduce_
==
BuildStrategy
::
ReduceStrategy
::
kReduce
)
{
strategy_
.
reduce_
==
BuildStrategy
::
ReduceStrategy
::
kReduce
)
{
return
;
return
;
}
}
if
(
strategy_
.
fuse_broadcast_op_
)
{
if
(
strategy_
.
fuse_broadcast_op
s
_
)
{
CreateFusedBroadcastOp
(
result
,
bcast_var_name_set_
);
CreateFusedBroadcastOp
(
result
,
bcast_var_name_set_
);
}
else
{
}
else
{
for
(
size_t
dev_id
=
0
;
dev_id
<
bcast_var_name_set_
.
size
();
++
dev_id
)
{
for
(
size_t
dev_id
=
0
;
dev_id
<
bcast_var_name_set_
.
size
();
++
dev_id
)
{
...
...
paddle/fluid/pybind/pybind.cc
浏览文件 @
b75a69ba
...
@@ -1299,7 +1299,20 @@ All parameter, weight, gradient are variables in Paddle.
...
@@ -1299,7 +1299,20 @@ All parameter, weight, gradient are variables in Paddle.
to fuse relu and depthwise_conv2d,
to fuse relu and depthwise_conv2d,
it will save GPU memory and may make the execution faster.
it will save GPU memory and may make the execution faster.
This options is only available in GPU devices.
This options is only available in GPU devices.
Default False)DOC"
)
Default False.)DOC"
)
.
def_property
(
"fuse_broadcast_ops"
,
[](
const
BuildStrategy
&
self
)
{
return
self
.
fuse_broadcast_ops_
;
},
[](
BuildStrategy
&
self
,
bool
b
)
{
PADDLE_ENFORCE
(
!
self
.
IsFinalized
(),
"BuildStrategy is finlaized."
);
self
.
fuse_broadcast_ops_
=
b
;
},
R"DOC(The type is BOOL, fuse_broadcast_op indicates whether
to fuse the broadcast ops. Note that, in Reduce mode,
fusing broadcast ops may make the program faster. Because
fusing broadcast OP equals delaying the execution of all
broadcast Ops, in this case, all nccl streams are used only
for NCCLReduce operations for a period of time. Default False.)DOC"
)
.
def_property
(
"fuse_all_optimizer_ops"
,
.
def_property
(
"fuse_all_optimizer_ops"
,
[](
const
BuildStrategy
&
self
)
{
[](
const
BuildStrategy
&
self
)
{
return
self
.
fuse_all_optimizer_ops_
;
return
self
.
fuse_all_optimizer_ops_
;
...
...
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