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
Crayon鑫
Paddle
提交
fbc16404
P
Paddle
项目概览
Crayon鑫
/
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
提交
fbc16404
编写于
8月 15, 2018
作者:
T
tensor-tang
浏览文件
操作
浏览文件
下载
差异文件
Merge remote-tracking branch 'ups/develop' into refine/op/fc
上级
0098a494
d96ee24f
变更
3
隐藏空白更改
内联
并排
Showing
3 changed file
with
139 addition
and
6 deletion
+139
-6
paddle/fluid/framework/ir/graph.h
paddle/fluid/framework/ir/graph.h
+32
-0
paddle/fluid/framework/ir/graph_test.cc
paddle/fluid/framework/ir/graph_test.cc
+95
-1
paddle/fluid/platform/profiler.cc
paddle/fluid/platform/profiler.cc
+12
-5
未找到文件。
paddle/fluid/framework/ir/graph.h
浏览文件 @
fbc16404
...
@@ -28,6 +28,38 @@ namespace paddle {
...
@@ -28,6 +28,38 @@ namespace paddle {
namespace
framework
{
namespace
framework
{
namespace
ir
{
namespace
ir
{
/*
* The graph is a Directed Acyclic Single Static Assignment Graph.
*
* In more detail, the following properties must hold:
*
* The graph shouldn't contain cycle. Each node is a black-box to the graph
* so the node itself could be a loop operator.
*
* Each Variable-type node has only one input (thus single static assignment).
*
* The output/input of operator is variable and the output/input of variable
* is operator.
*
* The following data harzards in Program are addressed in the Graph:
*
* Write-After-Read
* a = op1(x)
* x = op2(b)
* A control-dependency connection is created bettwen op1 and op2 such that
* op1->op2, so as to ensure correct order.
*
* Write-After-Write
* x = op1(a)
* x = op2(b)
* A control-dependency connection is created between op1 and op2 such that
* op1->op2, so as to ensure correct order.
*
* Other properties currently hold, but is not enforced yet:
*
* Variable-type node (not control dep) with the same variable name share
* the same underlying VarDesc.
*/
class
Graph
{
class
Graph
{
public:
public:
explicit
Graph
(
const
ProgramDesc
&
program
);
explicit
Graph
(
const
ProgramDesc
&
program
);
...
...
paddle/fluid/framework/ir/graph_test.cc
浏览文件 @
fbc16404
...
@@ -36,7 +36,7 @@ class SumOpMaker : public OpProtoAndCheckerMaker {
...
@@ -36,7 +36,7 @@ class SumOpMaker : public OpProtoAndCheckerMaker {
public:
public:
void
Make
()
{
void
Make
()
{
AddInput
(
"X"
,
""
).
AsDuplicable
();
AddInput
(
"X"
,
""
).
AsDuplicable
();
AddOutput
(
"Out"
,
""
);
AddOutput
(
"Out"
,
""
)
.
AsDuplicable
()
;
AddComment
(
""
);
AddComment
(
""
);
}
}
};
};
...
@@ -59,11 +59,27 @@ class SumOpVarTypeInference : public VarTypeInference {
...
@@ -59,11 +59,27 @@ class SumOpVarTypeInference : public VarTypeInference {
block
->
Var
(
out_var_name
)
->
SetType
(
default_var_type
);
block
->
Var
(
out_var_name
)
->
SetType
(
default_var_type
);
}
}
};
};
class
DummyOpMaker
:
public
OpProtoAndCheckerMaker
{
public:
void
Make
()
{
AddInput
(
"X"
,
""
).
AsDuplicable
();
AddOutput
(
"Out"
,
""
).
AsDuplicable
();
AddComment
(
""
);
}
};
class
DummyOpVarTypeInference
:
public
VarTypeInference
{
public:
void
operator
()(
const
OpDesc
&
op_desc
,
BlockDesc
*
block
)
const
override
{}
};
}
// namespace framework
}
// namespace framework
}
// namespace paddle
}
// namespace paddle
REGISTER_OPERATOR
(
sum
,
paddle
::
framework
::
NOP
,
paddle
::
framework
::
SumOpMaker
,
REGISTER_OPERATOR
(
sum
,
paddle
::
framework
::
NOP
,
paddle
::
framework
::
SumOpMaker
,
paddle
::
framework
::
SumOpVarTypeInference
);
paddle
::
framework
::
SumOpVarTypeInference
);
REGISTER_OPERATOR
(
dummy
,
paddle
::
framework
::
NOP
,
paddle
::
framework
::
SumOpMaker
,
paddle
::
framework
::
SumOpVarTypeInference
);
REGISTER_OPERATOR
(
sum_without_infer_var_type
,
paddle
::
framework
::
NOP
,
REGISTER_OPERATOR
(
sum_without_infer_var_type
,
paddle
::
framework
::
NOP
,
paddle
::
framework
::
SumOpMaker
);
paddle
::
framework
::
SumOpMaker
);
...
