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
e1913bc5
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 >>
提交
e1913bc5
编写于
9月 21, 2018
作者:
Y
Yu Yang
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Fix MixedVector
上级
6d2c6f96
变更
4
隐藏空白更改
内联
并排
Showing
4 changed file
with
92 addition
and
62 deletion
+92
-62
paddle/fluid/framework/details/cow_ptr.h
paddle/fluid/framework/details/cow_ptr.h
+14
-18
paddle/fluid/framework/mixed_vector.h
paddle/fluid/framework/mixed_vector.h
+75
-39
paddle/fluid/operators/lookup_table_op.cu
paddle/fluid/operators/lookup_table_op.cu
+2
-4
paddle/fluid/operators/sgd_op.cu
paddle/fluid/operators/sgd_op.cu
+1
-1
未找到文件。
paddle/fluid/framework/details/cow_ptr.h
浏览文件 @
e1913bc5
...
@@ -28,31 +28,27 @@ class COWPtr {
...
@@ -28,31 +28,27 @@ class COWPtr {
private:
private:
RefPtr
m_sp
;
RefPtr
m_sp
;
void
detach
()
{
T
*
tmp
=
m_sp
.
get
();
if
(
!
(
tmp
==
nullptr
||
m_sp
.
unique
()))
{
m_sp
=
RefPtr
(
new
T
(
*
tmp
));
}
}
public:
public:
COWPtr
()
:
m_sp
(
nullptr
)
{}
COWPtr
()
:
m_sp
(
nullptr
)
{}
explicit
COWPtr
(
T
*
t
)
:
m_sp
(
t
)
{}
explicit
COWPtr
(
T
*
t
)
:
m_sp
(
t
)
{}
explicit
COWPtr
(
const
RefPtr
&
refptr
)
:
m_sp
(
refptr
)
{}
const
T
&
Data
()
const
{
return
operator
*
()
;
}
const
T
&
Data
()
const
{
return
*
m_sp
;
}
T
*
MutableData
()
{
return
operator
->
();
}
T
*
MutableData
()
{
DetachIfNotUnique
();
return
m_sp
.
get
();
}
const
T
&
operator
*
()
const
{
return
*
m_sp
;
}
void
DetachIfNotUnique
()
{
T
&
operator
*
()
{
T
*
tmp
=
m_sp
.
get
();
detach
();
if
(
!
(
tmp
==
nullptr
||
m_sp
.
unique
()))
{
return
*
m_sp
;
Detach
();
}
}
}
const
T
*
operator
->
()
const
{
return
m_sp
.
operator
->
();
}
T
*
operator
->
()
{
void
Detach
()
{
detach
();
T
*
tmp
=
m_sp
.
get
();
return
m_sp
.
operator
->
(
);
m_sp
=
RefPtr
(
new
T
(
*
tmp
)
);
}
}
};
};
}
// namespace details
}
// namespace details
...
...
paddle/fluid/framework/mixed_vector.h
浏览文件 @
e1913bc5
...
@@ -17,6 +17,7 @@
...
@@ -17,6 +17,7 @@
#include <algorithm>
#include <algorithm>
#include <initializer_list>
#include <initializer_list>
#include <memory>
#include <memory>
#include <mutex> // NOLINT
#include <utility>
#include <utility>
#include <vector>
#include <vector>
#include "paddle/fluid/framework/details/cow_ptr.h"
#include "paddle/fluid/framework/details/cow_ptr.h"
...
@@ -51,6 +52,7 @@ struct CUDABuffer {
...
@@ -51,6 +52,7 @@ struct CUDABuffer {
ClearMemory
();
ClearMemory
();
place_
=
boost
::
get
<
platform
::
CUDAPlace
>
(
place
);
place_
=
boost
::
get
<
platform
::
CUDAPlace
>
(
place
);
data_
=
memory
::
Alloc
(
place_
,
size
);
data_
=
memory
::
Alloc
(
place_
,
size
);
PADDLE_ENFORCE_NOT_NULL
(
data_
);
size_
=
size
;
size_
=
size
;
}
}
...
@@ -62,7 +64,7 @@ struct CUDABuffer {
...
@@ -62,7 +64,7 @@ struct CUDABuffer {
private:
private:
void
ClearMemory
()
const
{
void
ClearMemory
()
const
{
if
(
data_
)
{
if
(
data_
!=
nullptr
)
{
memory
::
Free
(
place_
,
data_
);
memory
::
Free
(
place_
,
data_
);
}
}
}
}
...
@@ -89,6 +91,7 @@ class Vector {
...
