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
BaiXuePrincess
Paddle
提交
aae994fd
P
Paddle
项目概览
BaiXuePrincess
/
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
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
提交
aae994fd
编写于
7月 24, 2018
作者:
T
tensor-tang
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
refine im2col no padding
上级
03d70c1f
变更
1
隐藏空白更改
内联
并排
Showing
1 changed file
with
29 addition
and
5 deletion
+29
-5
paddle/fluid/operators/math/im2col.cc
paddle/fluid/operators/math/im2col.cc
+29
-5
未找到文件。
paddle/fluid/operators/math/im2col.cc
浏览文件 @
aae994fd
...
@@ -40,22 +40,46 @@ class Im2ColFunctor<paddle::operators::math::ColFormat::kCFO,
...
@@ -40,22 +40,46 @@ class Im2ColFunctor<paddle::operators::math::ColFormat::kCFO,
int
im_width
=
im
.
dims
()[
2
];
int
im_width
=
im
.
dims
()[
2
];
int
filter_height
=
col
->
dims
()[
1
];
int
filter_height
=
col
->
dims
()[
1
];
int
filter_width
=
col
->
dims
()[
2
];
int
filter_width
=
col
->
dims
()[
2
];
int
col
_height
=
col
->
dims
()[
3
];
int
output
_height
=
col
->
dims
()[
3
];
int
col
_width
=
col
->
dims
()[
4
];
int
output
_width
=
col
->
dims
()[
4
];
int
channels_col
=
im_channels
*
filter_height
*
filter_width
;
int
channels_col
=
im_channels
*
filter_height
*
filter_width
;
const
T
*
im_data
=
im
.
data
<
T
>
();
const
T
*
im_data
=
im
.
data
<
T
>
();
T
*
col_data
=
col
->
data
<
T
>
();
T
*
col_data
=
col
->
data
<
T
>
();
// TODO(TJ): change me to template
// further optimaze:
// 1. padding != 1
// 2. could also support stride_h != 1
if
(
stride
[
0
]
==
1
&&
stride
[
1
]
==
1
&&
dilation
[
0
]
==
1
&&
dilation
[
1
]
==
1
&&
padding
[
0
]
==
0
&&
padding
[
1
]
==
0
)
{
int
col_matrix_width
=
output_width
*
output_height
;
for
(
int
oh
=
0
;
oh
<
output_height
;
++
oh
)
{
const
T
*
im_data_start
=
im_data
+
oh
*
im_width
;
T
*
dst_data
=
col_data
+
oh
*
output_width
;
for
(
int
ic
=
0
;
ic
<
im_channels
;
++
ic
)
{
const
T
*
src_data
=
im_data_start
+
ic
*
im_height
*
im_width
;
for
(
int
kh
=
0
;
kh
<
filter_height
;
++
kh
)
{
for
(
int
kw
=
0
;
kw
<
filter_width
;
++
kw
)
{
std
::
memcpy
(
dst_data
,
src_data
+
kw
,
sizeof
(
T
)
*
output_width
);
dst_data
=
dst_data
+
col_matrix_width
;
}
src_data
=
src_data
+
im_width
;
}
}
}
return
;
}
for
(
int
c
=
0
;
c
<
channels_col
;
++
c
)
{
for
(
int
c
=
0
;
c
<
channels_col
;
++
c
)
{
int
w_offset
=
c
%
filter_width
;
int
w_offset
=
c
%
filter_width
;
int
h_offset
=
(
c
/
filter_width
)
%
filter_height
;
int
h_offset
=
(
c
/
filter_width
)
%
filter_height
;
int
c_im
=
c
/
(
filter_width
*
filter_height
);
int
c_im
=
c
/
(
filter_width
*
filter_height
);
for
(
int
h
=
0
;
h
<
col
_height
;
++
h
)
{
for
(
int
h
=
0
;
h
<
output
_height
;
++
h
)
{
int
im_row_idx
=
h
*
stride
[
0
]
-
padding
[
0
]
+
h_offset
*
dilation
[
0
];
int
im_row_idx
=
h
*
stride
[
0
]
-
padding
[
0
]
+
h_offset
*
dilation
[
0
];
for
(
int
w
=
0
;
w
<
col
_width
;
++
w
)
{
for
(
int
w
=
0
;
w
<
output
_width
;
++
w
)
{
int
im_col_idx
=
w
*
stride
[
1
]
-
padding
[
1
]
+
w_offset
*
dilation
[
1
];
int
im_col_idx
=
w
*
stride
[
1
]
-
padding
[
1
]
+
w_offset
*
dilation
[
1
];
int
col_idx
=
(
c
*
col_height
+
h
)
*
col
_width
+
w
;
int
col_idx
=
(
c
*
output_height
+
h
)
*
output
_width
+
w
;
int
im_idx
=
(
im_row_idx
+
c_im
*
im_height
)
*
im_width
+
im_col_idx
;
int
im_idx
=
(
im_row_idx
+
c_im
*
im_height
)
*
im_width
+
im_col_idx
;
col_data
[
col_idx
]
=
(
im_row_idx
<
0
||
im_row_idx
>=
im_height
||
col_data
[
col_idx
]
=
(
im_row_idx
<
0
||
im_row_idx
>=
im_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