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
机器未来
Paddle
提交
d558b8bb
P
Paddle
项目概览
机器未来
/
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
d558b8bb
编写于
6月 21, 2017
作者:
H
hedaoyuan
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Move the code in the GemmConvOpGpu.cu file into Im2ColOpGpu.cu.
上级
69271c92
变更
1
隐藏空白更改
内联
并排
Showing
1 changed file
with
172 addition
and
0 deletion
+172
-0
paddle/function/Im2ColOpGpu.cu
paddle/function/Im2ColOpGpu.cu
+172
-0
未找到文件。
paddle/function/Im2ColOpGpu.cu
浏览文件 @
d558b8bb
...
...
@@ -17,6 +17,178 @@ limitations under the License. */
namespace
paddle
{
template
<
class
T
>
__global__
void
im2col
(
const
T
*
data_im
,
int
numOuts
,
int
height
,
int
width
,
int
blockH
,
int
blockW
,
int
strideH
,
int
strideW
,
int
paddingH
,
int
paddingW
,
int
height_col
,
int
width_col
,
T
*
data_col
)
{
int
index
=
(
blockIdx
.
x
*
gridDim
.
y
+
blockIdx
.
y
)
*
blockDim
.
x
+
threadIdx
.
x
;
if
(
index
<
numOuts
)
{
int
w_out
=
index
%
width_col
;
index
/=
width_col
;
int
h_out
=
index
%
height_col
;
int
channel_in
=
index
/
height_col
;
int
channel_out
=
channel_in
*
blockH
*
blockW
;
int
h_in
=
h_out
*
strideH
;
int
w_in
=
w_out
*
strideW
;
data_col
+=
(
channel_out
*
height_col
+
h_out
)
*
width_col
+
w_out
;
for
(
int
i
=
0
;
i
<
blockH
;
++
i
)
{
for
(
int
j
=
0
;
j
<
blockW
;
++
j
)
{
int
rIdx
=
int
(
h_in
+
i
);
int
cIdx
=
int
(
w_in
+
j
);
if
((
rIdx
-
(
int
)
paddingH
)
>=
(
int
)
height
||
(
rIdx
-
(
int
)
paddingH
)
<
0
||
(
cIdx
-
(
int
)
paddingW
)
>=
(
int
)
width
||
(
cIdx
-
(
int
)
paddingW
)
<
0
)
{
*
data_col
=
0
;
}
else
{
rIdx
=
rIdx
+
channel_in
*
height
-
paddingH
;
cIdx
=
cIdx
-
paddingW
;
*
data_col
=
data_im
[
rIdx
*
width
+
cIdx
];
}
data_col
+=
height_col
*
width_col
;
}
}
}
}
template
<
class
T
>
class
Im2ColFunctor
<
kCFO
,
DEVICE_TYPE_GPU
,
T
>
{
public:
void
operator
()(
const
T
*
imData
,
const
TensorShape
&
imShape
,
T
*
colData
,
const
TensorShape
&
colShape
,
int
strideHeight
,
int
strideWidth
,
int
paddingHeight
,
int
paddingWidth
)
{
int
inputChannels
=
imShape
[
0
];
int
inputHeight
=
imShape
[
1
];
int
inputWidth
=
imShape
[
2
];
int
filterHeight
=
colShape
[
3
];
int
filterWidth
=
colShape
[
4
];
int
outputHeight
=
colShape
[
0
];
int
outputWidth
=
colShape
[
1
];
int
numKernels
=
inputChannels
*
outputHeight
*
outputWidth
;
int
blocks
=
(
numKernels
+
1024
-
1
)
/
1024
;
int
blockX
=
512
;
int
blockY
=
(
blocks
+
512
-
1
)
/
512
;
dim3
threads
(
1024
,
1
);
dim3
grid
(
blockX
,
blockY
);
im2col
<
T
><<<
grid
,
threads
,
0
,
STREAM_DEFAULT
>>>
(
imData
,
numKernels
,
inputHeight
,
inputWidth
,
filterHeight
,
filterWidth
,
strideHeight
,
strideWidth
,
paddingHeight
,
paddingWidth
,
outputHeight
,
outputWidth
,
colData
);
CHECK_SYNC
(
"Im2ColFunctor GPU failed"
);
}
};
template
<
class
T
>
__global__
void
col2im
(
size_t
n
,
const
T
*
data_col
,
size_t
height
,
size_t
width
,
size_t
channels
,
size_t
blockH
,
size_t
blockW
,
size_t
strideH
,
size_t
strideW
,
size_t
paddingH
,
size_t
paddingW
,
size_t
height_col
,
size_t
width_col
,
T
*
data_im
)
{
size_t
index
=
(
blockIdx
.
