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
urhero
Paddle
提交
fb52bc6e
P
Paddle
项目概览
urhero
/
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
fb52bc6e
编写于
9月 25, 2017
作者:
Y
Yibing Liu
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
revert code layout in multiplex_op
上级
7620efdf
变更
3
显示空白变更内容
内联
并排
Showing
3 changed file
with
94 addition
and
64 deletion
+94
-64
paddle/operators/multiplex_op.cc
paddle/operators/multiplex_op.cc
+3
-3
paddle/operators/multiplex_op.cu
paddle/operators/multiplex_op.cu
+74
-3
paddle/operators/multiplex_op.h
paddle/operators/multiplex_op.h
+17
-58
未找到文件。
paddle/operators/multiplex_op.cc
浏览文件 @
fb52bc6e
...
@@ -106,8 +106,8 @@ namespace ops = paddle::operators;
...
@@ -106,8 +106,8 @@ namespace ops = paddle::operators;
REGISTER_OP
(
multiplex
,
ops
::
MultiplexOp
,
ops
::
MultiplexOpMaker
,
multiplex_grad
,
REGISTER_OP
(
multiplex
,
ops
::
MultiplexOp
,
ops
::
MultiplexOpMaker
,
multiplex_grad
,
ops
::
MultiplexGradOp
);
ops
::
MultiplexGradOp
);
REGISTER_OP_CPU_KERNEL
(
multiplex
,
REGISTER_OP_CPU_KERNEL
(
ops
::
Multiplex
Kernel
<
paddle
::
platform
::
CPUPlace
,
float
>
);
multiplex
,
ops
::
MultiplexCPU
Kernel
<
paddle
::
platform
::
CPUPlace
,
float
>
);
REGISTER_OP_CPU_KERNEL
(
REGISTER_OP_CPU_KERNEL
(
multiplex_grad
,
multiplex_grad
,
ops
::
MultiplexGradKernel
<
paddle
::
platform
::
CPUPlace
,
float
>
);
ops
::
MultiplexGrad
CPU
Kernel
<
paddle
::
platform
::
CPUPlace
,
float
>
);
paddle/operators/multiplex_op.cu
浏览文件 @
fb52bc6e
...
@@ -15,10 +15,81 @@
...
@@ -15,10 +15,81 @@
#include "paddle/framework/op_registry.h"
#include "paddle/framework/op_registry.h"
#include "paddle/operators/multiplex_op.h"
#include "paddle/operators/multiplex_op.h"
namespace
paddle
{
namespace
operators
{
template
<
typename
Place
,
typename
T
>
class
MultiplexGPUKernel
:
public
framework
::
OpKernel
{
public:
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
{
auto
ins
=
ctx
.
MultiInput
<
framework
::
Tensor
>
(
"X"
);
auto
*
out
=
ctx
.
Output
<
framework
::
LoDTensor
>
(
"Out"
);
out
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
auto
rows
=
ins
[
1
]
->
dims
()[
0
];
auto
cols
=
ins
[
1
]
->
dims
()[
1
];
// copy index to cpu
framework
::
Tensor
index_t_cpu
;
index_t_cpu
.
CopyFrom
<
T
>
(
*
(
ins
[
0
]),
platform
::
CPUPlace
());
auto
*
index
=
index_t_cpu
.
data
<
T
>
();
auto
stream
=
reinterpret_cast
<
const
platform
::
CUDADeviceContext
&>
(
ctx
.
device_context
())
.
stream
();
Place
place
=
boost
::
get
<
Place
>
(
ctx
.
GetPlace
());
for
(
auto
i
=
0
;
i
<
rows
;
i
++
)
{
int
k
=
(
int
)
index
[
i
]
+
1
;
PADDLE_ENFORCE_LT
(
k
,
ins
.
size
(),
"index exceeds the number of candidate tensors."
);
memory
::
Copy
(
place
,
out
->
data
<
T
>
()
+
i
*
cols
,
place
,
ins
[
k
]
->
data
<
T
>
()
+
i
*
cols
,
cols
*
sizeof
(
T
),
stream
);
}
}
};
template
<
typename
Place
,
typename
T
>
class
MultiplexGradGPUKernel
:
public
framework
::
OpKernel
{
public:
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
{
auto
*
d_out
=
ctx
.
Input
<
framework
::
Tensor
>
(
framework
::
GradVarName
(
"Out"
));
auto
ins
=
ctx
.
MultiInput
<
framework
::
Tensor
>
(
"X"
);
auto
d_ins
=
ctx
.
MultiOutput
<
framework
::
Tensor
>
(
framework
::
GradVarName
(
"X"
));
for
(
size_t
i
=
1
;
i
<
d_ins
.
size
();
i
++
)
{
if
(
d_ins
[
i
])
{
d_ins
[
i
]
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
auto
t
=
framework
::
EigenVector
<
T
>::
Flatten
(
*
d_ins
[
i
]);
t
.
device
(
ctx
.
