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
wzpei9
Paddle
提交
07cde439
P
Paddle
项目概览
wzpei9
/
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
体验新版 GitCode,发现更多精彩内容 >>
提交
07cde439
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
Reconstruction of GemmConv Based on new im2col.
上级
eb0c7e5e
变更
1
隐藏空白更改
内联
并排
Showing
1 changed file
with
48 addition
and
137 deletion
+48
-137
paddle/function/GemmConvOp.cpp
paddle/function/GemmConvOp.cpp
+48
-137
未找到文件。
paddle/function/GemmConvOp.cpp
浏览文件 @
07cde439
...
...
@@ -12,101 +12,13 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "
Gemm
ConvOp.h"
#include "ConvOp.h"
#include "GemmFunctor.h"
#include "Im2Col.h"
#include "paddle/math/MemoryHandle.h"
namespace
paddle
{
/*
* imData = [input_channels, input_height, input_width]
* colData = [input_channels, filter_height, filter_width,
* output_height, output_width]
*/
template
<
class
T
>
class
Im2ColFunctor
<
DEVICE_TYPE_CPU
,
T
>
{
public:
void
operator
()(
const
T
*
imData
,
int
inputChannels
,
int
inputHeight
,
int
inputWidth
,
int
filterHeight
,
int
filterWidth
,
int
strideHeight
,
int
strideWidth
,
int
paddingHeight
,
int
paddingWidth
,
int
outputHeight
,
int
outputWidth
,
T
*
colData
)
{
int
channelsCol
=
inputChannels
*
filterHeight
*
filterWidth
;
for
(
int
c
=
0
;
c
<
channelsCol
;
++
c
)
{
int
wOffset
=
c
%
filterWidth
;
int
hOffset
=
(
c
/
filterWidth
)
%
filterHeight
;
int
c_im
=
c
/
filterWidth
/
filterHeight
;
for
(
int
h
=
0
;
h
<
outputHeight
;
++
h
)
{
for
(
int
w
=
0
;
w
<
outputWidth
;
++
w
)
{
int
imRowIdx
=
h
*
strideHeight
+
hOffset
;
int
imColIdx
=
w
*
strideWidth
+
wOffset
;
if
((
imRowIdx
-
paddingHeight
)
<
0
||
(
imRowIdx
-
paddingHeight
)
>=
inputHeight
||
(
imColIdx
-
paddingWidth
)
<
0
||
(
imColIdx
-
paddingWidth
)
>=
inputWidth
)
{
colData
[(
c
*
outputHeight
+
h
)
*
outputWidth
+
w
]
=
T
(
0
);
}
else
{
imRowIdx
+=
c_im
*
inputHeight
-
paddingHeight
;
imColIdx
-=
paddingWidth
;
colData
[(
c
*
outputHeight
+
h
)
*
outputWidth
+
w
]
=
imData
[
imRowIdx
*
inputWidth
+
imColIdx
];
}
}
}
}
}
};
template
<
class
T
>
class
Col2ImFunctor
<
DEVICE_TYPE_CPU
,
T
>
{
public:
void
operator
()(
const
T
*
colData
,
int
inputChannels
,
int
inputHeight
,
int
inputWidth
,
int
filterHeight
,
int
filterWidth
,
int
strideHeight
,
int
strideWidth
,
int
paddingHeight
,
int
paddingWidth
,
int
outputHeight
,
int
outputWidth
,
T
*
imData
)
{
int
channelsCol
=
inputChannels
*
filterHeight
*
filterWidth
;
for
(
int
c
=
0
;
c
<
channelsCol
;
++
c
)
{
int
wOffset
=
c
%
filterWidth
;
int
hOffset
=
(
c
/
filterWidth
)
%
filterHeight
;
int
c_im
=
c
/
filterWidth
/
filterHeight
;
for
(
int
h
=
0
;
h
<
outputHeight
;
++
h
)
{
for
(
int
w
=
0
;
w
<
outputWidth
;
++
w
)
{
int
imRowIdx
=
h
*
strideHeight
+
hOffset
;
int
imColIdx
=
w
*
strideWidth
+
wOffset
;
if
((
imRowIdx
-
paddingHeight
)
>=
0
&&
(
imRowIdx
-
paddingHeight
)
<
inputHeight
&&
(
imColIdx
-
paddingWidth
)
>=
0
&&
(
imColIdx
-
paddingWidth
)
<
inputWidth
)
{
imRowIdx
+=
c_im
*
inputHeight
-
paddingHeight
;
imColIdx
-=
paddingWidth
;
imData
[
imRowIdx
*
inputWidth
+
imColIdx
]
+=
colData
[(
c
*
outputHeight
+
h
)
*
outputWidth
+
w
];
}
}
}
}
}
};
/*
* \brief Forward calculation of convolution.
*/
...
...
@@ -155,15 +67,20 @@ public:
real
*
inputData
=
inputs
[
0
].
data
<
real
>
();
real
*
filterData
=
inputs
[
1
].
data
<
real
>
();
real
*
outputData
=
outputs
[
0
].
data
<
real
>
();
size_t
size
=
inputChannels
/
groups_
*
filterHeight
*
filterWidth
*
outputHeight
*
outputWidth
;
resizeBuffer
<
Device
>
(
size
);
TensorShape
imShape
=
TensorShape
({
inputChannels
/
groups_
,
inputHeight
,
inputWidth
});
TensorShape
colShape
=
TensorShape
({
inputChannels
/
groups_
,
filterHeight
,
filterWidth
,
outputHeight
,
outputWidth
});
resizeBuffer
<
Device
>
(
colShape
.
getElements
());
real
*
colData
=
reinterpret_cast
<
real
*>
(
memory_
->
getBuf
());
Im2ColFunctor
<
Device
,
real
>
im2col
;
Im2ColFunctor
<
kCFO
,
Device
,
real
>
im2col
;
GemmFunctor
<
Device
,
real
>
gemm
;
size_t
inputOffset
=
(
inputChannels
/
groups_
)
*
inputHeight
*
inputWidth
;
size_t
inputOffset
=
imShape
.
getElements
()
;
size_t
outputOffset
=
(
outputChannels
/
groups_
)
*
outputHeight
*
outputWidth
;
size_t
filterOffset
=
filter
.
getElements
()
/
groups_
;
...
