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
OPTHREE
Paddle
提交
72d2a180
P
Paddle
项目概览
OPTHREE
/
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
72d2a180
编写于
1月 07, 2019
作者:
T
tensor-tang
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
add seqpool concat fuse pass
test=develop
上级
f702f8fd
变更
5
隐藏空白更改
内联
并排
Showing
5 changed file
with
239 addition
and
1 deletion
+239
-1
paddle/fluid/framework/ir/CMakeLists.txt
paddle/fluid/framework/ir/CMakeLists.txt
+1
-0
paddle/fluid/framework/ir/seqpool_concat_fuse_pass.cc
paddle/fluid/framework/ir/seqpool_concat_fuse_pass.cc
+194
-0
paddle/fluid/framework/ir/seqpool_concat_fuse_pass.h
paddle/fluid/framework/ir/seqpool_concat_fuse_pass.h
+38
-0
paddle/fluid/inference/api/paddle_pass_builder.h
paddle/fluid/inference/api/paddle_pass_builder.h
+1
-0
paddle/fluid/inference/tests/api/analyzer_seq_pool1_tester.cc
...le/fluid/inference/tests/api/analyzer_seq_pool1_tester.cc
+5
-1
未找到文件。
paddle/fluid/framework/ir/CMakeLists.txt
浏览文件 @
72d2a180
...
...
@@ -42,6 +42,7 @@ pass_library(seq_concat_fc_fuse_pass inference)
pass_library
(
multi_batch_merge_pass base
)
pass_library
(
conv_bn_fuse_pass inference
)
pass_library
(
seqconv_eltadd_relu_fuse_pass inference
)
pass_library
(
seqpool_concat_fuse_pass inference
)
pass_library
(
is_test_pass base
)
pass_library
(
conv_elementwise_add_act_fuse_pass inference
)
pass_library
(
conv_elementwise_add2_act_fuse_pass inference
)
...
...
paddle/fluid/framework/ir/seqpool_concat_fuse_pass.cc
0 → 100644
浏览文件 @
72d2a180
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License. */
#include "paddle/fluid/framework/ir/seqpool_concat_fuse_pass.h"
#include <string>
#include <vector>
#include "paddle/fluid/framework/lod_tensor.h"
#define MAX_CONCAT_INPUTS 200
namespace
paddle
{
namespace
framework
{
namespace
ir
{
PDNode
*
BuildSeqPoolConcatPattern
(
PDPattern
*
pattern
,
const
std
::
string
&
name_scope
,
int
num_inputs
)
{
auto
is_concat_op_with_inputs
=
[](
Node
*
x
,
int
num
)
->
bool
{
return
x
&&
x
->
IsOp
()
&&
x
->
Op
()
->
Type
()
==
"concat"
&&
x
->
Op
()
->
Input
(
"X"
).
size
()
==
static_cast
<
size_t
>
(
num
);
};
auto
is_nth_input_var_of_concat
=
[
=
](
Node
*
x
,
int
idx
)
->
bool
{
return
x
&&
x
->
IsVar
()
&&
VarLinksToOp
(
x
,
"concat"
)
&&
x
->
outputs
.
size
()
==
1
&&
IsNthInput
(
x
,
x
->
outputs
[
0
],
"X"
,
idx
)
&&
is_concat_op_with_inputs
(
x
->
outputs
[
0
],
num_inputs
);
};
auto
is_seqpool_op_with_pootype_of_nth_input_of_concat
=
[
=
](
Node
*
x
,
const
std
::
string
&
type
,
int
idx
)
->
bool
{
bool
ok
=
x
&&
x
->
IsOp
()
&&
x
->
Op
()
->
Type
()
==
"sequence_pool"
&&
x
->
Op
()
->
HasAttr
(
"pooltype"
)
&&
boost
::
get
<
std
::
string
>
(
x
->
Op
()
->
GetAttr
(
"pooltype"
))
==
type
&&
x
->
outputs
.
size
()
==
2
;
// seqpool should only have 2 outputs
if
(
ok
)
{
// only one output of seqpool_op is nth_input_var of concat
// the other one should be unused empty var
if
(
is_nth_input_var_of_concat
(
x
->
outputs
[
0
],
idx
))
{
ok
=
ok
&&
x
->
outputs
[
1
]
->
IsVar
()
&&
x
->
outputs
[
1
]
->
outputs
.
size
()
==
0
;
}
else
{
ok
=
ok
&&
is_nth_input_var_of_concat
(
x
->
outputs
[
1
],
idx
)
&&
x
->
outputs
[
0
]
->
IsVar
()
&&
x
->
outputs
[
0
]
->
outputs
.