@@ -110,5 +126,83 @@ TEST(GraphTest, Basic) {
...
@@ -110,5 +126,83 @@ TEST(GraphTest, Basic) {
}
}
ASSERT_EQ
(
nodes
.
size
(),
5
);
ASSERT_EQ
(
nodes
.
size
(),
5
);
}
}
TEST
(
GraphTest
,
WriteAfterRead
)
{
// void Test() {
ProgramDesc
prog
;
auto
*
op
=
prog
.
MutableBlock
(
0
)
->
AppendOp
();
op
->
SetType
(
"sum"
);
op
->
SetInput
(
"X"
,
{
"a"
});
op
->
SetOutput
(
"Out"
,
{
"b"
});
op
->
SetAttr
(
"op_role"
,
1
);
op
=
prog
.
MutableBlock
(
0
)
->
AppendOp
();
op
->
SetType
(
"dummy"
);
op
->
SetInput
(
"X"
,
{
"c"
});
op
->
SetOutput
(
"Out"
,
{
"a"
});
op
->
SetAttr
(
"op_role"
,
1
);
prog
.
MutableBlock
(
0
)
->
Var
(
"a"
)
->
SetType
(
proto
::
VarType
::
LOD_TENSOR
);
prog
.
MutableBlock
(
0
)
->
Var
(
"b"
)
->
SetType
(
proto
::
VarType
::
LOD_TENSOR
);
prog
.
MutableBlock
(
0
)
->
Var
(
"c"
)
->
SetType
(
proto
::
VarType
::
LOD_TENSOR
);
std
::
unique_ptr
<
ir
::
Graph
>
g
(
new
ir
::
Graph
(
prog
));
ir
::
Node
*
control_dep1
=
nullptr
;
ir
::
Node
*
control_dep2
=
nullptr
;
for
(
ir
::
Node
*
n
:
g
->
Nodes
())
{
if
(
n
->
Name
()
==
"sum"
)
{
ASSERT_EQ
(
n
->
outputs
[
0
]
->
Name
(),
"b"
);
ASSERT_TRUE
(
ir
::
IsControlDepVar
(
*
n
->
outputs
[
1
]));
control_dep1
=
n
->
outputs
[
1
];
ASSERT_EQ
(
n
->
outputs
.
size
(),
2
);
}
if
(
n
->
Name
()
==
"dummy"
)
{
ASSERT_EQ
(
n
->
inputs
[
0
]
->
Name
(),
"c"
);
ASSERT_TRUE
(
ir
::
IsControlDepVar
(
*
n
->
inputs
[
1
]));
control_dep2
=
n
->
inputs
[
1
];
ASSERT_EQ
(
n
->
inputs
.
size
(),
2
);
}
}
ASSERT_EQ
(
control_dep1
,
control_dep2
);
}
TEST
(
GraphTest
,
WriteAfterWrite
)
{
// void Test() {
ProgramDesc
prog
;
auto
*
op
=
prog
.
MutableBlock
(
0
)
->
AppendOp
();
op
->
SetType
(
"sum"
);
op
->
SetInput
(
"X"
,
{
"a"
});
op
->
SetOutput
(
"Out"
,
{
"b"
});
op
->
SetAttr
(
"op_role"
,
1
);
op
=
prog
.
MutableBlock
(
0
)
->
AppendOp
();
op
->
SetType
(
"dummy"
);
op
->
SetInput
(
"X"
,
{
"c"
});
op
->
SetOutput
(
"Out"
,
{
"b"
});
op
->
SetAttr
(
"op_role"
,
1
);
prog
.
MutableBlock
(
0
)
->
Var
(
"a"
)
->
SetType
(
proto
::
VarType
::
LOD_TENSOR
);
prog
.
MutableBlock
(
0
)
->
Var
(
"b"
)
->
SetType
(
proto
::
VarType
::
LOD_TENSOR
);
prog
.
