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
afeixing77
Paddle
提交
0dffe68c
P
Paddle
项目概览
afeixing77
/
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
0dffe68c
编写于
8月 24, 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
Add NeonDepthwiseConvFunction.
上级
692259e0
变更
6
隐藏空白更改
内联
并排
Showing
6 changed file
with
398 addition
and
0 deletion
+398
-0
paddle/function/CMakeLists.txt
paddle/function/CMakeLists.txt
+2
-0
paddle/function/DepthwiseConvOpTest.cpp
paddle/function/DepthwiseConvOpTest.cpp
+5
-0
paddle/function/Im2Col.h
paddle/function/Im2Col.h
+92
-0
paddle/function/neon/NeonDepthwiseConv.cpp
paddle/function/neon/NeonDepthwiseConv.cpp
+227
-0
paddle/function/neon/NeonDepthwiseConv.h
paddle/function/neon/NeonDepthwiseConv.h
+25
-0
paddle/function/neon/neon_util.h
paddle/function/neon/neon_util.h
+47
-0
未找到文件。
paddle/function/CMakeLists.txt
浏览文件 @
0dffe68c
...
...
@@ -21,6 +21,8 @@ if(USE_NNPACK)
endif
()
endif
()
list
(
APPEND cpp_files neon/NeonDepthwiseConv.cpp
)
add_library
(
paddle_function STATIC
${
cpp_files
}
${
cu_objs
}
)
add_dependencies
(
paddle_function
${
external_project_dependencies
}
)
add_dependencies
(
paddle_function paddle_proto
)
...
...
paddle/function/DepthwiseConvOpTest.cpp
浏览文件 @
0dffe68c
...
...
@@ -34,4 +34,9 @@ TEST(DepthwiseConv, BackwardFilter) {
}
#endif
TEST
(
DepthwiseConv
,
Forward
)
{
DepthwiseConvolution
<
DEVICE_TYPE_CPU
,
DEVICE_TYPE_CPU
>
(
"GemmConv-CPU"
,
"NeonDepthwiseConv-CPU"
,
forward
);
}
}
// namespace paddle
paddle/function/Im2Col.h
浏览文件 @
0dffe68c
...
...
@@ -16,6 +16,7 @@ limitations under the License. */
#include "TensorShape.h"
#include "TensorType.h"
#include "neon/neon_util.h"
namespace
paddle
{
...
...
@@ -93,4 +94,95 @@ public:
int
paddingWidth
);
};
template
<
class
T
>
struct
Padding
{
static
void
run
(
const
T
*
src
,
T
*
dest
,
int
channels
,
int
inputHeight
,
int
inputWidth
,
int
paddingHeight
,
int
paddingWidth
)
{
const
int
destWidth
=
inputWidth
+
2
*
paddingWidth
;
for
(
int
c
=
0
;
c
<
channels
;
c
++
)
{
if
(
paddingHeight
>
0
)
{
memset
(
dest
,
0
,
destWidth
*
paddingHeight
*
sizeof
(
T
));
dest
+=
destWidth
*
paddingHeight
;
}
for
(
int
i
=
0
;
i
<
inputHeight
;
i
++
)
{
// padding head
for
(
int
j
=
0
;
j
<
paddingWidth
;
j
++
)
{
*
dest
++
=
T
(
0
);
}
memcpy
(
dest
,
src
,
inputWidth
*
sizeof
(
T
));
dest
+=
inputWidth
;
src
+=
inputWidth
;
// padding tail
for
(
int
j
=
0
;
j
<
paddingWidth
;
j
++
)
{
*
dest
++
=
T
(
0
);
}
}
if
(
paddingHeight
>
0
)
{
memset
(
dest
,
0
,
destWidth
*
paddingHeight
*
sizeof
(
T
));
dest
+=
destWidth
*
paddingHeight
;
}
}
}
};
#if defined(__ARM_NEON__) || defined(__ARM_NEON)
template
<
>
struct
Padding
<
float
>
{
static
void
run
(
const
float
*
src
,
float
*
dest
,
int
