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
af91d41a
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 >>
未验证
提交
af91d41a
编写于
10月 16, 2018
作者:
Z
Zeng Jinle
提交者:
GitHub
10月 16, 2018
浏览文件
操作
浏览文件
下载
差异文件
Merge pull request #13852 from sneaxiy/feature/eager_delete_tensor
Fix bug of eager deletion to support if_else_op
上级
93606c2c
d3ed070e
变更
7
隐藏空白更改
内联
并排
Showing
7 changed file
with
91 addition
and
80 deletion
+91
-80
paddle/fluid/framework/details/op_handle_base.h
paddle/fluid/framework/details/op_handle_base.h
+2
-1
paddle/fluid/framework/executor.cc
paddle/fluid/framework/executor.cc
+43
-41
paddle/fluid/framework/executor.h
paddle/fluid/framework/executor.h
+19
-25
paddle/fluid/framework/parallel_executor.cc
paddle/fluid/framework/parallel_executor.cc
+4
-0
paddle/fluid/framework/parallel_executor.h
paddle/fluid/framework/parallel_executor.h
+1
-1
paddle/fluid/framework/scope.cc
paddle/fluid/framework/scope.cc
+17
-12
paddle/fluid/framework/scope.h
paddle/fluid/framework/scope.h
+5
-0
未找到文件。
paddle/fluid/framework/details/op_handle_base.h
浏览文件 @
af91d41a
...
@@ -64,7 +64,8 @@ class OpHandleBase {
...
@@ -64,7 +64,8 @@ class OpHandleBase {
virtual
bool
IsMultiDeviceTransfer
()
{
return
false
;
}
virtual
bool
IsMultiDeviceTransfer
()
{
return
false
;
}
const
platform
::
DeviceContext
*
DeviceContext
(
platform
::
Place
place
)
{
const
platform
::
DeviceContext
*
DeviceContext
(
platform
::
Place
place
)
{
return
dev_ctxes_
[
place
];
auto
it
=
dev_ctxes_
.
find
(
place
);
return
it
!=
dev_ctxes_
.
end
()
?
it
->
second
:
nullptr
;
}
}
void
SetDeviceContext
(
platform
::
Place
place
,
platform
::
DeviceContext
*
ctx_
)
{
void
SetDeviceContext
(
platform
::
Place
place
,
platform
::
DeviceContext
*
ctx_
)
{
...
...
paddle/fluid/framework/executor.cc
浏览文件 @
af91d41a
...
@@ -46,6 +46,41 @@ ExecutorPrepareContext::~ExecutorPrepareContext() {
...
@@ -46,6 +46,41 @@ ExecutorPrepareContext::~ExecutorPrepareContext() {
VLOG
(
5
)
<<
"destroy ExecutorPrepareContext"
;
VLOG
(
5
)
<<
"destroy ExecutorPrepareContext"
;
}
}
template
<
typename
RefCntMap
>
static
void
DeleteUnusedTensors
(
const
Scope
&
scope
,
const
OperatorBase
*
op
,
GarbageCollector
<
Tensor
>*
gc
,
RefCntMap
*
ref_cnts
)
{
std
::
unordered_set
<
Tensor
*>
erase_tensors
;
auto
handler
=
[
&
](
const
VariableNameMap
&
name_map
)
{
for
(
auto
&
name_pair
:
name_map
)
{
for
(
auto
&
name
:
name_pair
.
second
)
{
auto
it
=
ref_cnts
->
find
(
name
);
if
(
it
==
ref_cnts
->
end
())
continue
;
if
((
it
->
second
)
--
==
1
)
{
auto
*
var
=
scope
.
FindVar
(
name
);
if
(
var
!=
nullptr
)
{
VLOG
(
10
)
<<
"Erase tensor
\'
"
<<
name
<<
"
\'
"
;
if
(
var
->
IsType
<
LoDTensor
>
())
{
erase_tensors
.
insert
(
var
->
GetMutable
<
LoDTensor
>
());
}
else
if
(
var
->
IsType
<
SelectedRows
>
())
{
erase_tensors
.
insert
(
var
->
GetMutable
<
SelectedRows
>
()
->
mutable_value
());
}
}
}
}
}
};
handler
(
op
->
Inputs
());
handler
(
op
->
Outputs
());
if
(
!
erase_tensors
.
empty
())
{
gc
->
Add
(
erase_tensors
);
}
}
Executor
::
Executor
(
const
platform
::
Place
&
place
)
:
place_
(
place
)
{}
Executor
::
Executor
(
const
platform
::
Place
&
place
)
:
place_
(
place
)
{}
void
Executor
::
Close
()
{
void
Executor
::
Close
()
{
...
