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
2b32484a
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
2b32484a
编写于
5月 26, 2020
作者:
H
hanbuhe
浏览文件
操作
浏览文件
下载
差异文件
merge ..
上级
e5f779de
31150e99
变更
5
隐藏空白更改
内联
并排
Showing
5 changed file
with
423 addition
and
69 deletion
+423
-69
lite/backends/arm/math/pooling.cc
lite/backends/arm/math/pooling.cc
+401
-67
lite/backends/opencl/cl_kernel/buffer/sigmoid_kernel.cl
lite/backends/opencl/cl_kernel/buffer/sigmoid_kernel.cl
+22
-0
lite/core/mir/fusion/fc_fuser.cc
lite/core/mir/fusion/fc_fuser.cc
+0
-2
lite/core/program.cc
lite/core/program.cc
+0
-0
lite/core/tensor.h
lite/core/tensor.h
+0
-0
未找到文件。
lite/backends/arm/math/pooling.cc
浏览文件 @
2b32484a
...
@@ -206,6 +206,20 @@ void pooling_basic(const float* din,
...
@@ -206,6 +206,20 @@ void pooling_basic(const float* din,
"ld2 {v0.4s, v1.4s}, [%[dr0]], #32\n"
/* load q0-q1, dr0, 0-7*/
\
"ld2 {v0.4s, v1.4s}, [%[dr0]], #32\n"
/* load q0-q1, dr0, 0-7*/
\
"ld2 {v2.4s, v3.4s}, [%[dr1]], #32\n"
/* load q2-q3, dr1, 0-7*/
"ld2 {v2.4s, v3.4s}, [%[dr1]], #32\n"
/* load q2-q3, dr1, 0-7*/
#define P2x2S2P1_MAX \
"ext v6.16b, %[vzero].16b, v1.16b, #12\n"
/* 1357-0135 */
\
"ext v8.16b, %[vzero].16b, v3.16b, #12\n"
/* 1357-0135 */
\
"sub %[dr0], %[dr0], #4\n"
/* sub */
\
"sub %[dr1], %[dr1], #4\n"
/* sub */
\
"fmax v4.4s, v0.4s, v6.4s\n"
/* max */
\
"fmax v5.4s, v2.4s, v8.4s\n"
/* max */
\
"ld2 {v0.4s, v1.4s}, [%[dr0]], #32\n"
/* load q0-q1, dr0, 0-7*/
\
"ld2 {v2.4s, v3.4s}, [%[dr1]], #32\n"
/* load q2-q3, dr1, 0-7*/
\
"fmax v6.4s, v4.4s, v5.4s\n"
/* max reduce */
\
"subs %w[cnt_num], %w[cnt_num], #1\n"
/* subs cnt_num, #1*/
\
"st1 {v6.4s}, [%[dr_out]], #16\n"
/* store 4 out, dr_out */
\
"ble 2f\n"
/* bne s3_max_loop_mid */
#define P2x2S2P0_MAX \
#define P2x2S2P0_MAX \
"1: \n" \
"1: \n" \
"fmax v4.4s, v0.4s, v1.4s\n"
/* max */
\
"fmax v4.4s, v0.4s, v1.4s\n"
/* max */
\
...
@@ -217,6 +231,21 @@ void pooling_basic(const float* din,
...
@@ -217,6 +231,21 @@ void pooling_basic(const float* din,
"st1 {v6.4s}, [%[dr_out]], #16\n"
/* store 4 out, dr_out */
\
"st1 {v6.4s}, [%[dr_out]], #16\n"
/* store 4 out, dr_out */
\
"bne 1b\n"
/* bne s3_max_loop_mid */
"bne 1b\n"
/* bne s3_max_loop_mid */
#define P2x2S2P1_AVG \
"ext v6.16b, %[vzero].16b, v1.16b, #12\n"
/* 1357-0135 */
\
"ext v8.16b, %[vzero].16b, v3.16b, #12\n"
/* 1357-0135 */
\
"sub %[dr0], %[dr0], #4\n"
/* sub */
\
"sub %[dr1], %[dr1], #4\n"
/* sub */
\
"fadd v4.4s, v0.4s, v6.4s\n"
/* add 0, 2, 4, 6 and 1, 3, 5, 7 */
\
"fadd v5.4s, v2.4s, v8.4s\n"
/* add 0, 2, 4, 6 and 1, 3, 5, 7 */
\
"ld2 {v0.4s, v1.4s}, [%[dr0]], #32\n"
/* load q0-q1, dr0, 0-7*/
\
"ld2 {v2.4s, v3.4s}, [%[dr1]], #32\n"
/* load q2-q3, dr1, 0-7*/
\
"fadd v6.4s, v4.4s, v5.4s\n"
/* add reduce */
\
"subs %w[cnt_num], %w[cnt_num], #1\n"
/* subs cnt_num, #1*/
\
"fmul v4.4s, v6.4s, %[vcoef_left].4s\n"
/* mul coef */
\
"st1 {v4.4s}, [%[dr_out]], #16\n"
/* store 4 out, dr_out */
\
"ble 2f\n"
/* bne s3_max_loop_mid */
#define P2x2S2P0_AVG \
#define P2x2S2P0_AVG \
"1: \n"
/* load bias to q2, q3*/
\
"1: \n"
/* load bias to q2, q3*/
\
"fadd v4.4s, v0.4s, v1.4s\n"
/* add 0, 2, 4, 6 and 1, 3, 5, 7 */
\
"fadd v4.4s, v0.4s, v1.4s\n"
/* add 0, 2, 4, 6 and 1, 3, 5, 7 */
\
...
