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
PaddlePaddle
Paddle-Lite
提交
3076c54f
P
Paddle-Lite
项目概览
PaddlePaddle
/
Paddle-Lite
通知
331
Star
4
Fork
1
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
271
列表
看板
标记
里程碑
合并请求
78
Wiki
0
Wiki
分析
仓库
DevOps
项目成员
Pages
P
Paddle-Lite
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
271
Issue
271
列表
看板
标记
里程碑
合并请求
78
合并请求
78
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
3076c54f
编写于
11月 10, 2018
作者:
H
hjchen2
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Fix im2col bug and support do winograd with multi-threads
上级
23275903
变更
4
隐藏空白更改
内联
并排
Showing
4 changed file
with
60 addition
and
67 deletion
+60
-67
src/operators/kernel/central-arm-func/conv_arm_func.h
src/operators/kernel/central-arm-func/conv_arm_func.h
+2
-1
src/operators/math/im2col.cpp
src/operators/math/im2col.cpp
+36
-26
src/operators/math/winograd/winograd.cpp
src/operators/math/winograd/winograd.cpp
+4
-0
src/operators/math/winograd/winograd_transform_f6k3.cpp
src/operators/math/winograd/winograd_transform_f6k3.cpp
+18
-40
未找到文件。
src/operators/kernel/central-arm-func/conv_arm_func.h
浏览文件 @
3076c54f
...
...
@@ -166,7 +166,8 @@ void ConvCompute(const ConvParam<CPU> ¶m) {
param
.
Strides
()[
0
]
==
param
.
Strides
()[
1
]
&&
param
.
Dilations
()[
0
]
==
param
.
Dilations
()[
1
]
&&
param
.
Filter
()
->
dims
()[
2
]
==
3
&&
param
.
Strides
()[
0
]
==
1
&&
param
.
Dilations
()[
0
]
==
1
&&
param
.
Input
()
->
dims
()[
1
]
>=
16
)
{
param
.
Dilations
()[
0
]
==
1
&&
param
.
Output
()
->
dims
()[
1
]
>=
16
&&
param
.
Output
()
->
dims
()[
2
]
>=
16
)
{
BatchConv3x3Winograd
(
param
);
}
else
{
ConvBasic
<
float
,
float
>
(
param
);
...
...
src/operators/math/im2col.cpp
浏览文件 @
3076c54f
...
...
@@ -41,48 +41,48 @@ void ExtractToImg(const float *im_data, float *col_data, const int im_height,
im_data
+=
start_height
*
im_width
+
start_width
;
col_data
+=
col_start_height
*
col_width
+
col_start_width
;
#pragma omp parallel for
for
(
int
i
=
start_height
;
i
<
end_height
;
i
+=
stride_h
)
{
const
float
*
local_im_data
=
im_data
+
i
*
im_width
*
stride_h
;
float
*
local_col_data
=
col_data
+
col_width
;
if
(
stride_w
==
1
)
{
memcpy
(
local_col_data
,
local_
im_data
,
extract
*
sizeof
(
float
));
memcpy
(
col_data
,
im_data
,
extract
*
sizeof
(
float
));
}
else
if
(
stride_w
==
2
)
{
int
s
=
0
;
#if __ARM_NEON
for
(;
s
<
extract
-
15
;
s
+=
16
)
{
float32x4x2_t
img
=
vld2q_f32
(
local_
im_data
+
s
*
2
);
vst1q_f32
(
local_
col_data
+
s
,
img
.
val
[
0
]);
for
(;
s
<
extract
-
3
;
s
+=
4
)
{
float32x4x2_t
img
=
vld2q_f32
(
im_data
+
s
*
2
);
vst1q_f32
(
col_data
+
s
,
img
.
val
[
0
]);
}
#endif
for
(;
s
<
extract
;
++
s
)
{
local_col_data
[
s
]
=
local_
im_data
[
s
*
2
];
col_data
[
s
]
=
im_data
[
s
*
2
];
}
}
else
if
(
stride_w
==
3
)
{
int
s
=
0
;
#if __ARM_NEON
for
(;
s
<
extract
-
15
;
s
+=
16
)
{
float32x4x3_t
img
=
vld3q_f32
(
local_
im_data
+
s
*
3
);
vst1q_f32
(
local_
col_data
+
s
,
img
.
val
[
0
]);
for
(;
s
<
extract
-
3
;
s
+=
4
)
{
float32x4x3_t
img
=
vld3q_f32
(
im_data
+
s
*
3
);
vst1q_f32
(
col_data
+
s
,
img
.