@@ -331,9 +366,13 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
...
@@ -331,9 +366,13 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
}
}
int64_t
max_memory_size
=
GetEagerDeletionThreshold
();
int64_t
max_memory_size
=
GetEagerDeletionThreshold
();
std
::
unique_ptr
<
GarbageCollector
<
Tensor
>>
gc
;
std
::
unique_ptr
<
GarbageCollector
<
Tensor
>>
gc
;
if
(
max_memory_size
>=
0
)
{
// WhileOp would set keep_kids to false
// WhileGradOp would need the scopes created in WhileOp
// Perhaps, we should not perform eager deletion in WhileOp
// The scopes and variables created by WhileOp would be deleted
// in WhileGradOp.
if
(
max_memory_size
>=
0
&&
!
keep_kids
)
{
ctx
->
ResetReferenceCount
();
ctx
->
ResetReferenceCount
();
#ifdef PADDLE_WITH_CUDA
#ifdef PADDLE_WITH_CUDA
if
(
platform
::
is_gpu_place
(
place_
))
{
if
(
platform
::
is_gpu_place
(
place_
))
{
...
@@ -352,45 +391,8 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
...
@@ -352,45 +391,8 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
op
->
Run
(
*
local_scope
,
place_
);
op
->
Run
(
*
local_scope
,
place_
);
if
(
gc
!=
nullptr
)
{
if
(
gc
!=
nullptr
)
{
std
::
vector
<
std
::
string
>
erase_vars
;
DeleteUnusedTensors
(
*
local_scope
,
op
.
get
(),
gc
.
get
(),
for
(
auto
&
input
:
op
->
Inputs
())
{
&
(
ctx
->
cur_ref_cnts_
));
for
(
auto
&
input_name
:
input
.
second
)
{
auto
it
=
ctx
->
cur_ref_cnts_
.
find
(
input_name
);
if
(
it
==
ctx
->
cur_ref_cnts_
.
end
())
continue
;
if
(
it
->
second
==
1
)
{
// should delete it
erase_vars
.
emplace_back
(
input_name
);
ctx
->
cur_ref_cnts_
.
erase
(
input_name
);
}
else
{
--
(
it
->
second
);
}
}
}
for
(
auto
&
output
:
op
->
Outputs
())
{
for
(
auto
&
output_name
:
output
.
second
)
{
auto
it
=
ctx
->
cur_ref_cnts_
.
find
(
output_name
);
if
(
it
==
ctx
->
cur_ref_cnts_
.
end
())
continue
;
if
(
it
->
second
==
1
)
{
erase_vars
.
emplace_back
(
output_name
);
ctx
->
cur_ref_cnts_
.
erase
(
output_name
);
}
else
{
--
(
it
->
second
);
}
}
}
if
(
!
erase_vars
.
empty
())
{
std
::
vector
<
framework
::
LoDTensor
*>
erase_tensors
;
for
(
auto
&
name
:
erase_vars
)
{
auto
*
var
=
local_scope
->
FindVar
(
name
);
if
(
var
==
nullptr
)
continue
;
if
(
var
->
IsType
<
framework
::
LoDTensor
>
())
{
auto
*
tensor
=
var
->
GetMutable
<
framework
::
LoDTensor
>
();
erase_tensors
.
push_back
(
tensor
);
}
}
if
(
!
erase_tensors
.
empty
())
gc
->
Add
(
erase_tensors
);
}
}
}
if
(
FLAGS_benchmark
)
{
if
(
FLAGS_benchmark
)
{
...
...
paddle/fluid/framework/executor.h
浏览文件 @
af91d41a
...
@@ -32,38 +32,32 @@ template <typename T>
...
@@ -32,38 +32,32 @@ template <typename T>
std
::
unordered_map
<
std
::
string
,
T
>
GetNonPersistableReferenceCount
(
std
::
unordered_map
<
std
::
string
,
T
>
GetNonPersistableReferenceCount
(
const
ProgramDesc
&
prog
,
size_t
block_id
)
{
const
ProgramDesc
&
prog
,
size_t
block_id
)
{
auto
&
block
=
prog
.
Block
(
block_id
);
auto
&
block
=
prog
.
Block
(
block_id
);
std
::
unordered_set
<
std
::
string
>
ignored_vars
;
std
::
unordered_map
<
std
::
string
,
T
>
ref_cnts
;
std
::
unordered_map
<
std
::
string
,
T
>
ref_cnts
;
for
(
auto
var_desc
:
block
.