@@ -228,6 +257,7 @@ void pooling_basic(const float* din,
...
@@ -228,6 +257,7 @@ void pooling_basic(const float* din,
"fmul v4.4s, v6.4s, %[vcoef].4s\n"
/* mul coef */
\
"fmul v4.4s, v6.4s, %[vcoef].4s\n"
/* mul coef */
\
"st1 {v4.4s}, [%[dr_out]], #16\n"
/* store 4 out, dr_out */
\
"st1 {v4.4s}, [%[dr_out]], #16\n"
/* store 4 out, dr_out */
\
"bne 1b\n"
/* bne s3_max_loop_mid */
"bne 1b\n"
/* bne s3_max_loop_mid */
#define P3x3S1_INIT \
#define P3x3S1_INIT \
"ldr q0, [%[dr0]], #16\n"
/* load q0, dr0, 0-3*/
\
"ldr q0, [%[dr0]], #16\n"
/* load q0, dr0, 0-3*/
\
"ldr q1, [%[dr1]], #16\n"
/* load q1, dr1, 0-3*/
\
"ldr q1, [%[dr1]], #16\n"
/* load q1, dr1, 0-3*/
\
...
@@ -518,16 +548,45 @@ void pooling_basic(const float* din,
...
@@ -518,16 +548,45 @@ void pooling_basic(const float* din,
"vld2.f32 {d0-d3}, [%[dr0]]! @ load \n" \
"vld2.f32 {d0-d3}, [%[dr0]]! @ load \n" \
"vld2.f32 {d4-d7}, [%[dr1]]! @ load \n"
"vld2.f32 {d4-d7}, [%[dr1]]! @ load \n"
#define P2x2S2P1_MAX \
"vext.32 q4, %q[vzero], q1, #3 @ 1357-0135\n" \
"vext.32 q5, %q[vzero], q3, #3 @ 1357-0135\n" \
"sub %[dr0], #4 @sub \n" \
"sub %[dr1], #4 @sub \n" \
"vmax.f32 q8, q0, q4 @ max \n" \
"vmax.f32 q9, q2, q5 @ max \n" \
"vld2.f32 {d0-d3}, [%[dr0]]! @ load \n" \
"vld2.f32 {d4-d7}, [%[dr1]]! @ load \n" \
"vmax.f32 q5, q9, q8 @ max reduce\n" \
"subs %[cnt_num], #1 @ subs cnt_num \n" \
"vst1.f32 {d10-d11}, [%[dr_out]]! @ store 4 out \n" \
"ble 2f @ bne \n"
#define P2x2S2P0_MAX \
#define P2x2S2P0_MAX \
"1: @ main loop\n" \
"1: @ main loop\n" \
"vmax.f32 q4, q0, q1 @ max \n" \
"vmax.f32 q4, q0, q1 @ max \n" \
"vmax.f32 q5, q2, q3 @ max \n" \
"vmax.f32 q5, q2, q3 @ max \n" \
"vld2.f32 {d0-d3}, [%[dr0]]! @ load \n" \
"vld2.f32 {d0-d3}, [%[dr0]]! @ load \n" \
"vld2.f32 {d4-d7}, [%[dr1]]! @ load \n" \
"vld2.f32 {d4-d7}, [%[dr1]]! @ load \n" \
"vmax.f32 q
6
, q4, q5 @ max reduce\n" \
"vmax.f32 q
8
, q4, q5 @ max reduce\n" \
"subs %[cnt_num], #1 @ subs cnt_num \n" \
"subs %[cnt_num], #1 @ subs cnt_num \n" \
"vst1.f32 {d12-d13}, [%[dr_out]]! @ store 4 out \n" \
"vst1.f32 {d16-d17}, [%[dr_out]]! @ store 4 out \n" \
"bne 1b @ bne "
"bne 1b @ bne \n"
#define P2x2S2P1_AVG \
"vext.32 q4, %q[vzero], q1, #3 @ 1357-0135\n" \
"vext.32 q5, %q[vzero], q3, #3 @ 1357-0135\n" \
"sub %[dr0], #4 @sub \n" \
"sub %[dr1], #4 @sub \n" \
"vadd.f32 q9, q0, q4 @ max \n" \
"vadd.f32 q8, q2, q5 @ max \n" \
"vld2.f32 {d0-d3}, [%[dr0]]! @ load \n" \
"vld2.f32 {d4-d7}, [%[dr1]]! @ load \n" \
"vadd.f32 q5, q9, q8 @ max reduce\n" \
"subs %[cnt_num], #1 @ subs cnt_num \n" \
"vmul.f32 q4, q5, %q[vcoef_left] @ mul coef \n" \
"vst1.f32 {d8-d9}, [%[dr_out]]! @ store 4 out \n" \
"ble 2f @ bne\n"
#define P2x2S2P0_AVG \
#define P2x2S2P0_AVG \
"1: @ main loop\n" \
"1: @ main loop\n" \
...