size
()
==
0
;
}
}
return
ok
;
};
auto
*
concat_op
=
pattern
->
NewNode
(
[
=
](
Node
*
x
)
{
return
is_concat_op_with_inputs
(
x
,
num_inputs
);
},
name_scope
+
"/concat_op"
);
concat_op
->
assert_op_attr
<
int
>
(
"axis"
,
1
);
auto
*
concat_out_var
=
pattern
->
NewNode
(
[
=
](
Node
*
x
)
{
return
x
&&
x
->
IsVar
()
&&
VarLinksFromOp
(
x
,
"concat"
)
&&
x
->
inputs
.
size
()
==
1
&&
is_concat_op_with_inputs
(
x
->
inputs
[
0
],
num_inputs
);
},
name_scope
+
"/concat_out_var"
);
concat_out_var
->
assert_is_only_output_of_op
(
"concat"
);
std
::
vector
<
PDNode
*>
seqpool_ops_input_var
(
num_inputs
);
std
::
vector
<
PDNode
*>
seqpool_ops_output_var
(
num_inputs
);
std
::
vector
<
PDNode
*>
seqpool_ops
(
num_inputs
);
for
(
int
i
=
0
;
i
<
num_inputs
;
++
i
)
{
seqpool_ops_output_var
[
i
]
=
pattern
->
NewNode
(
[
=
](
Node
*
x
)
{
return
x
&&
x
->
IsVar
()
&&
is_nth_input_var_of_concat
(
x
,
i
)
&&
x
->
inputs
.
size
()
==
1
&&
is_seqpool_op_with_pootype_of_nth_input_of_concat
(
x
->
inputs
[
0
],
"SUM"
,
i
);
},
name_scope
+
"/sequence_pool_out_"
+
std
::
to_string
(
i
));
seqpool_ops
[
i
]
=
pattern
->
NewNode
(
[
=
](
Node
*
x
)
{
return
x
&&
x
->
IsOp
()
&&
is_seqpool_op_with_pootype_of_nth_input_of_concat
(
x
,
"SUM"
,
i
);
},
name_scope
+
"/sequence_pool_op_"
+
std
::
to_string
(
i
));
seqpool_ops_input_var
[
i
]
=
pattern
->
NewNode
(
[
=
](
Node
*
x
)
{
return
x
&&
x
->
IsVar
()
&&
x
->
outputs
.
size
()
>=
1
&&
is_seqpool_op_with_pootype_of_nth_input_of_concat
(
x
->
outputs
[
0
],
"SUM"
,
i
);
},
name_scope
+
"/sequence_pool_in_"
+
std
::
to_string
(
i
));
// Links
seqpool_ops
[
i
]
->
LinksFrom
({
seqpool_ops_input_var
[
i
]})
.
LinksTo
({
seqpool_ops_output_var
[
i
]});
}
concat_op
->
LinksFrom
(
seqpool_ops_output_var
).
LinksTo
({
concat_out_var
});
return
concat_out_var
;
}
int
BuildFusion
(
Graph
*
graph
,
const
std
::
string
&
name_scope
,
Scope
*
scope
,
int
num_inputs
)
{
GraphPatternDetector
gpd
;
auto
*
pattern
=
gpd
.
mutable_pattern
();
BuildSeqPoolConcatPattern
(
pattern
,
name_scope
,
num_inputs
);
auto
retrieve_node
=
[](
const
std
::
string
&
name
,
const
GraphPatternDetector
::
subgraph_t
&
subgraph
,
const
PDPattern
&
pat
)
->
Node
*
{
PADDLE_ENFORCE
(
subgraph
.
count
(
pat
.
RetrieveNode
(
name
)),
"pattern has no Node called %s"
,
name
.
c_str
());
Node
*
p
=
subgraph
.
at
(
pat
.
RetrieveNode
(
name
));
PADDLE_ENFORCE_NOT_NULL
(
p
,
"subgraph has no node %s"
,
name
.
c_str
());
return
p
;
};
int
fusion_count
{
0
};
auto
handler
=
[
&
](
const
GraphPatternDetector
::
subgraph_t
&
subgraph
,
Graph
*
g
)
{
VLOG
(
4
)
<<
"handle SeqPool Concat fuse"
;
std
::
vector
<
std
::
string
>
input_names
(
num_inputs
);
std
::
vector
<
Node
*>
input_vars
(
num_inputs
);
auto
&
fused_pattern
=
gpd
.
pattern
();
for
(
int
i
=
0
;
i
<
num_inputs
;
++
i
)
{
input_vars
[
i
]
=
retrieve_node
(
name_scope
+
"/sequence_pool_in_"
+
std
::
to_string
(
i
),
subgraph
,
fused_pattern
);
input_names
[
i
]
=
input_vars
[
i
]
->
Name
();
}
auto
*
concat_op
=
retrieve_node
(
name_scope
+
"/concat_op"
,
subgraph
,
fused_pattern
);
auto
*
concat_out_var
=
retrieve_node
(
name_scope
+
"/concat_out_var"
,
subgraph
,
fused_pattern
);
auto
*
seqpool_op0
=
retrieve_node
(
name_scope
+
"/sequence_pool_op_0"
,
subgraph
,
fused_pattern
);
// Create New OpDesc
OpDesc
op_desc
;
op_desc
.