channels
,
int
inputHeight
,
int
inputWidth
,
int
paddingHeight
,
int
paddingWidth
)
{
const
int
destWidth
=
inputWidth
+
2
*
paddingWidth
;
for
(
int
c
=
0
;
c
<
channels
;
c
++
)
{
if
(
paddingHeight
>
0
)
{
memset
(
dest
,
0
,
destWidth
*
paddingHeight
*
sizeof
(
float
));
dest
+=
destWidth
*
paddingHeight
;
}
for
(
int
i
=
0
;
i
<
inputHeight
;
i
++
)
{
// padding head
for
(
int
j
=
0
;
j
<
paddingWidth
;
j
++
)
{
*
dest
++
=
float
(
0
);
}
int
step
=
inputWidth
>>
2
;
int
remain
=
inputWidth
&
3
;
for
(
int
s
=
0
;
s
<
step
;
s
++
)
{
float32x4_t
s0
=
vld1q_f32
(
src
);
vst1q_f32
(
dest
,
s0
);
src
+=
4
;
dest
+=
4
;
}
for
(
int
r
=
0
;
r
<
remain
;
r
++
)
{
*
dest
++
=
*
src
++
;
}
// padding tail
for
(
int
j
=
0
;
j
<
paddingWidth
;
j
++
)
{
*
dest
++
=
float
(
0
);
}
}
if
(
paddingHeight
>
0
)
{
memset
(
dest
,
0
,
destWidth
*
paddingHeight
*
sizeof
(
float
));
dest
+=
destWidth
*
paddingHeight
;
}
}
}
};
#endif
}
// namespace paddle
paddle/function/neon/NeonDepthwiseConv.cpp
0 → 100644
浏览文件 @
0dffe68c
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
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 "neon_util.h"
#include "paddle/function/ConvOp.h"
#include "paddle/function/Im2Col.h"
namespace
paddle
{
namespace
neon
{
#if defined(__ARM_NEON__) || defined(__ARM_NEON)
template
<
int
filterSize
,
int
stride
>
struct
DepthwiseConvKernel
{};
inline
float32_t
conv3x3
(
float32x4_t
r0
,
float32x4_t
r1
,
float32x4_t
r2
,
float32x4_t
k0
,
float32x4_t
k1
,
float32x4_t
k2
)
{
float32x4_t
tmp
;
tmp
=
vmulq_f32
(
r0
,
k0
);
tmp
=
vmlaq_f32
(
tmp
,
r1
,
k1
);
tmp
=
vmlaq_f32
(
tmp
,
r2
,
k2
);
return
vaddvq_f32
(
tmp
);
}
/**
* Each step calculates four elements of the output.
* First step:
* R0[0, 1, 2, 3...] * K[0][0]
* R0[1, 2, 3, 4...] * K[0][1]
* R0[2, 3, 4, 5...] * K[0][2]
* R1[0, 1, 2, 3...] * K[1][0]
* R1[1, 2, 3, 4...] * K[1][1]
* R1[2, 3, 4, 5...] * K[1][2]
* R2[0, 1, 2, 3...] * K[2][0]
* R2[1, 2, 3, 4...] * K[2][1]
* + R2[2, 3, 4, 5...] * K[2][2]
* ------------------------------
* Output[0, 1, 2, 3]
*/
template
<
>
struct
DepthwiseConvKernel
<
3
,
1
>
{
static
void
run
(
const
float
*
inputData
,
const
float
*
filterData
,
int
inputHeight
,
int
inputWidth
,
int
outputChannels
,
int
outputHeight
,
int
outputWidth
,
int
filterMultiplier
,
float
*
outputData
)
{
const
int
steps
=
outputWidth
>>
2
;
const
int
remain
=
outputWidth
&
3
;
for
(
int
c
=
0
;
c
<
outputChannels
;
c
++
,
filterData
+=
9
)
{
// Load the filters
float32x4_t
k
[
3
];
k
[
0
]
=
vld1q_f32
(
filterData
);
k
[
1
]
=
vld1q_f32
(
filterData
+
3
);
k
[
2
]
=
vld1q_f32
(
filterData
+
6
);
k
[
0
]
=
vsetq_lane_f32
(
0.
f
,
k
[
0
],
3
);
k
[
1
]
=
vsetq_lane_f32
(
0.
f
,
k
[
1
],
3
);
k
[
2
]
=
vsetq_lane_f32
(
0.