@@ -535,9 +594,9 @@ void pooling_basic(const float* din,
...
@@ -535,9 +594,9 @@ void pooling_basic(const float* din,
"vadd.f32 q5, q2, q3 @ add 0, 2, 4, 6 \n" \
"vadd.f32 q5, q2, q3 @ add 0, 2, 4, 6 \n" \
"vld2.f32 {d0-d3}, [%[dr0]]! @ load d0-d3 \n" \
"vld2.f32 {d0-d3}, [%[dr0]]! @ load d0-d3 \n" \
"vld2.f32 {d4-d7}, [%[dr1]]! @ load d4-d7 \n" \
"vld2.f32 {d4-d7}, [%[dr1]]! @ load d4-d7 \n" \
"vadd.f32 q
6
, q4, q5 @ add reduce \n" \
"vadd.f32 q
8
, q4, q5 @ add reduce \n" \
"subs %[cnt_num], #1 @ subs \n" \
"subs %[cnt_num], #1 @ subs \n" \
"vmul.f32 q4, q
6
, %q[vcoef] @ mul coef \n" \
"vmul.f32 q4, q
8
, %q[vcoef] @ mul coef \n" \
"vst1.f32 {d8-d9}, [%[dr_out]]! @ store 4 out \n" \
"vst1.f32 {d8-d9}, [%[dr_out]]! @ store 4 out \n" \
"bne 1b @ bne \n"
"bne 1b @ bne \n"
...
@@ -1037,17 +1096,17 @@ void pooling1x1s2p0_max(const float* din,
...
@@ -1037,17 +1096,17 @@ void pooling1x1s2p0_max(const float* din,
TargetFree
(
TARGET
(
kARM
),
write_ptr
);
TargetFree
(
TARGET
(
kARM
),
write_ptr
);
}
}
void
pooling2x2s2_max
(
const
float
*
din
,
void
pooling2x2s2
p0
_max
(
const
float
*
din
,
float
*
dout
,
float
*
dout
,
int
num
,
int
num
,
int
chout
,
int
chout
,
int
hout
,
int
hout
,
int
wout
,
int
wout
,
int
chin
,
int
chin
,
int
hin
,
int
hin
,
int
win
,
int
win
,
int
pad_bottom
,
int
pad_bottom
,
int
pad_right
)
{
int
pad_right
)
{
int
size_channel_out
=
wout
*
hout
;
int
size_channel_out
=
wout
*
hout
;
int
size_channel_in
=
win
*
hin
;
int
size_channel_in
=
win
*
hin
;
auto
data_out
=
static_cast
<
float
*>
(
dout
);
auto
data_out
=
static_cast
<
float
*>
(
dout
);
...
@@ -1095,7 +1154,7 @@ void pooling2x2s2_max(const float* din,
...
@@ -1095,7 +1154,7 @@ void pooling2x2s2_max(const float* din,
[
dr_out
]
"+r"
(
dr_out
),
[
dr_out
]
"+r"
(
dr_out
),
[
cnt_num
]
"+r"
(
cnt_num
)
[
cnt_num
]
"+r"
(
cnt_num
)
:
:
:
"cc"
,
"memory"
,
"q0"
,
"q1"
,
"q2"
,
"q3"
,
"q4"
,
"q5"
,
"q
6
"
);
:
"cc"
,
"memory"
,
"q0"
,
"q1"
,
"q2"
,
"q3"
,
"q4"
,
"q5"
,
"q
8
"
);
#endif
#endif
dr0
-=
8
;
dr0
-=
8
;
dr1
-=
8
;
dr1
-=
8
;
...