val
[
0
]);
}
#endif
for
(;
s
<
extract
;
++
s
)
{
local_col_data
[
s
]
=
local_
im_data
[
s
*
3
];
col_data
[
s
]
=
im_data
[
s
*
3
];
}
}
else
if
(
stride_w
==
4
)
{
int
s
=
0
;
#if __ARM_NEON
for
(;
s
<
extract
-
15
;
s
+=
16
)
{
float32x4x4_t
img
=
vld4q_f32
(
local_
im_data
+
s
*
4
);
vst1q_f32
(
local_
col_data
+
s
,
img
.
val
[
0
]);
for
(;
s
<
extract
-
3
;
s
+=
4
)
{
float32x4x4_t
img
=
vld4q_f32
(
im_data
+
s
*
4
);
vst1q_f32
(
col_data
+
s
,
img
.
val
[
0
]);
}
#endif
for
(;
s
<
extract
;
++
s
)
{
local_col_data
[
s
]
=
local_
im_data
[
s
*
4
];
col_data
[
s
]
=
im_data
[
s
*
4
];
}
}
else
{
PADDLE_MOBILE_THROW_EXCEPTION
(
"stride_w must be one of 1, 2, 3 and 4."
);
}
im_data
+=
im_width
*
stride_h
;
col_data
+=
col_width
;
}
}
...
...
@@ -428,18 +428,23 @@ void Im2ColFunctor<ColFormat::kCFO, CPU, float>::operator()(
im_data
+=
isize
*
isize
;
}
}
else
if
(
stride
[
0
]
<=
4
&&
dilation
[
0
]
==
1
&&
dilation
[
0
]
==
dilation
[
1
])
{
int
im_spatial_size
=
im_height
*
im_width
;
int
col_spatial_size
=
col_height
*
col_width
;
// pad 0
memset
(
col_data
,
0
,
col
->
numel
()
*
sizeof
(
float
));
#pragma omp parallel for
for
(
int
ic
=
0
;
ic
<
im_channels
;
++
ic
)
{
const
float
*
local_im_data
=
im_data
+
ic
*
im_spatial_size
;
float
*
local_col_data
=
col_data
+
ic
*
filter_height
*
filter_width
*
col_spatial_size
;
for
(
int
kh
=
0
;
kh
<
filter_height
;
++
kh
)
{
for
(
int
kw
=
0
;
kw
<
filter_width
;
++
kw
)
{
ExtractToImg
(
im_data
,
col_data
,
im_height
,
im_width
,
col_height
,
col_
width
,
padding
[
0
],
padding
[
1
],
stride
[
0
],
stride
[
1
],
kh
,
kw
);
col_data
+=
col_height
*
col_width
;
ExtractToImg
(
local_im_data
,
local_col_data
,
im_height
,
im_width
,
col_
height
,
col_width
,
padding
[
0
],
padding
[
1
],
stride
[
0
],
stride
[
1
],
kh
,
kw
);
local_col_data
+=
col_spatial_size
;
}
}
im_data
+=
im_height
*
im_width
;
}
}
else
{
#endif
...
...
@@ -553,18 +558,23 @@ void Im2ColFunctor<ColFormat::kCFO, CPU, int8_t>::operator()(
int8_t
*
col_data
=
col
->
mutable_data
<
int8_t
>
();
#if defined(__ARM_NEON__) || defined(__ARM_NEON)
if
(
stride
[
0
]
<=
4
&&
dilation
[
0
]
==
1
&&
dilation
[
0
]
==
dilation
[
1
])
{
int
im_spatial_size
=
im_height
*
im_width
;
int
col_spatial_size
=
col_height
*
col_width
;
// pad 0
memset
(
col_data
,
0
,
col
->
numel
()
*
sizeof
(
int8_t
));
#pragma omp parallel for
for
(
int
ic
=
0
;
ic
<
im_channels
;
++
ic
)
{
const
int8_t
*
local_im_data
=
im_data
+
ic
*
im_spatial_size
;
int8_t
*
local_col_data
=
col_data
+
ic
*
filter_height
*
filter_width
*
col_spatial_size
;
for
(
int
kh
=
0
;
kh
<
filter_height
;
++
kh
)
{
for
(
int
kw
=
0
;
kw
<
filter_width
;
++
kw
)
{
ExtractToImg
(
im_data
,
col_data
,
im_height
,
im_width
,
col_height
,
col_
width
,
padding
[
0
],
padding
[
1
],
stride
[
0
],
stride
[
1
],
kh
,
kw
);
col_data
+=
col_height
*
col_width
;
ExtractToImg
(
local_im_data
,
local_col_data
,
im_height
,
im_width
,
col_
height
,
col_width
,
padding
[
0
],
padding
[
1
],
stride
[
0
],
stride
[
1
],
kh
,
kw
);
local_col_data
+=
col_spatial_size
;
}
}
im_data
+=
im_height
*
im_width
;
}
}
else
{
#endif
...