GetEigenDevice
<
Place
>
())
=
t
.
constant
(
static_cast
<
T
>
(
0
));
}
}
auto
rows
=
ins
[
1
]
->
dims
()[
0
];
auto
cols
=
ins
[
1
]
->
dims
()[
1
];
// copy index to cpu
framework
::
Tensor
index_t_cpu
;
index_t_cpu
.
CopyFrom
<
T
>
(
*
(
ins
[
0
]),
platform
::
CPUPlace
());
auto
*
index
=
index_t_cpu
.
data
<
T
>
();
auto
stream
=
reinterpret_cast
<
const
platform
::
CUDADeviceContext
&>
(
ctx
.
device_context
())
.
stream
();
Place
place
=
boost
::
get
<
Place
>
(
ctx
.
GetPlace
());
for
(
auto
i
=
0
;
i
<
rows
;
i
++
)
{
int
k
=
(
int
)
index
[
i
]
+
1
;
if
(
d_ins
[
k
])
{
memory
::
Copy
(
place
,
d_ins
[
k
]
->
data
<
T
>
()
+
i
*
cols
,
place
,
d_out
->
data
<
T
>
()
+
i
*
cols
,
cols
*
sizeof
(
T
),
stream
);
}
}
}
};
}
// namespace operators
}
// namespace paddle
namespace
ops
=
paddle
::
operators
;
namespace
ops
=
paddle
::
operators
;
REGISTER_OP_GPU_KERNEL
(
multiplex
,
REGISTER_OP_GPU_KERNEL
(
ops
::
Multiplex
Kernel
<
paddle
::
platform
::
GPUPlace
,
float
>
);
multiplex
,
ops
::
MultiplexGPU
Kernel
<
paddle
::
platform
::
GPUPlace
,
float
>
);
REGISTER_OP_GPU_KERNEL
(
REGISTER_OP_GPU_KERNEL
(
multiplex_grad
,
multiplex_grad
,
ops
::
MultiplexGradKernel
<
paddle
::
platform
::
GPUPlace
,
float
>
);
ops
::
MultiplexGrad
GPU
Kernel
<
paddle
::
platform
::
GPUPlace
,
float
>
);
paddle/operators/multiplex_op.h
浏览文件 @
fb52bc6e
...
@@ -23,7 +23,7 @@ namespace paddle {
...
@@ -23,7 +23,7 @@ namespace paddle {
namespace
operators
{
namespace
operators
{
template
<
typename
Place
,
typename
T
>
template
<
typename
Place
,
typename
T
>
class
MultiplexKernel
:
public
framework
::
OpKernel
{
class
Multiplex
CPU
Kernel
:
public
framework
::
OpKernel
{
public:
public:
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
{
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
{
auto
ins
=
ctx
.
MultiInput
<
framework
::
Tensor
>
(
"X"
);
auto
ins
=
ctx
.
MultiInput
<
framework
::
Tensor
>
(
"X"
);
...
@@ -33,9 +33,8 @@ class MultiplexKernel : public framework::OpKernel {
...
@@ -33,9 +33,8 @@ class MultiplexKernel : public framework::OpKernel {
auto
rows
=
ins
[
1
]
->
dims
()[
0
];
auto
rows
=
ins
[
1
]
->
dims
()[
0
];
auto
cols
=
ins
[
1
]
->
dims
()[
1
];
auto
cols
=
ins
[
1
]
->
dims
()[
1
];
if
(
platform
::
is_cpu_place
(
ctx
.
GetPlace
()))
{
auto
*
index
=
ins
[
0
]
->
data
<
T
>
();
auto
*
index
=
ins
[
0
]
->
data
<
T
>
();
platform
::
CPUPlace
place
=
boost
::
get
<
platform
::
CPU
Place
>
(
ctx
.
GetPlace
());
Place
place
=
boost
::
get
<
Place
>
(
ctx
.
GetPlace
());
for
(
auto
i
=
0
;
i
<
rows
;
i
++
)
{
for
(
auto
i
=
0
;
i
<
rows
;
i
++
)
{
int
k
=
(
int
)
index
[
i
]
+
1
;
int
k
=
(
int
)
index
[
i
]
+
1
;
PADDLE_ENFORCE_LT
(
k
,
ins
.
size
(),
PADDLE_ENFORCE_LT
(
k
,
ins
.
size
(),
...
@@ -43,30 +42,11 @@ class MultiplexKernel : public framework::OpKernel {
...
@@ -43,30 +42,11 @@ class MultiplexKernel : public framework::OpKernel {
memory
::
Copy
(
place
,
out
->
data
<
T
>
()
+
i
*
cols
,
place
,
memory
::
Copy
(
place
,
out
->
data
<
T
>
()
+
i
*
cols
,
place
,
ins
[
k
]
->
data
<
T
>
()
+
i
*
cols
,
cols
*
sizeof
(
T
));
ins
[
k
]
->
data
<
T
>
()
+
i
*
cols
,
cols
*
sizeof
(
T
));
}
}
}
else
{
#ifndef PADDLE_ONLY_CPU
// copy index to cpu
framework
::
Tensor
index_t_cpu
;
index_t_cpu
.
