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
956cf921
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
956cf921
编写于
12月 26, 2018
作者:
H
hjchen2
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Fix conv_elementwise_add2_act pass
test=develop
上级
05d1121b
变更
3
显示空白变更内容
内联
并排
Showing
3 changed file
with
21 addition
and
20 deletion
+21
-20
paddle/fluid/framework/ir/conv_elementwise_add2_act_fuse_pass.cc
...fluid/framework/ir/conv_elementwise_add2_act_fuse_pass.cc
+14
-11
paddle/fluid/framework/ir/graph_pattern_detector.cc
paddle/fluid/framework/ir/graph_pattern_detector.cc
+5
-7
paddle/fluid/operators/conv_fusion_op.cu.cc
paddle/fluid/operators/conv_fusion_op.cu.cc
+2
-2
未找到文件。
paddle/fluid/framework/ir/conv_elementwise_add2_act_fuse_pass.cc
浏览文件 @
956cf921
...
@@ -40,18 +40,20 @@ framework::proto::OpDesc PrepareOpDesc(
...
@@ -40,18 +40,20 @@ framework::proto::OpDesc PrepareOpDesc(
const
std
::
string
&
output
)
{
const
std
::
string
&
output
)
{
auto
proto
=
base_desc
;
auto
proto
=
base_desc
;
framework
::
OpDesc
desc
(
proto
,
nullptr
);
framework
::
OpDesc
desc
(
proto
,
nullptr
);
desc
.
SetType
(
"conv2d_fusion"
);
desc
.
SetInput
(
"Bias"
,
{
bias
});
desc
.
SetInput
(
"Bias"
,
{
bias
});
desc
.
SetInput
(
"ResidualData"
,
{
bias1
});
desc
.
SetInput
(
"ResidualData"
,
{
bias1
});
desc
.
SetAttr
(
"activation"
,
activation
);
desc
.
SetAttr
(
"activation"
,
activation
);
desc
.
SetOutput
(
"Output"
,
{
output
});
desc
.
SetOutput
(
"Output"
,
{
output
});
desc
.
SetAttr
(
"is_test"
,
true
);
desc
.
SetAttr
(
"is_test"
,
true
);
desc
.
SetAttr
(
"use_cudnn"
,
false
);
desc
.
Flush
();
return
*
desc
.
Proto
();
return
*
desc
.
Proto
();
}
}
std
::
unique_ptr
<
ir
::
Graph
>
ConvElementwiseAdd2ActFusePass
::
ApplyImpl
(
std
::
unique_ptr
<
ir
::
Graph
>
ConvElementwiseAdd2ActFusePass
::
ApplyImpl
(
std
::
unique_ptr
<
ir
::
Graph
>
graph
)
const
{
std
::
unique_ptr
<
ir
::
Graph
>
graph
)
const
{
const
std
::
string
pattern_name
=
"conv_elementwise_add_act_fuse"
;
const
std
::
string
pattern_name
=
"conv_elementwise_add
2
_act_fuse"
;
FusePassBase
::
Init
(
pattern_name
,
graph
.
get
());
FusePassBase
::
Init
(
pattern_name
,
graph
.
get
());
GraphPatternDetector
gpd
;
GraphPatternDetector
gpd
;
...
@@ -76,22 +78,23 @@ std::unique_ptr<ir::Graph> ConvElementwiseAdd2ActFusePass::ApplyImpl(
...
@@ -76,22 +78,23 @@ std::unique_ptr<ir::Graph> ConvElementwiseAdd2ActFusePass::ApplyImpl(
framework
::
OpDesc
new_op_desc
(
new_op_proto
,
nullptr
);
framework
::
OpDesc
new_op_desc
(
new_op_proto
,
nullptr
);
// Create a new node for the fused op.
// Create a new node for the fused op.
graph
->
CreateOpNode
(
&
new_op_desc
);
auto
*
new_conv_op
=
graph
->
CreateOpNode
(
&
new_op_desc
);
// Link inputs and outputs.
// Link inputs and outputs.