...
src/operators/math/winograd/winograd.cpp
浏览文件 @
3076c54f
...
...
@@ -30,6 +30,9 @@ void winograd_f6k3(const framework::Tensor &input,
const
framework
::
Tensor
&
weight
,
framework
::
Tensor
*
output
)
{
framework
::
Tensor
transformed_input
;
framework
::
Tensor
transformed_weight
;
#if __aarch64__
// TODO(hjchen2)
#else
// transform weight
winograd_transform_weight
<
8
,
3
>
(
weight
,
&
transformed_weight
);
// tile input and transform
...
...
@@ -37,6 +40,7 @@ void winograd_f6k3(const framework::Tensor &input,
// caculate output
winograd_transform_output
<
8
,
3
>
(
transformed_input
,
transformed_weight
,
output
);
#endif
}
// F(4X4, 5X5)
...
...
src/operators/math/winograd/winograd_transform_f6k3.cpp
浏览文件 @
3076c54f
...
...
@@ -12,8 +12,12 @@ 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. */
// Inspired by https://arxiv.org/abs/1509.09308 and
// https://github.com/andravin/wincnn and refered from nnpack and ncnn project
// Inspired by https://arxiv.org/abs/1509.09308 and refered from nnpack and ncnn
// project.
#ifdef CONV_OP
#ifndef __aarch64__
#include "operators/math/pad.h"
#include "operators/math/winograd/winograd_transform.h"
...
...
@@ -47,12 +51,13 @@ void winograd_transform_weight<8, 3>(const framework::Tensor &weight,
const
float
transform_matrix
[
8
]
=
{
2.
f
,
-
2.
f
/
9
,
1.
f
/
90
,
1.
f
/
180
};
const
float
*
inptr
=
weight
.
data
<
float
>
();
int
remain_start
=
out_channel
&
0xFFFC
;
#if
def __aarch64__
#if
0
remain_start = 0;
#else
#pragma omp parallel for
for
(
int
oc
=
0
;
oc
<
out_channel
-
3
;
oc
+=
4
)
{
float
gw
[
96
];
// gw[3][8][4]
const
float
*
inptr0
=
inptr
+
oc
*
in_channel
*
9
;
//
float
gw
[
96
];
// gw[3][8][4]
const
float
*
inptr0
=
inptr
+
oc
*
in_channel
*
9
;
const
float
*
inptr1
=
inptr
+
(
oc
+
1
)
*
in_channel
*
9
;
const
float
*
inptr2
=
inptr
+
(
oc
+
2
)
*
in_channel
*
9
;
const
float
*
inptr3
=
inptr
+
(
oc
+
3
)
*
in_channel
*
9
;
...
...
@@ -252,9 +257,10 @@ void winograd_transform_weight<8, 3>(const framework::Tensor &weight,
"q13"
,
"r0"
);
}
}
#endif
// __aarch64__
#endif
// remain output channel
#pragma omp parallel for
for
(
int
oc
=
remain_start
;
oc
<
out_channel
;
++
oc
)
{
float
gw
[
3
][
8
];
// gw[3][8]
const
float
*
inptr0
=
inptr
+
oc
*
in_channel
*
9
;
//
...
...
@@ -301,10 +307,6 @@ void winograd_transform_weight<8, 3>(const framework::Tensor &weight,
outptr
+=
4
;
}
}
// for (int i = 0; i < output->numel(); ++i) {
// DLOG << "TransK[" << i << "] = " << trans_outptr[i];
// }
}
template
<
>
...
...
@@ -657,6 +659,7 @@ void winograd_transform_input<8, 3>(const framework::Tensor &input,
#endif
// remainer channels
#pragma omp parallel for
for
(
int
c
=
remain_c_start
;
c
<
channel
;
++
c
)
{
const
float
*
in
=
inptr
+
c
*
image_size
;
float
d_bt
[
64
];
// d * B_t
...
...