@@ -1121,18 +1180,18 @@ void pooling2x2s2_max(const float* din,
...
@@ -1121,18 +1180,18 @@ void pooling2x2s2_max(const float* din,
}
}
}
}
void
pooling2x2s2_avg
(
const
float
*
din
,
void
pooling2x2s2
p0
_avg
(
const
float
*
din
,
float
*
dout
,
float
*
dout
,
int
num
,
int
num
,
int
chout
,
int
chout
,
int
hout
,
int
hout
,
int
wout
,
int
wout
,
int
chin
,
int
chin
,
int
hin
,
int
hin
,
int
win
,
int
win
,
bool
exclusive
,
bool
exclusive
,
int
pad_bottom
,
int
pad_bottom
,
int
pad_right
)
{
int
pad_right
)
{
int
size_channel_out
=
wout
*
hout
;
int
size_channel_out
=
wout
*
hout
;
int
size_channel_in
=
win
*
hin
;
int
size_channel_in
=
win
*
hin
;
auto
data_out
=
static_cast
<
float
*>
(
dout
);
auto
data_out
=
static_cast
<
float
*>
(
dout
);
...
@@ -1158,12 +1217,14 @@ void pooling2x2s2_avg(const float* din,
...
@@ -1158,12 +1217,14 @@ void pooling2x2s2_avg(const float* din,
const
float
*
data_in_channel
=
data_in_batch
+
c
*
size_channel_in
;
const
float
*
data_in_channel
=
data_in_batch
+
c
*
size_channel_in
;
const
float
*
r0
=
data_in_channel
;
const
float
*
r0
=
data_in_channel
;
const
float
*
r1
=
r0
+
win
;
const
float
*
r1
=
r0
+
win
;
vcoef
=
vdupq_n_f32
(
0.25
f
);
for
(
int
h
=
0
;
h
<
hout
;
h
++
)
{
for
(
int
h
=
0
;
h
<
hout
;
h
++
)
{
float
*
dr_out
=
data_out_channel
;
float
*
dr_out
=
data_out_channel
;
auto
dr0
=
r0
;
auto
dr0
=
r0
;
auto
dr1
=
r1
;
auto
dr1
=
r1
;
if
(
h
*
S
+
K
-
P
>
hin
)
{
if
(
h
*
S
+
K
-
P
>
hin
)
{
dr1
=
zero_ptr
;
dr1
=
zero_ptr
;
vcoef
=
vdupq_n_f32
(
0.5
f
);
}
}
int
cnt_num
=
w_unroll_size
;
int
cnt_num
=
w_unroll_size
;
if
(
w_unroll_size
>
0
)
{
if
(
w_unroll_size
>
0
)
{
...
@@ -1184,7 +1245,7 @@ void pooling2x2s2_avg(const float* din,
...
@@ -1184,7 +1245,7 @@ void pooling2x2s2_avg(const float* din,
[
dr_out
]
"+r"
(
dr_out
),
[
dr_out
]
"+r"
(
dr_out
),
[
cnt_num
]
"+r"
(
cnt_num
)
[
cnt_num
]
"+r"
(
cnt_num
)
:
[
vcoef
]
"w"
(
vcoef
)
:
[
vcoef
]
"w"
(
vcoef
)
:
"cc"
,
"memory"
,
"q0"
,
"q1"
,
"q2"
,
"q3"
,
"q4"
,
"q5"
,
"q
6
"
);
:
"cc"
,
"memory"
,
"q0"
,
"q1"
,
"q2"
,
"q3"
,
"q4"
,
"q5"
,
"q
8
"
);
#endif
#endif
dr0
-=
8
;
dr0
-=
8
;
dr1
-=
8
;
dr1
-=
8
;
...
@@ -1194,8 +1255,14 @@ void pooling2x2s2_avg(const float* din,
...
@@ -1194,8 +1255,14 @@ void pooling2x2s2_avg(const float* din,
int
wstart
=
0
;
int
wstart
=
0
;
for
(
int
j
=
0
;
j
<
w_unroll_remian
;
++
j
)
{
for
(
int
j
=
0
;
j
<
w_unroll_remian
;
++
j
)
{
int
wend
=
std
::
min
(
wstart
+
K
,
rem
);
int
wend
=
std
::
min
(
wstart
+
K
,
rem
);
float
coef
=
0.
5
f
/
(
wend
-
wstart
)
;
float
coef
=
0.