AllVars
())
{
auto
update_ref_cnts
=
[
&
](
OpDesc
*
op_desc
,
const
VariableNameMap
&
name_map
)
{
auto
type
=
var_desc
->
Proto
()
->
type
().
type
();
for
(
auto
&
name_pair
:
name_map
)
{
if
(
type
!=
proto
::
VarType
::
LOD_TENSOR
||
var_desc
->
Persistable
())
{
for
(
auto
&
name
:
name_pair
.
second
)
{
ignored_vars
.
insert
(
var_desc
->
Name
());
// ignore persistable vars
auto
*
var_desc
=
block
.
FindVar
(
name
);
}
if
(
var_desc
==
nullptr
||
var_desc
->
Persistable
())
continue
;
}
auto
type
=
var_desc
->
Proto
()
->
type
().
type
();
if
(
type
!=
proto
::
VarType
::
LOD_TENSOR
&&
for
(
auto
op_desc
:
block
.
AllOps
())
{
type
!=
proto
::
VarType
::
SELECTED_ROWS
)
{
for
(
auto
&
input
:
op_desc
->
Inputs
())
{
continue
;
for
(
auto
&
input_name
:
input
.
second
)
{
if
(
!
ignored_vars
.
count
(
input_name
))
{
if
(
ref_cnts
.
count
(
input_name
))
++
ref_cnts
[
input_name
];
else
ref_cnts
[
input_name
]
=
1
;
}
}
}
}
for
(
auto
&
output
:
op_desc
->
Outputs
())
{
auto
it
=
ref_cnts
.
find
(
name
);
for
(
auto
output_name
:
output
.
second
)
{
if
(
it
!=
ref_cnts
.
end
())
{
if
(
!
ignored_vars
.
count
(
output_name
))
{
++
it
->
second
;
if
(
ref_cnts
.
count
(
output_name
))
}
else
{
++
ref_cnts
[
output_name
];
ref_cnts
[
name
]
=
1
;
else
ref_cnts
[
output_name
]
=
1
;
}
}
}
}
}
}
};
for
(
auto
op_desc
:
block
.
AllOps
())
{
update_ref_cnts
(
op_desc
,
op_desc
->
Inputs
());
update_ref_cnts
(
op_desc
,
op_desc
->
Outputs
());
}
}
return
ref_cnts
;
return
ref_cnts
;
}
}
...
...
paddle/fluid/framework/parallel_executor.cc
浏览文件 @
af91d41a
...
@@ -307,6 +307,10 @@ ParallelExecutor::~ParallelExecutor() {
...
@@ -307,6 +307,10 @@ ParallelExecutor::~ParallelExecutor() {
}
}
}
}
}
}
// member_ must be destructed before gcs_ since the destructor of
// ReferenceCountOpHandle use raw pointers of gcs_ inside.
member_
.
reset
();
}
}
}
// namespace framework
}
// namespace framework
...
...
paddle/fluid/framework/parallel_executor.h
浏览文件 @
af91d41a
...
@@ -75,7 +75,7 @@ class ParallelExecutor {
...
@@ -75,7 +75,7 @@ class ParallelExecutor {
private:
private:
void
BCastParamsToDevices
(
const
std
::
unordered_set
<
std
::
string
>
&
vars
)
const
;
void
BCastParamsToDevices
(
const
std
::
unordered_set
<
std
::
string
>
&
vars
)
const
;
ParallelExecutorPrivate
*
member_
;
std
::
unique_ptr
<
ParallelExecutorPrivate
>
member_
;
#ifdef PADDLE_WITH_CUDA
#ifdef PADDLE_WITH_CUDA
// ref_cnts_ is only initialized when ParallelExecutor constructs, and then
// ref_cnts_ is only initialized when ParallelExecutor constructs, and then
...
...
paddle/fluid/framework/scope.cc
浏览文件 @
af91d41a
...
@@ -49,18 +49,18 @@ int64_t GetEagerDeletionThreshold() {
...
@@ -49,18 +49,18 @@ int64_t GetEagerDeletionThreshold() {
Scope
::~
Scope
()
{
DropKids
();
}
Scope
::~
Scope
()
{
DropKids
();
}
Scope
&
Scope
::
NewScope
()
const
{
Scope
&
Scope
::
NewScope
()
const
{
std
::
unique_lock
<
std
::
mutex
>
lock
(
mutex_
);
std
::
lock_guard
<
std
::
mutex
>
lock
(
mutex_
);
kids_
.
push_back
(
new
Scope
(
this
));
kids_
.
push_back
(
new
Scope
(
this
));
return
*
kids_
.
back
();
return
*
kids_
.