CopyFrom
<
T
>
(
*
(
ins
[
0
]),
platform
::
CPUPlace
());
auto
*
index
=
index_t_cpu
.
data
<
T
>
();
auto
stream
=
reinterpret_cast
<
const
platform
::
CUDADeviceContext
&>
(
ctx
.
device_context
())
.
stream
();
platform
::
GPUPlace
place
=
boost
::
get
<
platform
::
GPUPlace
>
(
ctx
.
GetPlace
());
for
(
auto
i
=
0
;
i
<
rows
;
i
++
)
{
int
k
=
(
int
)
index
[
i
]
+
1
;
PADDLE_ENFORCE_LT
(
k
,
ins
.
size
(),
"index exceeds the number of candidate tensors."
);
memory
::
Copy
(
place
,
out
->
data
<
T
>
()
+
i
*
cols
,
place
,
ins
[
k
]
->
data
<
T
>
()
+
i
*
cols
,
cols
*
sizeof
(
T
),
stream
);
}
#endif
}
}
}
};
};
template
<
typename
Place
,
typename
T
>
template
<
typename
Place
,
typename
T
>
class
MultiplexGradKernel
:
public
framework
::
OpKernel
{
class
MultiplexGrad
CPU
Kernel
:
public
framework
::
OpKernel
{
public:
public:
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
{
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
{
auto
*
d_out
=
ctx
.
Input
<
framework
::
Tensor
>
(
framework
::
GradVarName
(
"Out"
));
auto
*
d_out
=
ctx
.
Input
<
framework
::
Tensor
>
(
framework
::
GradVarName
(
"Out"
));
...
@@ -83,9 +63,8 @@ class MultiplexGradKernel : public framework::OpKernel {
...
@@ -83,9 +63,8 @@ class MultiplexGradKernel : public framework::OpKernel {
auto
rows
=
ins
[
1
]
->
dims
()[
0
];
auto
rows
=
ins
[
1
]
->
dims
()[
0
];
auto
cols
=
ins
[
1
]
->
dims
()[
1
];
auto
cols
=
ins
[
1
]
->
dims
()[
1
];
if
(
platform
::
is_cpu_place
(
ctx
.
GetPlace
()))
{
auto
*
index
=
ins
[
0
]
->
data
<
T
>
();
auto
*
index
=
ins
[
0
]
->
data
<
T
>
();
platform
::
CPUPlace
place
=
boost
::
get
<
platform
::
CPU
Place
>
(
ctx
.
GetPlace
());
Place
place
=
boost
::
get
<
Place
>
(
ctx
.
GetPlace
());
for
(
auto
i
=
0
;
i
<
rows
;
i
++
)
{
for
(
auto
i
=
0
;
i
<
rows
;
i
++
)
{
int
k
=
(
int
)
index
[
i
]
+
1
;
int
k
=
(
int
)
index
[
i
]
+
1
;
if
(
d_ins
[
k
])
{
if
(
d_ins
[
k
])
{
...
@@ -93,26 +72,6 @@ class MultiplexGradKernel : public framework::OpKernel {
...
@@ -93,26 +72,6 @@ class MultiplexGradKernel : public framework::OpKernel {
d_out
->
data
<
T
>
()
+
i
*
cols
,
cols
*
sizeof
(
T
));
d_out
->
data
<
T
>
()
+
i
*
cols
,
cols
*
sizeof
(
T
));
}
}
}
}
}
else
{
#ifndef PADDLE_ONLY_CPU
// copy index to cpu
framework
::
Tensor
index_t_cpu
;
index_t_cpu
.
CopyFrom
<
T
>
(
*
(
ins
[
0
]),
platform
::
CPUPlace
());
auto
*
index
=
index_t_cpu
.
data
<
T
>
();
auto
stream
=
reinterpret_cast
<
const
platform
::
CUDADeviceContext
&>
(
ctx
.
device_context
())
.
stream
();
platform
::
GPUPlace
place
=
boost
::
get
<
platform
::
GPUPlace
>
(
ctx
.
GetPlace
());
for
(
auto
i
=
0
;
i
<
rows
;
i
++
)
{
int
k
=
(
int
)
index
[
i
]
+
1
;
if
(
d_ins
[
k
])
{
memory
::
Copy
(
place
,
d_ins
[
k
]
->
data
<
T
>
()
+
i
*
cols
,
place
,
d_out
->
data
<
T
>
()
+
i
*
cols
,
cols
*
sizeof
(
T
),
stream
);
}
}
#endif
}
}
}
};
};
}
}
...
...
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