PADDLE_ENFORCE
(
subgraph
.
count
(
x
));
PADDLE_ENFORCE
(
subgraph
.
count
(
x
));
auto
*
conv_in_node
=
subgraph
.
at
(
x
);
auto
*
conv_in_node
=
subgraph
.
at
(
x
);
IR_NODE_LINK_TO
(
conv_in_node
,
conv_op
);
// Input
IR_NODE_LINK_TO
(
conv_in_node
,
new_
conv_op
);
// Input
IR_NODE_LINK_TO
(
conv_filter
,
conv_op
);
// Filter
IR_NODE_LINK_TO
(
conv_filter
,
new_
conv_op
);
// Filter
IR_NODE_LINK_TO
(
conv_op
,
conv_out
);
// Output
IR_NODE_LINK_TO
(
elementwise_add_in_y
,
new_conv_op
);
// Bias
IR_NODE_LINK_TO
(
elementwise_add_in_y
,
conv_op
);
// Bias
IR_NODE_LINK_TO
(
elementwise_add_in_y
_1
,
new_conv_op
);
// Bias
IR_NODE_LINK_TO
(
elementwise_add_in_y_1
,
conv_op
);
// Bias
IR_NODE_LINK_TO
(
new_conv_op
,
act_out
);
// Output
// Delete the unneeded nodes.
// Delete the unneeded nodes.
GraphSafeRemoveNodes
(
graph
.
get
(),
GraphSafeRemoveNodes
(
{
conv_op
,
elementwise_add_op
,
elementwise_add_op_1
,
graph
.
get
(),
elementwise_add_out
});
{
conv_op
,
conv_out
,
elementwise_add_op
,
elementwise_add_op_1
,
elementwise_add_out
,
elementwise_add_out_1
,
act_op
});
};
};
gpd
(
graph
.
get
(),
handler
);
gpd
(
graph
.
get
(),
handler
);
return
graph
;
return
graph
;
...
...
paddle/fluid/framework/ir/graph_pattern_detector.cc
浏览文件 @
956cf921
...
@@ -1101,9 +1101,7 @@ PDNode *patterns::ElementwiseAdd::operator()(PDNode *x_var, PDNode *y_var) {
...
@@ -1101,9 +1101,7 @@ PDNode *patterns::ElementwiseAdd::operator()(PDNode *x_var, PDNode *y_var) {
return
out_var
;
return
out_var
;
}
}
std
::
unordered_set
<
std
::
string
>
conv_act_set
({
"identity"
,
"sigmoid"
,
"relu"
,
std
::
unordered_set
<
std
::
string
>
conv_act_set
({
"identity"
,
"relu"
});
"relu6"
,
"relux"
,
"tanh"
,
"band_pass"
});
PDNode
*
patterns
::
ConvElementwiseaddAct
::
operator
()(
PDNode
*
conv_in
)
{
PDNode
*
patterns
::
ConvElementwiseaddAct
::
operator
()(
PDNode
*
conv_in
)
{
conv_in
->
AsInput
();
conv_in
->
AsInput
();
...
@@ -1169,13 +1167,13 @@ PDNode *patterns::ConvElementwiseadd2Act::operator()(PDNode *conv_in) {
...
@@ -1169,13 +1167,13 @@ PDNode *patterns::ConvElementwiseadd2Act::operator()(PDNode *conv_in) {
->
AsInput
();
->
AsInput
();
auto
elementwise_add_out
=
pattern
->
NewNode
(
elementwise_add_out_repr
())
auto
elementwise_add_out
=
pattern
->
NewNode
(
elementwise_add_out_repr
())
->
assert_is_op_output
(
"elementwise_add"
)
->
assert_is_op_output
(
"elementwise_add"
)
->
assert_is_op_input
(
"elementwise_add"
,
"
X
"
)
->
assert_is_op_input
(
"elementwise_add"
,
"
Y
"
)
->
AsIntermediate
();
->
AsIntermediate
();
auto
elementwise_add_op_1
=
pattern
->
NewNode
(
elementwise_add_op_1_repr
())
auto
elementwise_add_op_1
=
pattern
->
NewNode
(
elementwise_add_op_1_repr
())
->
assert_is_op
(
"elementwise_add"
);
->
assert_is_op
(
"elementwise_add"
);
auto
elementwise_add_in_y_1
=
pattern
->
NewNode
(
elementwise_add_in_y_1_repr
())
auto
elementwise_add_in_y_1
=
pattern
->
NewNode
(
elementwise_add_in_y_1_repr
())
->
assert_is_op_input
(
"elementwise_add"
,
"
Y
"
)
->
assert_is_op_input
(
"elementwise_add"
,
"
X
"
)
->
AsInput
();
->
AsInput
();
auto
elementwise_add_out_1
=
pattern
->
NewNode
(
elementwise_add_out_1_repr
())
auto
elementwise_add_out_1
=
pattern
->
NewNode
(
elementwise_add_out_1_repr
())
->
assert_is_op_output
(
"elementwise_add"
)
->
assert_is_op_output
(
"elementwise_add"
)
...