f
,
k
[
2
],
3
);
const
float
*
r0
=
inputData
+
(
c
/
filterMultiplier
)
*
(
inputHeight
*
inputWidth
);
const
float
*
r1
=
r0
+
inputWidth
;
const
float
*
r2
=
r0
+
inputWidth
*
2
;
float32x4_t
input
[
3
][
3
];
for
(
int
h
=
0
;
h
<
outputHeight
;
h
++
)
{
for
(
int
s
=
0
;
s
<
steps
;
s
++
)
{
// Load the inputs
float32x4_t
tmp
;
input
[
0
][
0
]
=
vld1q_f32
(
r0
);
tmp
=
vld1q_f32
(
r0
+
4
);
input
[
0
][
1
]
=
vextq_f32
(
input
[
0
][
0
],
tmp
,
1
);
input
[
0
][
2
]
=
vextq_f32
(
input
[
0
][
0
],
tmp
,
2
);
input
[
1
][
0
]
=
vld1q_f32
(
r1
);
tmp
=
vld1q_f32
(
r1
+
4
);
input
[
1
][
1
]
=
vextq_f32
(
input
[
1
][
0
],
tmp
,
1
);
input
[
1
][
2
]
=
vextq_f32
(
input
[
1
][
0
],
tmp
,
2
);
input
[
2
][
0
]
=
vld1q_f32
(
r2
);
tmp
=
vld1q_f32
(
r2
+
4
);
input
[
2
][
1
]
=
vextq_f32
(
input
[
2
][
0
],
tmp
,
1
);
input
[
2
][
2
]
=
vextq_f32
(
input
[
2
][
0
],
tmp
,
2
);
float32x4_t
tmp1
=
vdupq_n_f32
(
0.
f
);
float32x4_t
tmp2
=
vdupq_n_f32
(
0.
f
);
tmp1
=
vmlaq_laneq_f32
(
tmp1
,
input
[
0
][
0
],
k
[
0
],
0
);
tmp2
=
vmlaq_laneq_f32
(
tmp2
,
input
[
0
][
1
],
k
[
0
],
1
);
tmp1
=
vmlaq_laneq_f32
(
tmp1
,
input
[
0
][
2
],
k
[
0
],
2
);
tmp2
=
vmlaq_laneq_f32
(
tmp2
,
input
[
1
][
0
],
k
[
1
],
0
);
tmp1
=
vmlaq_laneq_f32
(
tmp1
,
input
[
1
][
1
],
k
[
1
],
1
);
tmp2
=
vmlaq_laneq_f32
(
tmp2
,
input
[
1
][
2
],
k
[
1
],
2
);
tmp1
=
vmlaq_laneq_f32
(
tmp1
,
input
[
2
][
0
],
k
[
2
],
0
);
tmp2
=
vmlaq_laneq_f32
(
tmp2
,
input
[
2
][
1
],
k
[
2
],
1
);
tmp1
=
vmlaq_laneq_f32
(
tmp1
,
input
[
2
][
2
],
k
[
2
],
2
);
tmp1
=
vaddq_f32
(
tmp1
,
tmp2
);
vst1q_f32
(
outputData
,
tmp1
);
r0
+=
4
;
r1
+=
4
;
r2
+=
4
;
outputData
+=
4
;
}
for
(
int
r
=
0
;
r
<
remain
;
r
++
)
{
float32x4_t
i0
=
vld1q_f32
(
r0
);
float32x4_t
i1
=
vld1q_f32
(
r1
);
float32x4_t
i2
=
vld1q_f32
(
r2
);
*
outputData
=
conv3x3
(
i0
,
i1
,
i2
,
k
[
0
],
k
[
1
],
k
[
2
]);
r0
++
;
r1
++
;
r2
++
;
outputData
++
;
}
r0
+=
2
;
r1
+=
2
;
r2
+=
2
;
}
}
}
};
template
<
DeviceType
Device
>
class
NeonDepthwiseConvFunction
:
public
ConvFunctionBase
{
public:
void
init
(
const
FuncConfig
&
config
)
override
{
ConvFunctionBase
::
init
(
config
);
}
void
check
(
const
BufferArgs
&
inputs
,
const
BufferArgs
&
outputs
)
override
{
const
TensorShape
&
input
=
inputs
[
0
].
shape
();
const
TensorShape
&
filter
=
inputs
[
1
].
shape
();
const
TensorShape
&
output
=
outputs
[
0
].
shape
();
checkShape
(
input
,
filter
,
output
);
}
void
calc
(
const
BufferArgs
&
inputs
,
const
BufferArgs
&
outputs
)
override
{
CHECK_EQ
(
numInputs_
,
inputs
.
size
());
CHECK_EQ
(
numOutputs_
,
outputs
.