MutableBlock
(
0
)
->
Var
(
"c"
)
->
SetType
(
proto
::
VarType
::
LOD_TENSOR
);
std
::
unique_ptr
<
ir
::
Graph
>
g
(
new
ir
::
Graph
(
prog
));
ir
::
Node
*
control_dep1
=
nullptr
;
ir
::
Node
*
control_dep2
=
nullptr
;
for
(
ir
::
Node
*
n
:
g
->
Nodes
())
{
if
(
n
->
Name
()
==
"sum"
)
{
ASSERT_EQ
(
n
->
outputs
[
0
]
->
Name
(),
"b"
);
ASSERT_TRUE
(
ir
::
IsControlDepVar
(
*
n
->
outputs
[
1
]));
ASSERT_EQ
(
n
->
outputs
.
size
(),
2
);
control_dep1
=
n
->
outputs
[
1
];
}
if
(
n
->
Name
()
==
"dummy"
)
{
ASSERT_EQ
(
n
->
inputs
[
0
]
->
Name
(),
"c"
);
ASSERT_TRUE
(
ir
::
IsControlDepVar
(
*
n
->
inputs
[
1
]));
control_dep2
=
n
->
inputs
[
1
];
ASSERT_EQ
(
n
->
inputs
.
size
(),
2
);
ASSERT_EQ
(
control_dep1
,
control_dep2
);
}
}
}
}
// namespace framework
}
// namespace framework
}
// namespace paddle
}
// namespace paddle
paddle/fluid/platform/profiler.cc
浏览文件 @
fbc16404
...
@@ -270,12 +270,13 @@ struct EventItem {
...
@@ -270,12 +270,13 @@ struct EventItem {
double
min_time
;
double
min_time
;
double
max_time
;
double
max_time
;
double
ave_time
;
double
ave_time
;
float
ratio
;
};
};
// Print results
// Print results
void
PrintProfiler
(
const
std
::
vector
<
std
::
vector
<
EventItem
>>&
events_table
,
void
PrintProfiler
(
const
std
::
vector
<
std
::
vector
<
EventItem
>>&
events_table
,
const
std
::
string
&
sorted_domain
,
const
size_t
name_width
,
const
std
::
string
&
sorted_domain
,
const
size_t
name_width
,
const
size_t
data_width
)
{
const
size_t
data_width
,
double
total
)
{
// Output header information
// Output header information
std
::
cout
<<
"
\n
------------------------->"
std
::
cout
<<
"
\n
------------------------->"
<<
" Profiling Report "
<<
" Profiling Report "
...
@@ -300,7 +301,8 @@ void PrintProfiler(const std::vector<std::vector<EventItem>>& events_table,
...
@@ -300,7 +301,8 @@ void PrintProfiler(const std::vector<std::vector<EventItem>>& events_table,
std
::
cout
<<
std
::
setw
(
name_width
)
<<
"Event"
<<
std
::
setw
(
data_width
)
std
::
cout
<<
std
::
setw
(
name_width
)
<<
"Event"
<<
std
::
setw
(
data_width
)
<<
"Calls"
<<
std
::
setw
(
data_width
)
<<
"Total"
<<
"Calls"
<<
std
::
setw
(
data_width
)
<<
"Total"
<<
std
::
setw
(
data_width
)
<<
"Min."
<<
std
::
setw
(
data_width
)
<<
std
::
setw
(
data_width
)
<<
"Min."
<<
std
::
setw
(
data_width
)
<<
"Max."
<<
std
::
setw
(
data_width
)
<<
"Ave."
<<
std
::
endl
;
<<
"Max."
<<
std
::
setw
(
data_width
)
<<
"Ave."
<<
std
::
setw
(
data_width
)
<<
"Ratio."
<<
std
::
endl
;
for
(
size_t
i
=
0
;
i
<
events_table
.
size
();
++
i
)
{
for
(
size_t
i
=
0
;
i
<
events_table
.
size
();
++
i
)
{
for
(
size_t
j
=
0
;
j
<
events_table
[
i
].
size
();
++
j
)
{
for
(
size_t
j
=
0
;
j
<
events_table
[
i
].
size
();
++
j
)
{
const
EventItem
&
event_item
=
events_table
[
i
][
j
];
const
EventItem
&
event_item
=
events_table
[
i
][
j
];
...
@@ -309,7 +311,9 @@ void PrintProfiler(const std::vector<std::vector<EventItem>>& events_table,
...