25
f
;
float
tmp
=
0.
f
;
float
tmp
=
0.
f
;
if
(
wend
-
wstart
==
1
&&
pad_right
==
0
)
{
coef
*=
2
;
}
if
(
h
*
S
+
K
-
P
>
hin
&&
pad_bottom
==
0
)
{
coef
*=
2
;
}
for
(
int
i
=
wstart
;
i
<
wend
;
i
++
)
{
for
(
int
i
=
wstart
;
i
<
wend
;
i
++
)
{
tmp
+=
dr0
[
i
]
+
dr1
[
i
];
tmp
+=
dr0
[
i
]
+
dr1
[
i
];
}
}
...
@@ -1212,6 +1279,235 @@ void pooling2x2s2_avg(const float* din,
...
@@ -1212,6 +1279,235 @@ void pooling2x2s2_avg(const float* din,
TargetFree
(
TARGET
(
kARM
),
zero_ptr
);
TargetFree
(
TARGET
(
kARM
),
zero_ptr
);
}
}
void
pooling2x2s2p1_max
(
const
float
*
din
,
float
*
dout
,
int
num
,
int
chout
,
int
hout
,
int
wout
,
int
chin
,
int
hin
,
int
win
,
int
pad_bottom
,
int
pad_right
)
{
int
size_channel_out
=
wout
*
hout
;
int
size_channel_in
=
win
*
hin
;
auto
data_out
=
static_cast
<
float
*>
(
dout
);
auto
data_in
=
static_cast
<
const
float
*>
(
din
);
const
int
K
=
2
;
const
int
P
=
1
;
const
int
S
=
2
;
int
w_unroll_size
=
wout
/
4
;
int
w_unroll_remian
=
wout
-
w_unroll_size
*
4
;
float32x4_t
vzero
=
vdupq_n_f32
(
std
::
numeric_limits
<
float
>::
lowest
());
if
(
w_unroll_remian
==
0
)
{
w_unroll_size
-=
1
;
w_unroll_remian
=
wout
-
w_unroll_size
*
4
;
}
for
(
int
n
=
0
;
n
<
num
;
++
n
)
{
float
*
data_out_batch
=
data_out
+
n
*
chout
*
size_channel_out
;
const
float
*
data_in_batch
=
data_in
+
n
*
chin
*
size_channel_in
;
#pragma omp parallel for
for
(
int
c
=
0
;
c
<
chout
;
c
++
)
{
float
*
data_out_channel
=
data_out_batch
+
c
*
size_channel_out
;
const
float
*
data_in_channel
=
data_in_batch
+
c
*
size_channel_in
;
const
float
*
r0
=
data_in_channel
;
const
float
*
r1
=
r0
+
win
;
for
(
int
h
=
0
;
h
<
hout
;
h
++
)
{
float
*
dr_out
=
data_out_channel
;
auto
dr0
=
r0
;
auto
dr1
=
r1
;
if
(
h
==
0
)
{
dr0
=
r0
;
dr1
=
r0
;
r0
=
r1
;
r1
=
r0
+
win
;
}
else
{
r0
=
r1
+
win
;
r1
=
r0
+
win
;
}
if
(
h
*
S
+
K
-
P
>
hin
)
{
dr1
=
dr0
;
if
(
h
*
S
+
K
-
P
>
hin
+
1
)
{
memset
(
dr_out
,
0
,
wout
*
sizeof
(
float
));
continue
;
}
}
int
cnt_num
=
w_unroll_size
;
if
(
w_unroll_size
>
0
)
{
#ifdef __aarch64__
asm
volatile
(
P2x2S2_INIT
P2x2S2P1_MAX
P2x2S2P0_MAX
"2:
\n
"
/* end */
:
[
dr0
]
"+r"
(
dr0
),
[
dr1
]
"+r"
(
dr1
),
[
dr_out
]
"+r"
(
dr_out
),
[
cnt_num
]
"+r"
(
cnt_num
)
:
[
vzero
]
"w"
(
vzero
)
:
"cc"
,
"memory"
,
"v0"
,
"v1"
,
"v2"
,
"v3"
,
"v4"
,
"v5"
,
"v6"
,
"v8"
);
#else
asm
volatile
(
P2x2S2_INIT
P2x2S2P1_MAX
P2x2S2P0_MAX
"2:
\n
"
/* end */
:
[
dr0
]
"+r"
(
dr0
),
[
dr1
]
"+r"
(
dr1
),
[
dr_out
]
"+r"
(
dr_out
),
[
cnt_num
]
"+r"
(
cnt_num
)
:
[
vzero
]
"w"
(
vzero
)
:
"cc"
,
"memory"
,
"q0"
,
"q1"
,
"q2"
,
"q3"
,
"q4"
,
"q5"
,
"q8"
,
"q9"
);
#endif
dr0
-=
8
;
dr1
-=
8
;
}
// deal with right pad
int
wstart
=
w_unroll_size
*
4
*
S
-
P
;
for
(
int
j
=
0
;
j
<
w_unroll_remian
;
++
j
)
{
int
wend
=
std
::
min
(
wstart
+
K
,
win
);
int
st
=
wstart
>
0
?
wstart
:
0
;
float
tmp
=
wend
==
st
?