back
();
}
}
Variable
*
Scope
::
Var
(
const
std
::
string
&
name
)
{
Variable
*
Scope
::
Var
(
const
std
::
string
&
name
)
{
std
::
unique_lock
<
std
::
mutex
>
lock
(
mutex_
);
std
::
lock_guard
<
std
::
mutex
>
lock
(
mutex_
);
return
VarInternal
(
name
);
return
VarInternal
(
name
);
}
}
Variable
*
Scope
::
Var
(
std
::
string
*
name
)
{
Variable
*
Scope
::
Var
(
std
::
string
*
name
)
{
std
::
unique_lock
<
std
::
mutex
>
lock
(
mutex_
);
std
::
lock_guard
<
std
::
mutex
>
lock
(
mutex_
);
auto
new_name
=
string
::
Sprintf
(
"%p.%d"
,
this
,
vars_
.
size
());
auto
new_name
=
string
::
Sprintf
(
"%p.%d"
,
this
,
vars_
.
size
());
if
(
name
!=
nullptr
)
{
if
(
name
!=
nullptr
)
{
*
name
=
new_name
;
*
name
=
new_name
;
...
@@ -69,29 +69,34 @@ Variable* Scope::Var(std::string* name) {
...
@@ -69,29 +69,34 @@ Variable* Scope::Var(std::string* name) {
}
}
Variable
*
Scope
::
FindVar
(
const
std
::
string
&
name
)
const
{
Variable
*
Scope
::
FindVar
(
const
std
::
string
&
name
)
const
{
std
::
unique_lock
<
std
::
mutex
>
lock
(
mutex_
);
std
::
lock_guard
<
std
::
mutex
>
lock
(
mutex_
);
return
FindVarInternal
(
name
);
return
FindVarInternal
(
name
);
}
}
Variable
*
Scope
::
FindLocalVar
(
const
std
::
string
&
name
)
const
{
std
::
lock_guard
<
std
::
mutex
>
lock
(
mutex_
);
return
FindVarLocally
(
name
);
}
const
Scope
*
Scope
::
FindScope
(
const
Variable
*
var
)
const
{
const
Scope
*
Scope
::
FindScope
(
const
Variable
*
var
)
const
{
std
::
unique_lock
<
std
::
mutex
>
lock
(
mutex_
);
std
::
lock_guard
<
std
::
mutex
>
lock
(
mutex_
);
return
FindScopeInternal
(
var
);
return
FindScopeInternal
(
var
);
}
}
void
Scope
::
DropKids
()
{
void
Scope
::
DropKids
()
{
std
::
unique_lock
<
std
::
mutex
>
lock
(
mutex_
);
std
::
lock_guard
<
std
::
mutex
>
lock
(
mutex_
);
for
(
Scope
*
s
:
kids_
)
delete
s
;
for
(
Scope
*
s
:
kids_
)
delete
s
;
kids_
.
clear
();
kids_
.
clear
();
}
}
bool
Scope
::
HasKid
(
const
Scope
*
scope
)
const
{
bool
Scope
::
HasKid
(
const
Scope
*
scope
)
const
{
std
::
unique_lock
<
std
::
mutex
>
lock
(
mutex_
);
std
::
lock_guard
<
std
::
mutex
>
lock
(
mutex_
);
auto
it
=
std
::
find
(
this
->
kids_
.
begin
(),
this
->
kids_
.
end
(),
scope
);
auto
it
=
std
::
find
(
this
->
kids_
.
begin
(),
this
->
kids_
.
end
(),
scope
);
return
it
!=
this
->
kids_
.
end
();
return
it
!=
this
->
kids_
.
end
();
}
}
std
::
vector
<
std
::
string
>
Scope
::
LocalVarNames
()
const
{
std
::
vector
<
std
::
string
>
Scope
::
LocalVarNames
()
const
{
std
::
unique_lock
<
std
::
mutex
>
lock
(
mutex_
);
std
::
lock_guard
<
std
::
mutex
>
lock
(
mutex_
);
std
::
vector
<
std
::
string
>
known_vars
;
std
::
vector
<
std
::
string
>
known_vars
;
known_vars
.
reserve
(
this
->
vars_
.
size
());
known_vars
.
reserve
(
this
->
vars_
.
size
());
for
(
auto
&
p
:
vars_
)
{
for
(
auto
&
p
:
vars_
)
{
...
@@ -101,7 +106,7 @@ std::vector<std::string> Scope::LocalVarNames() const {
...