@@ -89,6 +91,7 @@ class Vector {
template
<
typename
U
>
template
<
typename
U
>
explicit
VectorData
(
const
std
::
vector
<
U
>
&
dat
)
explicit
VectorData
(
const
std
::
vector
<
U
>
&
dat
)
:
cpu_
(
dat
),
flag_
(
kDataInCPU
)
{}
:
cpu_
(
dat
),
flag_
(
kDataInCPU
)
{}
~
VectorData
()
{}
VectorData
(
const
VectorData
&
o
)
{
VectorData
(
const
VectorData
&
o
)
{
o
.
ImmutableCPU
();
o
.
ImmutableCPU
();
...
@@ -215,7 +218,7 @@ class Vector {
...
@@ -215,7 +218,7 @@ class Vector {
size_t
capacity
()
const
{
return
cpu_
.
capacity
();
}
size_t
capacity
()
const
{
return
cpu_
.
capacity
();
}
// reserve data
// reserve data
void
reserve
(
size_t
size
)
{
cpu_
.
reserve
(
size
);
}
void
reserve
(
size_t
size
)
const
{
cpu_
.
reserve
(
size
);
}
// implicit cast operator. Vector can be cast to std::vector implicitly.
// implicit cast operator. Vector can be cast to std::vector implicitly.
operator
std
::
vector
<
T
>
()
const
{
operator
std
::
vector
<
T
>
()
const
{
...
@@ -229,6 +232,17 @@ class Vector {
...
@@ -229,6 +232,17 @@ class Vector {
return
cpu_
==
other
.
cpu_
;
return
cpu_
==
other
.
cpu_
;
}
}
std
::
mutex
&
Mutex
()
const
{
return
mtx_
;
}
std
::
unique_ptr
<
platform
::
CUDAPlace
>
CUDAPlace
()
const
{
if
(
gpu_
.
data_
==
nullptr
)
{
return
nullptr
;
}
else
{
return
std
::
unique_ptr
<
platform
::
CUDAPlace
>
(
new
platform
::
CUDAPlace
(
gpu_
.
place_
));
}
}
private:
private:
enum
DataFlag
{
enum
DataFlag
{
kDataInCPU
=
0x01
,
kDataInCPU
=
0x01
,
...
@@ -239,10 +253,15 @@ class Vector {
...
@@ -239,10 +253,15 @@ class Vector {
void
CopyToCPU
()
const
{
void
CopyToCPU
()
const
{
// COPY GPU Data To CPU
// COPY GPU Data To CPU
auto
*
dev_ctx
=
static_cast
<
platform
::
CUDADeviceContext
*>
(
platform
::
DeviceContextPool
::
Instance
().
Get
(
platform
::
Place
(
gpu_
.
place_
)));
auto
stream
=
dev_ctx
->
stream
();
void
*
src
=
gpu_
.
data_
;
void
*
src
=
gpu_
.
data_
;
void
*
dst
=
cpu_
.
data
();
void
*
dst
=
cpu_
.
data
();
memory
::
Copy
(
platform
::
CPUPlace
(),
dst
,
gpu_
.
place_
,
src
,
gpu_
.
size_
,
memory
::
Copy
(
platform
::
CPUPlace
(),
dst
,
gpu_
.
place_
,
src
,
gpu_
.
size_
,
nullptr
);
stream
);
dev_ctx
->
Wait
();
}
}
void
MutableCPU
()
{
void
MutableCPU
()
{
...
@@ -260,7 +279,7 @@ class Vector {
...
@@ -260,7 +279,7 @@ class Vector {
SetFlag
(
kDataInCUDA
);
SetFlag
(
kDataInCUDA
);
}
else
if
(
IsInCUDA
()
&&
}
else
if
(
IsInCUDA
()
&&
!
(
boost
::
get
<
platform
::
CUDAPlace
>
(
place
)
==
gpu_
.
place_
))
{
!
(
boost
::
get
<
platform
::
CUDAPlace
>
(
place
)
==
gpu_
.
place_
))
{
CopyCUDADataToAnotherPlace
(
place
);
PADDLE_THROW
(
"This situation should not happen"
);
// Still dirty
// Still dirty
}
else
{
}
else
{
// Dirty && DataInCUDA && Device is same
// Dirty && DataInCUDA && Device is same
...
@@ -272,28 +291,21 @@ class Vector {
...
@@ -272,28 +291,21 @@ class Vector {
CopyCPUDataToCUDA
(
place
);
CopyCPUDataToCUDA
(
place
);
SetFlag
(
kDataInCUDA
);
SetFlag
(
kDataInCUDA
);
}
else
if
(
!