@@ -309,7 +311,9 @@ void PrintProfiler(const std::vector<std::vector<EventItem>>& events_table,
<<
std
::
setw
(
data_width
)
<<
event_item
.
total_time
<<
std
::
setw
(
data_width
)
<<
event_item
.
total_time
<<
std
::
setw
(
data_width
)
<<
event_item
.
min_time
<<
std
::
setw
(
data_width
)
<<
event_item
.
min_time
<<
std
::
setw
(
data_width
)
<<
event_item
.
max_time
<<
std
::
setw
(
data_width
)
<<
event_item
.
max_time
<<
std
::
setw
(
data_width
)
<<
event_item
.
ave_time
<<
std
::
endl
;
<<
std
::
setw
(
data_width
)
<<
event_item
.
ave_time
<<
std
::
setw
(
data_width
)
<<
event_item
.
total_time
/
total
<<
std
::
endl
;
}
}
}
}
std
::
cout
<<
std
::
endl
;
std
::
cout
<<
std
::
endl
;
...
@@ -359,6 +363,7 @@ void ParseEvents(const std::vector<std::vector<Event>>& events,
...
@@ -359,6 +363,7 @@ void ParseEvents(const std::vector<std::vector<Event>>& events,
std
::
vector
<
std
::
vector
<
EventItem
>>
events_table
;
std
::
vector
<
std
::
vector
<
EventItem
>>
events_table
;
size_t
max_name_width
=
0
;
size_t
max_name_width
=
0
;
double
total
=
0.
;
// the total time
for
(
size_t
i
=
0
;
i
<
events
.
size
();
i
++
)
{
for
(
size_t
i
=
0
;
i
<
events
.
size
();
i
++
)
{
std
::
list
<
Event
>
pushed_events
;
std
::
list
<
Event
>
pushed_events
;
std
::
vector
<
EventItem
>
event_items
;
std
::
vector
<
EventItem
>
event_items
;
...
@@ -379,6 +384,7 @@ void ParseEvents(const std::vector<std::vector<Event>>& events,
...
@@ -379,6 +384,7 @@ void ParseEvents(const std::vector<std::vector<Event>>& events,
g_state
==
ProfilerState
::
kAll
)
g_state
==
ProfilerState
::
kAll
)
?
rit
->
CudaElapsedMs
(
events
[
i
][
j
])
?
rit
->
CudaElapsedMs
(
events
[
i
][
j
])
:
rit
->
CpuElapsedMs
(
events
[
i
][
j
]);
:
rit
->
CpuElapsedMs
(
events
[
i
][
j
]);
total
+=
event_time
;
std
::
string
event_name
=
std
::
string
event_name
=
"thread"
+
std
::
to_string
(
rit
->
thread_id
())
+
"::"
+
rit
->
name
();
"thread"
+
std
::
to_string
(
rit
->
thread_id
())
+
"::"
+
rit
->
name
();
...
@@ -387,7 +393,8 @@ void ParseEvents(const std::vector<std::vector<Event>>& events,
...
@@ -387,7 +393,8 @@ void ParseEvents(const std::vector<std::vector<Event>>& events,
if
(
event_idx
.
find
(
event_name
)
==
event_idx
.
end
())
{
if
(
event_idx
.
find
(
event_name
)
==
event_idx
.
end
())
{
event_idx
[
event_name
]
=
event_items
.
size
();
event_idx
[
event_name
]
=
event_items
.
size
();
EventItem
event_item
=
{
event_name
,
1
,
event_time
,
EventItem
event_item
=
{
event_name
,
1
,
event_time
,
event_time
,
event_time
,
event_time
};
event_time
,
event_time
,
event_time
,
0.
};
event_items
.
push_back
(
event_item
);
event_items
.
push_back
(
event_item
);
}
else
{
}
else
{
int
index
=
event_idx
[
event_name
];
int
index
=
event_idx
[
event_name
];
...
@@ -431,7 +438,7 @@ void ParseEvents(const std::vector<std::vector<Event>>& events,
...
@@ -431,7 +438,7 @@ void ParseEvents(const std::vector<std::vector<Event>>& events,
}
}
// Print report
// Print report
PrintProfiler
(
events_table
,
sorted_domain
,
max_name_width
+
4
,
12
);
PrintProfiler
(
events_table
,
sorted_domain
,
max_name_width
+
4
,
12
,
total
);
}
}
void
DisableProfiler
(
EventSortingKey
sorted_key
,
void
DisableProfiler
(
EventSortingKey
sorted_key
,
...
...
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