@@ -101,7 +106,7 @@ std::vector<std::string> Scope::LocalVarNames() const {
}
}
void
Scope
::
DeleteScope
(
Scope
*
scope
)
const
{
void
Scope
::
DeleteScope
(
Scope
*
scope
)
const
{
std
::
unique_lock
<
std
::
mutex
>
lock
(
mutex_
);
std
::
lock_guard
<
std
::
mutex
>
lock
(
mutex_
);
auto
it
=
std
::
find
(
this
->
kids_
.
begin
(),
this
->
kids_
.
end
(),
scope
);
auto
it
=
std
::
find
(
this
->
kids_
.
begin
(),
this
->
kids_
.
end
(),
scope
);
PADDLE_ENFORCE
(
it
!=
this
->
kids_
.
end
(),
"Cannot find %p as kid scope"
,
scope
);
PADDLE_ENFORCE
(
it
!=
this
->
kids_
.
end
(),
"Cannot find %p as kid scope"
,
scope
);
this
->
kids_
.
erase
(
it
);
this
->
kids_
.
erase
(
it
);
...
@@ -114,7 +119,7 @@ void Scope::DeleteScope(Scope* scope) const {
...
@@ -114,7 +119,7 @@ void Scope::DeleteScope(Scope* scope) const {
}
}
void
Scope
::
EraseVars
(
const
std
::
vector
<
std
::
string
>&
var_names
)
{
void
Scope
::
EraseVars
(
const
std
::
vector
<
std
::
string
>&
var_names
)
{
std
::
unique_lock
<
std
::
mutex
>
lock
(
mutex_
);
std
::
lock_guard
<
std
::
mutex
>
lock
(
mutex_
);
std
::
set
<
std
::
string
>
var_set
(
var_names
.
begin
(),
var_names
.
end
());
std
::
set
<
std
::
string
>
var_set
(
var_names
.
begin
(),
var_names
.
end
());
for
(
auto
it
=
vars_
.
begin
();
it
!=
vars_
.
end
();)
{
for
(
auto
it
=
vars_
.
begin
();
it
!=
vars_
.
end
();)
{
if
(
var_set
.
find
(
it
->
first
)
!=
var_set
.
end
())
{
if
(
var_set
.
find
(
it
->
first
)
!=
var_set
.
end
())
{
...
@@ -127,12 +132,12 @@ void Scope::EraseVars(const std::vector<std::string>& var_names) {
...
@@ -127,12 +132,12 @@ void Scope::EraseVars(const std::vector<std::string>& var_names) {
void
Scope
::
Rename
(
const
std
::
string
&
origin_name
,
void
Scope
::
Rename
(
const
std
::
string
&
origin_name
,
const
std
::
string
&
new_name
)
const
{
const
std
::
string
&
new_name
)
const
{
std
::
unique_lock
<
std
::
mutex
>
lock
(
mutex_
);
std
::
lock_guard
<
std
::
mutex
>
lock
(
mutex_
);
RenameInternal
(
origin_name
,
new_name
);
RenameInternal
(
origin_name
,
new_name
);
}
}
std
::
string
Scope
::
Rename
(
const
std
::
string
&
origin_name
)
const
{
std
::
string
Scope
::
Rename
(
const
std
::
string
&
origin_name
)
const
{
std
::
unique_lock
<
std
::
mutex
>
lock
(
mutex_
);
std
::
lock_guard
<
std
::
mutex
>
lock
(
mutex_
);
auto
new_name
=
string
::
Sprintf
(
"%p.%d"
,
this
,
vars_
.
size
());
auto
new_name
=
string
::
Sprintf
(
"%p.%d"
,
this
,
vars_
.
size
());
RenameInternal
(
origin_name
,
new_name
);
RenameInternal
(
origin_name
,
new_name
);
return
new_name
;
return
new_name
;
...
...
paddle/fluid/framework/scope.h
浏览文件 @
af91d41a
...
@@ -63,6 +63,11 @@ class Scope {
...
@@ -63,6 +63,11 @@ class Scope {
/// Caller doesn't own the returned Variable.
/// Caller doesn't own the returned Variable.
Variable
*
FindVar
(
const
std
::
string
&
name
)
const
;
Variable
*
FindVar
(
const
std
::
string
&
name
)
const
;
/// Find a variable in the current scope.
/// Return nullptr if cannot find.
/// Caller doesn't own the returned Variable.
Variable
*
FindLocalVar
(
const
std
::
string
&
name
)
const
;
const
Scope
*
parent
()
const
{
return
parent_
;
}
const
Scope
*
parent
()
const
{
return
parent_
;
}
/// Find the scope or an ancestor scope that contains the given variable.
/// Find the scope or an ancestor scope that contains the given variable.
...
...
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