@@ -1203,8 +1201,8 @@ PDNode *patterns::ConvElementwiseadd2Act::operator()(PDNode *conv_in) {
...
@@ -1203,8 +1201,8 @@ PDNode *patterns::ConvElementwiseadd2Act::operator()(PDNode *conv_in) {
conv_op
->
LinksFrom
({
conv_in
,
conv_filter
}).
LinksTo
({
conv_out
});
conv_op
->
LinksFrom
({
conv_in
,
conv_filter
}).
LinksTo
({
conv_out
});
elementwise_add_op
->
LinksFrom
({
conv_out
,
elementwise_add_in_y
})
elementwise_add_op
->
LinksFrom
({
conv_out
,
elementwise_add_in_y
})
.
LinksTo
({
elementwise_add_out
});
.
LinksTo
({
elementwise_add_out
});
elementwise_add_op_1
->
LinksFrom
(
elementwise_add_op_1
->
LinksFrom
(
{
elementwise_add_out
,
elementwise_add_in_y_1
})
{
elementwise_add_out
,
elementwise_add_in_y
_1
});
.
LinksTo
({
elementwise_add_out
_1
});
act_op
->
LinksFrom
({
elementwise_add_out_1
}).
LinksTo
({
act_out
});
act_op
->
LinksFrom
({
elementwise_add_out_1
}).
LinksTo
({
act_out
});
return
act_out
;
return
act_out
;
}
}
...
...
paddle/fluid/operators/conv_fusion_op.cu.cc
浏览文件 @
956cf921
...
@@ -22,7 +22,7 @@ DECLARE_bool(cudnn_exhaustive_search);
...
@@ -22,7 +22,7 @@ DECLARE_bool(cudnn_exhaustive_search);
namespace
paddle
{
namespace
paddle
{
namespace
operators
{
namespace
operators
{
#if CUDNN_VERSION >= 7
001
#if CUDNN_VERSION >= 7
100
using
Tensor
=
framework
::
Tensor
;
using
Tensor
=
framework
::
Tensor
;
using
ScopedTensorDescriptor
=
platform
::
ScopedTensorDescriptor
;
using
ScopedTensorDescriptor
=
platform
::
ScopedTensorDescriptor
;
using
ScopedFilterDescriptor
=
platform
::
ScopedFilterDescriptor
;
using
ScopedFilterDescriptor
=
platform
::
ScopedFilterDescriptor
;
...
@@ -204,7 +204,7 @@ class CUDNNConvFusionOpKernel : public framework::OpKernel<T> {
...
@@ -204,7 +204,7 @@ class CUDNNConvFusionOpKernel : public framework::OpKernel<T> {
}
// namespace operators
}
// namespace operators
}
// namespace paddle
}
// namespace paddle
#if CUDNN_VERSION >= 7
001
#if CUDNN_VERSION >= 7
100
namespace
ops
=
paddle
::
operators
;
namespace
ops
=
paddle
::
operators
;
REGISTER_OP_CUDA_KERNEL
(
conv2d_fusion
,
ops
::
CUDNNConvFusionOpKernel
<
float
>
,
REGISTER_OP_CUDA_KERNEL
(
conv2d_fusion
,
ops
::
CUDNNConvFusionOpKernel
<
float
>
,
ops
::
CUDNNConvFusionOpKernel
<
double
>
);
ops
::
CUDNNConvFusionOpKernel
<
double
>
);
...
...
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