...
@@ -171,18 +88,13 @@ public:
for
(
size_t
i
=
0
;
i
<
batchSize
;
i
++
)
{
for
(
size_t
g
=
0
;
g
<
groups_
;
g
++
)
{
im2col
(
inputData
+
g
*
inputOffset
,
inputChannels
/
groups_
,
inputHeight
,
inputWidth
,
filterHeight
,
filterWidth
,
imShape
,
colData
,
colShape
,
strideH
(),
strideW
(),
paddingH
(),
paddingW
(),
outputHeight
,
outputWidth
,
colData
);
paddingW
());
int
M
=
outputChannels
/
groups_
;
int
N
=
outputHeight
*
outputWidth
;
...
...
@@ -249,15 +161,20 @@ public:
real
*
outputGrad
=
inputs
[
0
].
data
<
real
>
();
real
*
filterData
=
inputs
[
1
].
data
<
real
>
();
real
*
inputGrad
=
outputs
[
0
].
data
<
real
>
();
size_t
size
=
inputChannels
/
groups_
*
filterHeight
*
filterWidth
*
outputHeight
*
outputWidth
;
resizeBuffer
<
Device
>
(
size
);
TensorShape
imShape
=
TensorShape
({
inputChannels
/
groups_
,
inputHeight
,
inputWidth
});
TensorShape
colShape
=
TensorShape
({
inputChannels
/
groups_
,
filterHeight
,
filterWidth
,
outputHeight
,
outputWidth
});
resizeBuffer
<
Device
>
(
colShape
.
getElements
());
real
*
colData
=
reinterpret_cast
<
real
*>
(
memory_
->
getBuf
());
Col2ImFunctor
<
Device
,
real
>
col2im
;
Col2ImFunctor
<
kCFO
,
Device
,
real
>
col2im
;
GemmFunctor
<
Device
,
real
>
gemm
;
size_t
inputOffset
=
(
inputChannels
/
groups_
)
*
inputHeight
*
inputWidth
;
size_t
inputOffset
=
imShape
.
getElements
()
;
size_t
outputOffset
=
(
outputChannels
/
groups_
)
*
outputHeight
*
outputWidth
;
size_t
filterOffset
=
filter
.
getElements
()
/
groups_
;
...
...
@@ -280,20 +197,14 @@ public:
0.0
f
,
colData
,
N
);
col2im
(
colData
,
inputChannels
/
groups_
,
inputHeight
,
inputWidth
,
filterHeight
,
filterWidth
,
col2im
(
inputGrad
+
g
*
inputOffset
,
imShape
,
colData
,
colShape
,
strideH
(),
strideW
(),
paddingH
(),
paddingW
(),
outputHeight
,
outputWidth
,
inputGrad
+
g
*
inputOffset
);
paddingW
());
}
inputGrad
+=
inputChannels
*
inputHeight
*
inputWidth
;
outputGrad
+=
outputChannels
*
outputHeight
*
outputWidth
;
...
...
@@ -347,33 +258,33 @@ public:
real
*
outputGrad
=
inputs
[
0
].
data
<
real
>
();
real
*
inputData
=
inputs
[
1
].
data
<
real
>
();
real
*
filterGrad
=
outputs
[
0
].
data
<
real
>
();
size_t
size
=
inputChannels
/
groups_
*
filterHeight
*
filterWidth
*
outputHeight
*
outputWidth
;
resizeBuffer
<
Device
>
(
size
);
TensorShape
imShape
=
TensorShape
({
inputChannels
/
groups_
,
inputHeight
,
inputWidth
});
TensorShape
colShape
=
TensorShape
({
inputChannels
/
groups_
,
filterHeight
,
filterWidth
,
outputHeight
,
outputWidth
});
resizeBuffer
<
Device
>
(
colShape
.
getElements
());
real
*
colData
=
reinterpret_cast
<
real
*>
(
memory_
->
getBuf
());
Im2ColFunctor
<
Device
,
real
>
im2col
;
Im2ColFunctor
<
kCFO
,
Device
,
real
>
im2col
;
GemmFunctor
<
Device
,
real
>
gemm
;
size_t
inputOffset
=
(
inputChannels
/
groups_
)
*
inputHeight
*
inputWidth
;
size_t
inputOffset
=
imShape
.
getElements
()
;
size_t
outputOffset
=
(
outputChannels
/
groups_
)
*
outputHeight
*
outputWidth
;
size_t
filterOffset
=
filter
.
getElements
()
/
groups_
;
for
(
size_t
i
=
0
;
i
<
batchSize
;
i
++
)
{
for
(
size_t
g
=
0
;
g
<
groups_
;
g
++
)
{
im2col
(
inputData
+
g
*
inputOffset
,
inputChannels
/
groups_
,
inputHeight
,
inputWidth
,
filterHeight
,
filterWidth
,
imShape
,
colData
,
colShape
,
strideH
(),
strideW
(),
paddingH
(),
paddingW
(),
outputHeight
,
outputWidth
,
colData
);
paddingW
());
int
M
=
outputChannels
/
groups_
;
int
K
=
outputHeight
*
outputWidth
;
...
...
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