(
boost
::
get
<
platform
::
CUDAPlace
>
(
place
)
==
gpu_
.
place_
))
{
}
else
if
(
!
(
boost
::
get
<
platform
::
CUDAPlace
>
(
place
)
==
gpu_
.
place_
))
{
CopyCUDADataToAnotherPlace
(
place
);
PADDLE_THROW
(
"This situation should not happen."
);
}
else
{
}
else
{
// Not Dirty && DataInCUDA && Device is same
// Not Dirty && DataInCUDA && Device is same
// Do nothing.
// Do nothing.
}
}
}
}
}
}
void
CopyCUDADataToAnotherPlace
(
const
platform
::
Place
&
place
)
const
{
details
::
CUDABuffer
tmp
(
place
,
gpu_
.
size_
);
const
void
*
src
=
gpu_
.
data_
;
void
*
dst
=
tmp
.
data_
;
memory
::
Copy
(
tmp
.
place_
,
dst
,
gpu_
.
place_
,
src
,
gpu_
.
size_
,
nullptr
);
gpu_
.
Swap
(
tmp
);
}
void
CopyCPUDataToCUDA
(
const
platform
::
Place
&
place
)
const
{
void
CopyCPUDataToCUDA
(
const
platform
::
Place
&
place
)
const
{
void
*
src
=
cpu_
.
data
();
void
*
src
=
cpu_
.
data
();
gpu_
.
Resize
(
place
,
cpu_
.
size
()
*
sizeof
(
T
));
gpu_
.
Resize
(
place
,
cpu_
.
size
()
*
sizeof
(
T
));
void
*
dst
=
gpu_
.
data_
;
void
*
dst
=
gpu_
.
data_
;
auto
stream
=
static_cast
<
platform
::
CUDADeviceContext
*>
(
auto
*
dev_ctx
=
static_cast
<
platform
::
CUDADeviceContext
*>
(
platform
::
DeviceContextPool
::
Instance
().
Get
(
place
))
platform
::
DeviceContextPool
::
Instance
().
Get
(
place
));
->
stream
();
auto
stream
=
dev_ctx
->
stream
();
memory
::
Copy
(
gpu_
.
place_
,
dst
,
platform
::
CPUPlace
(),
src
,
gpu_
.
size_
,
memory
::
Copy
(
gpu_
.
place_
,
dst
,
platform
::
CPUPlace
(),
src
,
gpu_
.
size_
,
stream
);
stream
);
}
}
...
@@ -319,6 +331,8 @@ class Vector {
...
@@ -319,6 +331,8 @@ class Vector {
mutable
std
::
vector
<
T
>
cpu_
;
mutable
std
::
vector
<
T
>
cpu_
;
mutable
details
::
CUDABuffer
gpu_
;
mutable
details
::
CUDABuffer
gpu_
;
mutable
int
flag_
;
mutable
int
flag_
;
mutable
std
::
mutex
mtx_
;
};
};
public:
public:
...
@@ -350,81 +364,103 @@ class Vector {
...
@@ -350,81 +364,103 @@ class Vector {
Vector
(
Vector
<
T
>
&&
other
)
{
m_
=
std
::
move
(
other
.
m_
);
}
Vector
(
Vector
<
T
>
&&
other
)
{
m_
=
std
::
move
(
other
.
m_
);
}
// CPU data access method. Mutable.
// CPU data access method. Mutable.
T
&
operator
[](
size_t
i
)
{
return
(
*
m_
)[
i
];
}
T
&
operator
[](
size_t
i
)
{
return
(
*
m_
.
MutableData
()
)[
i
];
}
// CPU data access method. Immutable.
// CPU data access method. Immutable.
const
T
&
operator
[](
size_t
i
)
const
{
return
(
*
m_
)[
i
];
}
const
T
&
operator
[](
size_t
i
)
const
{
return
m_
.
Data
(
)[
i
];
}
// std::vector iterator methods. Based on CPU data access method
// std::vector iterator methods. Based on CPU data access method
size_t
size
()
const
{
return
m_
->
size
();
}
size_t
size
()
const
{
return
m_
.
Data
().
size
();
}
iterator
begin
()
{
return
m_
->
begin
();
}
iterator
begin
()
{
return
m_
.
MutableData
()
->
begin
();
}
iterator
end
()
{
return
m_
->
end
();
}
iterator
end
()
{
return
m_
.
MutableData
()
->
end
();
}
T
&
front
()
{
return
m_
->
front
();
}
T
&
front
()
{
return
m_
.
MutableData
()
->
front
();
}
T
&
back
()
{
return
m_
->
back
();
}
T
&
back
()
{
return
m_
.