0.
f
:
dr0
[
0
];
for
(
int
i
=
0
;
i
<
wend
-
st
;
i
++
)
{
tmp
=
std
::
max
(
tmp
,
dr0
[
i
]);
tmp
=
std
::
max
(
tmp
,
dr1
[
i
]);
}
*
(
dr_out
++
)
=
tmp
;
dr0
+=
S
-
(
st
-
wstart
);
dr1
+=
S
-
(
st
-
wstart
);
wstart
+=
S
;
}
data_out_channel
+=
wout
;
}
}
}
}
void
pooling2x2s2p1_avg
(
const
float
*
din
,
float
*
dout
,
int
num
,
int
chout
,
int
hout
,
int
wout
,
int
chin
,
int
hin
,
int
win
,
bool
exclusive
,
int
pad_bottom
,
int
pad_right
)
{
int
size_channel_out
=
wout
*
hout
;
int
size_channel_in
=
win
*
hin
;
auto
data_out
=
static_cast
<
float
*>
(
dout
);
auto
data_in
=
static_cast
<
const
float
*>
(
din
);
const
int
K
=
2
;
const
int
P
=
1
;
const
int
S
=
2
;
int
w_unroll_size
=
wout
/
4
;
int
w_unroll_remian
=
wout
-
w_unroll_size
*
4
;
auto
zero_ptr
=
static_cast
<
float
*>
(
TargetMalloc
(
TARGET
(
kARM
),
win
*
sizeof
(
float
)));
float32x4_t
vzero
=
vdupq_n_f32
(
0.
f
);
memset
(
zero_ptr
,
0
,
win
*
sizeof
(
float
));
if
(
w_unroll_remian
==
0
)
{
w_unroll_size
-=
1
;
w_unroll_remian
=
wout
-
w_unroll_size
*
4
;
}
for
(
int
n
=
0
;
n
<
num
;
++
n
)
{
float
*
data_out_batch
=
data_out
+
n
*
chout
*
size_channel_out
;
const
float
*
data_in_batch
=
data_in
+
n
*
chin
*
size_channel_in
;
#pragma omp parallel for
for
(
int
c
=
0
;
c
<
chout
;
c
++
)
{
float
*
data_out_channel
=
data_out_batch
+
c
*
size_channel_out
;
const
float
*
data_in_channel
=
data_in_batch
+
c
*
size_channel_in
;
const
float
*
r0
=
data_in_channel
;
const
float
*
r1
=
r0
+
win
;
for
(
int
h
=
0
;
h
<
hout
;
h
++
)
{
float
*
dr_out
=
data_out_channel
;
auto
dr0
=
r0
;
auto
dr1
=
r1
;
float
coef_h
=
0.5
f
;
if
(
h
==
0
)
{
dr0
=
zero_ptr
;
dr1
=
r0
;
r0
=
r1
;
r1
=
r0
+
win
;
if
(
exclusive
)
{
coef_h
=
1.
f
;
}
}
else
{
r0
=
r1
+
win
;
r1
=
r0
+
win
;
}
if
(
h
*
S
+
K
-
P
>
hin
)
{
dr1
=
zero_ptr
;
if
(
exclusive
)
{
coef_h
=
1.
f
;
}
if
(
h
*
S
+
K
-
P
>
hin
+
1
)
{
memset
(
dr_out
,
0
,
wout
*
sizeof
(
float
));
continue
;
}
}
float
coef_left_most
=
exclusive
?