SetType
(
"fusion_seqpool_concat"
);
op_desc
.
SetInput
(
"X"
,
input_names
);
op_desc
.
SetAttr
(
"pooltype"
,
seqpool_op0
->
Op
()
->
GetAttr
(
"pooltype"
));
op_desc
.
SetAttr
(
"axis"
,
concat_op
->
Op
()
->
GetAttr
(
"axis"
));
op_desc
.
SetOutput
(
"Out"
,
{
concat_out_var
->
Name
()});
auto
*
op
=
graph
->
CreateOpNode
(
&
op_desc
);
for
(
size_t
i
=
0
;
i
<
input_vars
.
size
();
++
i
)
{
IR_NODE_LINK_TO
(
input_vars
[
i
],
op
);
}
IR_NODE_LINK_TO
(
op
,
concat_out_var
);
std
::
unordered_set
<
const
Node
*>
marked_nodes
;
for
(
auto
&
item
:
subgraph
)
{
marked_nodes
.
insert
(
item
.
second
);
}
for
(
size_t
i
=
0
;
i
<
input_vars
.
size
();
++
i
)
{
marked_nodes
.
erase
(
input_vars
[
i
]);
}
marked_nodes
.
erase
(
concat_out_var
);
GraphSafeRemoveNodes
(
graph
,
marked_nodes
);
++
fusion_count
;
};
gpd
(
graph
,
handler
);
return
fusion_count
;
}
std
::
unique_ptr
<
ir
::
Graph
>
SeqPoolConcatFusePass
::
ApplyImpl
(
std
::
unique_ptr
<
ir
::
Graph
>
graph
)
const
{
FusePassBase
::
Init
(
name_scope_
,
graph
.
get
());
int
fusion_count
=
0
;
for
(
int
i
=
MAX_CONCAT_INPUTS
;
i
>
0
;
--
i
)
{
fusion_count
+=
BuildFusion
(
graph
.
get
(),
name_scope_
+
"/"
+
std
::
to_string
(
i
),
param_scope
(),
i
);
}
AddStatis
(
fusion_count
);
return
graph
;
}
}
// namespace ir
}
// namespace framework
}
// namespace paddle
REGISTER_PASS
(
seqpool_concat_fuse_pass
,
paddle
::
framework
::
ir
::
SeqPoolConcatFusePass
);
paddle/fluid/framework/ir/seqpool_concat_fuse_pass.h
0 → 100644
浏览文件 @
72d2a180
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License. */
#pragma once
#include <string>
#include "paddle/fluid/framework/ir/fuse_pass_base.h"
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/ir/graph_pattern_detector.h"
namespace
paddle
{
namespace
framework
{
namespace
ir
{
class
SeqPoolConcatFusePass
:
public
FusePassBase
{
public:
virtual
~
SeqPoolConcatFusePass
()
{}
protected:
std
::
unique_ptr
<
ir
::
Graph
>
ApplyImpl
(
std
::
unique_ptr
<
ir
::
Graph
>
graph
)
const
;
const
std
::
string
name_scope_
{
"seqpool_concat_fuse"
};
};
}
// namespace ir
}
// namespace framework
}
// namespace paddle
paddle/fluid/inference/api/paddle_pass_builder.h
浏览文件 @
72d2a180
...
...
@@ -89,6 +89,7 @@ class CpuPassStrategy : public PassStrategy {
passes_
.
assign
({
"infer_clean_graph_pass"
,
//
"attention_lstm_fuse_pass"
,
//
"seqpool_concat_fuse_pass"
,
//
"seqconv_eltadd_relu_fuse_pass"
,
//
// "embedding_fc_lstm_fuse_pass", //
"fc_lstm_fuse_pass"
,
//
...
...
paddle/fluid/inference/tests/api/analyzer_seq_pool1_tester.cc
浏览文件 @
72d2a180
...
...
@@ -177,8 +177,12 @@ TEST(Analyzer_seq_pool1, fuse_statis) {
auto
predictor
=
CreatePaddlePredictor
<
AnalysisConfig
>
(
cfg
);
auto
fuse_statis
=
GetFuseStatis
(
static_cast
<
AnalysisPredictor
*>
(
predictor
.
get
()),
&
num_ops
);
ASSERT_TRUE
(
fuse_statis
.
count
(
"seqpool_concat_fuse"
));
EXPECT_EQ
(
fuse_statis
.
at
(
"seqpool_concat_fuse"
),
2
);
LOG
(
INFO
)
<<
"num_ops: "
<<
num_ops
;
EXPECT_EQ
(
num_ops
,
349
);
EXPECT_EQ
(
num_ops
,
195
);
}
}
// namespace analysis
...
...
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