MutableData
()
->
back
();
}
const_iterator
begin
()
const
{
return
m_
->
begin
();
}
const_iterator
begin
()
const
{
return
m_
.
Data
().
begin
();
}
const_iterator
end
()
const
{
return
m_
->
end
();
}
const_iterator
end
()
const
{
return
m_
.
Data
().
end
();
}
const_iterator
cbegin
()
const
{
return
begin
();
}
const_iterator
cbegin
()
const
{
return
begin
();
}
const_iterator
cend
()
const
{
return
end
();
}
const_iterator
cend
()
const
{
return
end
();
}
const
T
&
back
()
const
{
return
m_
->
back
();
}
const
T
&
back
()
const
{
return
m_
.
Data
().
back
();
}
T
*
data
()
{
return
m_
->
data
();
}
T
*
data
()
{
return
m_
.
MutableData
()
->
data
();
}
const
T
*
data
()
const
{
return
m_
->
data
();
}
const
T
*
data
()
const
{
return
m_
.
Data
().
data
();
}
const
T
&
front
()
const
{
return
m_
->
front
();
}
const
T
&
front
()
const
{
return
m_
.
Data
().
front
();
}
// end of std::vector iterator methods
// end of std::vector iterator methods
// assign this from iterator.
// assign this from iterator.
// NOTE: the iterator must support `end-begin`
// NOTE: the iterator must support `end-begin`
template
<
typename
Iter
>
template
<
typename
Iter
>
void
assign
(
Iter
begin
,
Iter
end
)
{
void
assign
(
Iter
begin
,
Iter
end
)
{
m_
->
assign
(
begin
,
end
);
m_
.
MutableData
()
->
assign
(
begin
,
end
);
}
}
// push_back. If the previous capacity is not enough, the memory will
// push_back. If the previous capacity is not enough, the memory will
// double.
// double.
void
push_back
(
T
elem
)
{
m_
->
push_back
(
elem
);
}
void
push_back
(
T
elem
)
{
m_
.
MutableData
()
->
push_back
(
elem
);
}
// extend a vector by iterator.
// extend a vector by iterator.
// NOTE: the iterator must support end-begin
// NOTE: the iterator must support end-begin
template
<
typename
It
>
template
<
typename
It
>
void
Extend
(
It
begin
,
It
end
)
{
void
Extend
(
It
begin
,
It
end
)
{
m_
->
Extend
(
begin
,
end
);
m_
.
MutableData
()
->
Extend
(
begin
,
end
);
}
}
// resize the vector
// resize the vector
void
resize
(
size_t
size
)
{
void
resize
(
size_t
size
)
{
if
(
m_
.
Data
().
size
()
!=
size
)
{
if
(
m_
.
Data
().
size
()
!=
size
)
{
m_
->
resize
(
size
);
m_
.
MutableData
()
->
resize
(
size
);
}
}
}
}
// get cuda ptr. immutable
// get cuda ptr. immutable
const
T
*
CUDAData
(
platform
::
Place
place
)
const
{
const
T
*
CUDAData
(
platform
::
Place
place
)
const
{
return
m_
.
Data
().
CUDAData
(
place
);
{
auto
&
mtx
=
m_
.
Data
().
Mutex
();
std
::
lock_guard
<
std
::
mutex
>
guard
(
mtx
);
auto
cuda_place
=
m_
.
Data
().
CUDAPlace
();
if
(
cuda_place
==
nullptr
||
*
cuda_place
==
boost
::
get
<
platform
::
CUDAPlace
>
(
place
))
{
return
m_
.
Data
().
CUDAData
(
place
);
}
}
// If m_ contains CUDAData in a different place. Detach manually.
m_
.
Detach
();
return
CUDAData
(
place
);
}
}
// get cuda ptr. mutable
// get cuda ptr. mutable
T
*
CUDAMutableData
(
platform
::
Place
place
)
{
T
*
CUDAMutableData
(
platform
::
Place
place
)
{
return
m_
->
CUDAMutableData
(
place
);
{
auto
&
mtx
=
m_
.
Data
().
Mutex
();
std
::
lock_guard
<
std
::
mutex
>
guard
(
mtx
);
auto
cuda_place
=
m_
.
Data
().
CUDAPlace
();
if
(
cuda_place
==
nullptr
||
*
cuda_place
==
boost
::
get
<
platform
::
CUDAPlace
>
(
place
))
{
return
m_
.
MutableData
()
->
CUDAMutableData
(
place
);
}
}
// If m_ contains CUDAData in a different place. Detach manually.
m_
.