coef_h
:
coef_h
/
2
;
float32x4_t
vcoef
=
vdupq_n_f32
(
coef_h
/
2
);
float
coef_left
[
4
]
=
{
coef_left_most
,
coef_h
/
2
,
coef_h
/
2
,
coef_h
/
2
};
float32x4_t
vcoef_left
=
vld1q_f32
(
coef_left
);
int
cnt_num
=
w_unroll_size
;
if
(
w_unroll_size
>
0
)
{
#ifdef __aarch64__
asm
volatile
(
P2x2S2_INIT
P2x2S2P1_AVG
P2x2S2P0_AVG
"2:
\n
"
:
[
dr0
]
"+r"
(
dr0
),
[
dr1
]
"+r"
(
dr1
),
[
dr_out
]
"+r"
(
dr_out
),
[
cnt_num
]
"+r"
(
cnt_num
)
:
[
vcoef
]
"w"
(
vcoef
),
[
vzero
]
"w"
(
vzero
),
[
vcoef_left
]
"w"
(
vcoef_left
)
:
"cc"
,
"memory"
,
"v0"
,
"v1"
,
"v2"
,
"v3"
,
"v4"
,
"v5"
,
"v6"
,
"v8"
);
#else
asm
volatile
(
P2x2S2_INIT
P2x2S2P1_AVG
P2x2S2P0_AVG
"2:
\n
"
:
[
dr0
]
"+r"
(
dr0
),
[
dr1
]
"+r"
(
dr1
),
[
dr_out
]
"+r"
(
dr_out
),
[
cnt_num
]
"+r"
(
cnt_num
)
:
[
vcoef
]
"w"
(
vcoef
),
[
vzero
]
"w"
(
vzero
),
[
vcoef_left
]
"w"
(
vcoef_left
)
:
"cc"
,
"memory"
,
"q0"
,
"q1"
,
"q2"
,
"q3"
,
"q4"
,
"q5"
,
"q8"
,
"q9"
);
#endif
dr0
-=
8
;
dr1
-=
8
;
}
// deal with right pad
int
wstart
=
w_unroll_size
*
4
*
S
-
P
;
for
(
int
j
=
0
;
j
<
w_unroll_remian
;
++
j
)
{
int
wend
=
std
::
min
(
wstart
+
K
,
win
);
int
st
=
wstart
>
0
?
wstart
:
0
;
float
tmp
=
0.
f
;
float
coef
=
coef_h
/
2
;
if
(
exclusive
&&
wend
-
st
==
1
)
{
coef
=
coef_h
;
}
for
(
int
i
=
0
;
i
<
wend
-
st
;
i
++
)
{
tmp
+=
dr0
[
i
]
+
dr1
[
i
];
}
*
(
dr_out
++
)
=
tmp
*
coef
;
dr0
+=
S
-
(
st
-
wstart
);
dr1
+=
S
-
(
st
-
wstart
);
wstart
+=
S
;
}
data_out_channel
+=
wout
;
}
}
}
TargetFree
(
TARGET
(
kARM
),
zero_ptr
);
}
void
pooling3x3s1p1_max
(
const
float
*
din
,
void
pooling3x3s1p1_max
(
const
float
*
din
,
float
*
dout
,
float
*
dout
,
int
num
,
int
num
,
...
@@ -2240,6 +2536,9 @@ void pooling3x3s2p0_max(const float* din,
...
@@ -2240,6 +2536,9 @@ void pooling3x3s2p0_max(const float* din,
w_unroll_remian
=
wout
-
w_unroll_size
*
4
;
w_unroll_remian
=
wout
-
w_unroll_size
*
4
;
}
}
int
remain
=
w_unroll_remian
-
1
;
int
right
=
wout
*
2
+
1
-
win
;
// if need right pad
for
(
int
n
=
0
;
n
<
num
;
++
n
)
{
for
(
int
n
=
0
;
n
<
num
;
++
n
)
{
float
*
data_out_batch
=
data_out
+
n
*
chout
*
size_channel_out
;
float
*
data_out_batch
=
data_out
+
n
*
chout
*
size_channel_out
;
const
float
*
data_in_batch
=
data_in
+
n
*
chin
*
size_channel_in
;
const
float
*
data_in_batch
=
data_in
+
n
*
chin
*
size_channel_in
;
...
@@ -2266,6 +2565,7 @@ void pooling3x3s2p0_max(const float* din,
...
@@ -2266,6 +2565,7 @@ void pooling3x3s2p0_max(const float* din,
}
}
}
}
int
cnt_num
=
w_unroll_size
;
int
cnt_num
=
w_unroll_size
;
int
cnt_remain
=
remain
;
if
(
w_unroll_size
>
0
)
{
if
(
w_unroll_size
>
0
)
{
#ifdef __aarch64__
#ifdef __aarch64__
asm
volatile
(
P3x3S2P0_INIT
P3x3S2P0_MAX
asm
volatile
(
P3x3S2P0_INIT
P3x3S2P0_MAX
...
@@ -2289,46 +2589,80 @@ void pooling3x3s2p0_max(const float* din,
...