@@ -867,15 +870,6 @@ void winograd_transform_input<8, 3>(const framework::Tensor &input,
}
}
}
// for (int c = 0; c < channel; ++c) {
// for (int tile = 0; tile < output->numel()/channel/64; ++tile) {
// for (int i = 0; i < 64; ++i) {
// int offset = (((tile / 8) * 64 + i) * channel + c) * 8 + (tile % 8);
// DLOG << "TransInput[" << i << "] = " << outptr[offset];
// }
// }
// }
}
template
<
>
...
...
@@ -897,6 +891,7 @@ void winograd_transform_output<8, 3>(const framework::Tensor &input,
const
float
*
input_ptr
=
input
.
data
<
float
>
();
const
float
*
weight_ptr
=
weight
.
data
<
float
>
();
#pragma omp parallel for
for
(
int
i
=
0
;
i
<
out_channel
;
++
i
)
{
float
*
uv_ptr
=
uv_trans_ptr
+
(
i
*
tiles
*
64
*
32
);
for
(
int
k
=
0
;
k
<
64
;
++
k
)
{
...
...
@@ -1017,15 +1012,6 @@ void winograd_transform_output<8, 3>(const framework::Tensor &input,
}
}
// for (int c = 0; c < 4 * out_channel; ++c) {
// for (int tile = 0; tile < 8 * tiles; ++tile) {
// for (int i = 0; i < 64; ++i) {
// int offset = (c * 8 * tiles + tile) * 64 + i;
// DLOG << "uv_trans[" << i << "] = " << uv_trans_ptr[offset];
// }
// }
// }
/*
* s0 = m0 + (m1 + m2) + (m3 + m4) + 32 * (m5 + m6)
* s1 = (m1 - m2) + 2 * (m3 - m4) + 16 * (m5 - m6)
...
...
@@ -1045,12 +1031,7 @@ void winograd_transform_output<8, 3>(const framework::Tensor &input,
int
uv_image_size
=
uv_trans
.
dims
()[
1
]
*
64
;
float
transform_matrix
[
8
]
=
{
2.
f
,
4.
f
,
8.
f
,
16.
f
};
// DLOG << "out_channel: " << out_channel;
// DLOG << "h_tiles: " << h_tiles;
// DLOG << "w_tiles: " << w_tiles;
// DLOG << "remain_h: " << remain_h;
// DLOG << "remain_w: " << remain_w;
#pragma omp parallel for
for
(
int
oc
=
0
;
oc
<
out_channel
;
++
oc
)
{
float
at_m
[
48
];
// [6][8]
float
output_tmp
[
36
];
// [6][6], temporarily restore results
...
...
@@ -1118,9 +1099,6 @@ void winograd_transform_output<8, 3>(const framework::Tensor &input,
:
[
tm_ptr
]
"r"
((
float
*
)
transform_matrix
)
:
"cc"
,
"memory"
,
"q0"
,
"q1"
,
"q2"
,
"q3"
,
"q4"
,
"q5"
,
"q6"
,
"q7"
,
"q8"
,
"q9"
,
"q10"
,
"q11"
,
"q12"
,
"q13"
,
"q14"
,
"q15"
,
"r0"
);
// for (int i = 0; i < 48; ++i) {
// DLOG << "at_m[" << i << "] = " << at_m[i];
// }
float
*
at_m_ptr0
=
at_m
;
float
*
at_m_ptr1
=
at_m
+
24
;
...
...
@@ -1252,9 +1230,6 @@ void winograd_transform_output<8, 3>(const framework::Tensor &input,
float
*
out_ptr
=
output_ptr
+
offset
;
int
remain_row
=
(
tile_h
<
h_tiles
-
1
)
?
6
:
remain_h
;
int
remain_col
=
(
tile_w
<
w_tiles
-
1
)
?
6
:
remain_w
;
// for (int i = 0; i < 36; ++i) {
// DLOG << "output_tmp[" << i << "] = " << output_tmp[i];
// }
for
(
int
i
=
0
;
i
<
remain_row
;
++
i
,
out_ptr
+=
out_w
)
{
memcpy
(
out_ptr
,
output_tmp
+
i
*
6
,
remain_col
*
sizeof
(
float
));
}
...
...
@@ -1391,3 +1366,6 @@ void winograd_transform_output<8, 3>(const framework::Tensor &input,
}
// namespace math
}
// namespace operators
}
// namespace paddle_mobile
#endif // __aarch64__
#endif // CONV_OP
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