Detach
();
return
CUDAMutableData
(
place
);
}
}
// clear
// clear
void
clear
()
{
m_
->
clear
();
}
void
clear
()
{
m_
.
MutableData
()
->
clear
();
}
size_t
capacity
()
const
{
return
m_
->
capacity
();
}
size_t
capacity
()
const
{
return
m_
.
Data
().
capacity
();
}
// reserve data
// reserve data
void
reserve
(
size_t
size
)
{
m_
->
reserve
(
size
);
}
void
reserve
(
size_t
size
)
{
m_
.
Data
().
reserve
(
size
);
}
// the unify method to access CPU or CUDA data. immutable.
// the unify method to access CPU or CUDA data. immutable.
const
T
*
Data
(
platform
::
Place
place
)
const
{
const
T
*
Data
(
platform
::
Place
place
)
const
{
...
@@ -445,7 +481,7 @@ class Vector {
...
@@ -445,7 +481,7 @@ class Vector {
}
}
// implicit cast operator. Vector can be cast to std::vector implicitly.
// implicit cast operator. Vector can be cast to std::vector implicitly.
operator
std
::
vector
<
T
>
()
const
{
return
*
m_
;
}
operator
std
::
vector
<
T
>
()
const
{
return
m_
.
Data
()
;
}
bool
operator
==
(
const
Vector
<
T
>
&
other
)
const
{
bool
operator
==
(
const
Vector
<
T
>
&
other
)
const
{
if
(
size
()
!=
other
.
size
())
return
false
;
if
(
size
()
!=
other
.
size
())
return
false
;
...
@@ -463,7 +499,7 @@ class Vector {
...
@@ -463,7 +499,7 @@ class Vector {
private:
private:
// Vector is an COW object.
// Vector is an COW object.
details
::
COWPtr
<
VectorData
>
m_
;
mutable
details
::
COWPtr
<
VectorData
>
m_
;
};
};
#else // PADDLE_WITH_CUDA
#else // PADDLE_WITH_CUDA
...
...
paddle/fluid/operators/lookup_table_op.cu
浏览文件 @
e1913bc5
...
@@ -127,10 +127,8 @@ class LookupTableGradCUDAKernel : public framework::OpKernel<T> {
...
@@ -127,10 +127,8 @@ class LookupTableGradCUDAKernel : public framework::OpKernel<T> {
auto
gpu_place
=
boost
::
get
<
platform
::
CUDAPlace
>
(
context
.
GetPlace
());
auto
gpu_place
=
boost
::
get
<
platform
::
CUDAPlace
>
(
context
.
GetPlace
());
// TODO(yuyang18): Strange code here.
// TODO(yuyang18): Strange code here.
memory
::
Copy
(
platform
::
CPUPlace
(),
memory
::
Copy
(
gpu_place
,
new_rows
.
CUDAMutableData
(
context
.
GetPlace
()),
new_rows
.
CUDAMutableData
(
context
.
GetPlace
()),
gpu_place
,
gpu_place
,
ids_data
,
ids_num
*
sizeof
(
int64_t
),
stream
);
ids_data
,
ids_num
*
sizeof
(
int64_t
),
stream
);
d_table
->
set_rows
(
new_rows
);
d_table
->
set_rows
(
new_rows
);
auto
*
d_table_value
=
d_table
->
mutable_value
();
auto
*
d_table_value
=
d_table
->
mutable_value
();
...
...
paddle/fluid/operators/sgd_op.cu
浏览文件 @
e1913bc5
...
@@ -89,7 +89,7 @@ class SGDOpCUDAKernel : public framework::OpKernel<T> {
...
@@ -89,7 +89,7 @@ class SGDOpCUDAKernel : public framework::OpKernel<T> {
PADDLE_ENFORCE_EQ
(
in_height
,
out_dims
[
0
]);
PADDLE_ENFORCE_EQ
(
in_height
,
out_dims
[
0
]);
auto
&
in_value
=
grad
->
value
();
auto
&
in_value
=
grad
->
value
();
framework
::
Vector
<
int64_t
>
in_rows
(
grad
->
rows
()
);
auto
&
in_rows
=
grad
->
rows
(
);
int64_t
in_row_numel
=
in_value
.
numel
()
/
in_rows
.
size
();
int64_t
in_row_numel
=
in_value
.
numel
()
/
in_rows
.
size
();
PADDLE_ENFORCE_EQ
(
in_row_numel
,
param_out
->
numel
()
/
in_height
);
PADDLE_ENFORCE_EQ
(
in_row_numel
,
param_out
->
numel
()
/
in_height
);
...
...
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