size
());
check
(
inputs
,
outputs
);
const
TensorShape
&
input
=
inputs
[
0
].
shape
();
const
TensorShape
&
filter
=
inputs
[
1
].
shape
();
const
TensorShape
&
output
=
outputs
[
0
].
shape
();
size_t
batchSize
=
input
[
0
];
size_t
inputChannels
=
input
[
1
];
size_t
inputHeight
=
input
[
2
];
size_t
inputWidth
=
input
[
3
];
size_t
filterHeight
=
getFilterHeight
(
filter
);
size_t
filterWidth
=
getFilterWidth
(
filter
);
size_t
outputChannels
=
output
[
1
];
size_t
outputHeight
=
output
[
2
];
size_t
outputWidth
=
output
[
3
];
size_t
filterMultiplier
=
outputChannels
/
groups_
;
CHECK_EQ
(
inputChannels
,
groups_
);
// only support
CHECK_EQ
(
strideH
(),
strideW
());
CHECK_EQ
(
filterHeight
,
filterWidth
);
CHECK_EQ
(
filterHeight
,
size_t
(
3
));
CHECK_LT
(
strideH
(),
size_t
(
3
));
float
*
inputData
=
inputs
[
0
].
data
<
float
>
();
float
*
filterData
=
inputs
[
1
].
data
<
float
>
();
float
*
outputData
=
outputs
[
0
].
data
<
float
>
();
// padding the input
float
*
inputPadding
=
inputData
;
if
(
paddingH
()
>
0
||
paddingW
()
>
0
)
{
int
newSize
=
batchSize
*
inputChannels
*
(
inputHeight
+
2
*
paddingH
())
*
(
inputWidth
+
2
*
paddingW
());
resizeBuffer
<
Device
>
(
newSize
);
inputPadding
=
reinterpret_cast
<
float
*>
(
memory_
->
getBuf
());
Padding
<
float
>::
run
(
inputData
,
inputPadding
,
batchSize
*
inputChannels
,
inputHeight
,
inputWidth
,
paddingH
(),
paddingW
());
// height and width of padding data
inputHeight
+=
2
*
paddingH
();
inputWidth
+=
2
*
paddingW
();
}
for
(
size_t
i
=
0
;
i
<
batchSize
;
i
++
)
{
DepthwiseConvKernel
<
3
,
1
>::
run
(
inputPadding
,
filterData
,
inputHeight
,
inputWidth
,
outputChannels
,
outputHeight
,
outputWidth
,
filterMultiplier
,
outputData
);
inputPadding
+=
inputChannels
*
inputHeight
*
inputWidth
;
outputData
+=
outputChannels
*
outputHeight
*
outputWidth
;
}
}
};
REGISTER_TYPED_FUNC
(
NeonDepthwiseConv
,
CPU
,
NeonDepthwiseConvFunction
);
#endif
}
// namespace neon
}
// namespace paddle
paddle/function/neon/NeonDepthwiseConv.h
0 → 100644
浏览文件 @
0dffe68c
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
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. */
#pragma once
namespace
paddle
{
namespace
neon
{
template
<
int
filterSize
,
int
stride
>
struct
DepthwiseConvKernel
{};
}
// namespace neon
}
// namespace paddle
paddle/function/neon/neon_util.h
0 → 100644
浏览文件 @
0dffe68c
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
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. */
#pragma once
#if defined(__ARM_NEON__) || defined(__ARM_NEON)
#include <arm_neon.h>
namespace
paddle
{
namespace
neon
{
inline
float32x4_t
vld1q_f32_aligned
(
const
float
*
p
)
{
return
vld1q_f32
(
(
const
float
*
)
__builtin_assume_aligned
(
p
,
sizeof
(
float32x4_t
)));
}
#ifndef __aarch64__
inline
float32_t
vaddvq_f32
(
float32x4_t
a
)
{
float32x2_t
v
=
vadd_f32
(
vget_high_f32
(
a
),
vget_low_f32
(
a
));
return
vget_lane_f32
(
vpadd_f32
(
v
,
v
),
0
);
}
inline
float32x4_t
vmlaq_laneq_f32
(
float32x4_t
a
,
float32x4_t
b
,
float32x4_t
v
,
const
int
lane
)
{
return
vmlaq_n_f32
(
a
,
b
,
vgetq_lane_f32
(
v
,
lane
));
}
#endif
}
// namespace neon
}
// namespace paddle
#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