x
*
gridDim
.
y
+
blockIdx
.
y
)
*
blockDim
.
x
+
threadIdx
.
x
;
if
(
index
<
n
)
{
T
val
=
0
;
int
w
=
int
(
index
%
width
);
int
h
=
int
((
index
/
width
)
%
height
);
int
c
=
int
(
index
/
(
width
*
height
));
if
((
w
-
(
int
)
paddingW
)
>=
0
&&
(
w
-
(
int
)
paddingW
)
<
(
width
-
2
*
paddingW
)
&&
(
h
-
(
int
)
paddingH
)
>=
0
&&
(
h
-
paddingH
)
<
(
height
-
2
*
paddingH
))
{
// compute the start and end of the output
int
w_col_start
=
(
w
<
(
int
)
blockW
)
?
0
:
(
w
-
int
(
blockW
))
/
(
int
)
strideW
+
1
;
int
w_col_end
=
min
((
int
)(
w
/
(
int
)
strideW
+
1
),
(
int
)(
width_col
));
int
h_col_start
=
(
h
<
(
int
)
blockH
)
?
0
:
(
h
-
(
int
)
blockH
)
/
(
int
)
strideH
+
1
;
int
h_col_end
=
min
(
int
(
h
/
strideH
+
1
),
int
(
height_col
));
for
(
int
h_col
=
h_col_start
;
h_col
<
h_col_end
;
++
h_col
)
{
for
(
int
w_col
=
w_col_start
;
w_col
<
w_col_end
;
++
w_col
)
{
// the col location: [c * width * height + h_out, w_out]
int
c_col
=
int
(
c
*
blockH
*
blockW
)
+
\
(
h
-
h_col
*
(
int
)
strideH
)
*
(
int
)
blockW
+
(
w
-
w_col
*
(
int
)
strideW
);
val
+=
data_col
[(
c_col
*
height_col
+
h_col
)
*
width_col
+
w_col
];
}
}
h
-=
paddingH
;
w
-=
paddingW
;
data_im
[
c
*
((
width
-
2
*
paddingW
)
*
(
height
-
2
*
paddingH
))
+
h
*
(
width
-
2
*
paddingW
)
+
w
]
+=
val
;
}
}
}
template
<
class
T
>
class
Col2ImFunctor
<
kCFO
,
DEVICE_TYPE_GPU
,
T
>
{
public:
void
operator
()(
T
*
imData
,
const
TensorShape
&
imShape
,
const
T
*
colData
,
const
TensorShape
&
colShape
,
int
strideHeight
,
int
strideWidth
,
int
paddingHeight
,
int
paddingWidth
)
{
int
inputChannels
=
imShape
[
0
];
int
inputHeight
=
imShape
[
1
];
int
inputWidth
=
imShape
[
2
];
int
filterHeight
=
colShape
[
3
];
int
filterWidth
=
colShape
[
4
];
int
outputHeight
=
colShape
[
0
];
int
outputWidth
=
colShape
[
1
];
size_t
numKernels
=
inputChannels
*
(
inputHeight
+
2
*
paddingHeight
)
*
(
inputWidth
+
2
*
paddingWidth
);
size_t
blocks
=
(
numKernels
+
1024
-
1
)
/
1024
;
size_t
blockX
=
512
;
size_t
blockY
=
(
blocks
+
512
-
1
)
/
512
;
dim3
threads
(
1024
,
1
);
dim3
grid
(
blockX
,
blockY
);
// To avoid involving atomic operations, we will launch one kernel per
// bottom dimension, and then in the kernel add up the top dimensions.
col2im
<
T
><<<
grid
,
threads
,
0
,
STREAM_DEFAULT
>>>
(
numKernels
,
colData
,
inputHeight
+
2
*
paddingHeight
,
inputWidth
+
2
*
paddingWidth
,
inputChannels
,
filterHeight
,
filterWidth
,
strideHeight
,
strideWidth
,
paddingHeight
,
paddingWidth
,
outputHeight
,
outputWidth
,
imData
);
CHECK_SYNC
(
"Col2ImFunctor GPU failed"
);
}
};
template
class
Im2ColFunctor
<
kCFO
,
DEVICE_TYPE_GPU
,
float
>;
template
class
Im2ColFunctor
<
kCFO
,
DEVICE_TYPE_GPU
,
double
>;
template
class
Col2ImFunctor
<
kCFO
,
DEVICE_TYPE_GPU
,
float
>;
template
class
Col2ImFunctor
<
kCFO
,
DEVICE_TYPE_GPU
,
double
>;
template
<
class
T
>
__global__
void
im2colOCF
(
const
T
*
imData
,
T
*
colData
,
...
...
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