@@ -2289,46 +2589,80 @@ void pooling3x3s2p0_max(const float* din,
"v9"
,
"v9"
,
"v10"
,
"v10"
,
"v11"
);
"v11"
);
#else
asm
volatile
(
P3x3S2P0_INIT
P3x3S2P0_MAX
:
[
dr0
]
"+r"
(
dr0
),
[
dr1
]
"+r"
(
dr1
),
[
dr2
]
"+r"
(
dr2
),
[
dr_out
]
"+r"
(
dr_out
),
[
cnt_num
]
"+r"
(
cnt_num
)
:
:
"cc"
,
"memory"
,
"q0"
,
"q1"
,
"q2"
,
"q3"
,
"q4"
,
"q5"
,
"q6"
,
"q7"
,
"q8"
,
"q9"
,
"q10"
,
"q11"
);
#endif
dr0
-=
8
;
dr0
-=
8
;
dr1
-=
8
;
dr1
-=
8
;
dr2
-=
8
;
dr2
-=
8
;
}
int
rem
=
win
-
(
w_unroll_size
*
4
)
*
S
;
// deal with right pad
int
wstart
=
0
;
int
rem
=
win
-
(
w_unroll_size
*
4
)
*
S
;
for
(
int
j
=
0
;
j
<
w_unroll_remian
;
++
j
)
{
int
wstart
=
0
;
int
wend
=
std
::
min
(
wstart
+
K
,
rem
);
for
(
int
j
=
0
;
j
<
w_unroll_remian
;
++
j
)
{
float
tmp
=
dr0
[
wstart
];
// std::numeric_limits<float>::min();
int
wend
=
std
::
min
(
wstart
+
K
,
rem
);
for
(
int
i
=
wstart
;
i
<
wend
;
i
++
)
{
float
tmp
=
dr0
[
wstart
];
// std::numeric_limits<float>::min();
tmp
=
std
::
max
(
tmp
,
dr0
[
i
]);
for
(
int
i
=
wstart
;
i
<
wend
;
i
++
)
{
tmp
=
std
::
max
(
tmp
,
dr1
[
i
]);
tmp
=
std
::
max
(
tmp
,
dr0
[
i
]);
tmp
=
std
::
max
(
tmp
,
dr2
[
i
]);
tmp
=
std
::
max
(
tmp
,
dr1
[
i
]);
}
tmp
=
std
::
max
(
tmp
,
dr2
[
i
]);
*
(
dr_out
++
)
=
tmp
;
wstart
+=
S
;
}
}
*
(
dr_out
++
)
=
tmp
;
#else
wstart
+=
S
;
asm
volatile
(
P3x3S2P0_INIT
P3x3S2P0_MAX
"cmp %[remain], #0 @cmp cnt_num
\n
"
"sub %[dr0], #32 @sub - 8
\n
"
"sub %[dr1], #32 @sub - 8
\n
"
"sub %[dr2], #32 @sub - 8
\n
"
"ble 4f @ble exit1
\n
"
"2: @mid loop
\n
"
"vld1.f32 {d0-d1}, [%[dr0]]! @load
\n
"
"vld1.f32 {d2-d3}, [%[dr1]]! @load
\n
"
"vld1.f32 {d4-d5}, [%[dr2]]! @load
\n
"
"vmov.f32 s3,s2 @mov
\n
"
"vmov.f32 s7,s6 @mov
\n
"
"vmov.f32 s11,s10 @mov
\n
"
"vmax.f32 q0, q0, q1 @max n"
"sub %[dr0], #8 @add w
\n
"
"sub %[dr1], #8 @add w
\n
"
"sub %[dr2], #8 @add w
\n
"
"vmax.f32 q0, q0, q2 @max
\n
"
"vpmax.f32 d0, d0, d1 @pmax
\n
"
"vpmax.f32 d0, d0, d0 @pmax
\n
"
"subs %[remain], #1 @subs
\n
"
"vst1.f32 d0[0], [%[dr_out]]! @vst
\n
"
"bne 2b @bne
\n
"
"4: @exit
\n
"
:
[
dr0
]
"+r"
(
dr0
),
[
dr1
]
"+r"
(
dr1
),
[
dr2
]
"+r"
(
dr2
),
[
dr_out
]
"+r"
(
dr_out
),
[
remain
]
"+r"
(
cnt_remain
),
[
cnt_num
]
"+r"
(
cnt_num
)
:
:
"cc"
,
"memory"
,
"q0"
,
"q1"
,
"q2"
,
"q3"
,
"q4"
,
"q5"
,
"q6"
,
"q7"
,
"q8"
,
"q9"
,
"q10"
,
"q11"
);
if
(
right
)
{
int
wstart
=
(
w_unroll_size
*
4
+
remain
)
*
S
;
int
wend
=
std
::
min
(
wstart
+
K
,
win
);
float
tmp
=
dr0
[
wstart
];
// std::numeric_limits<float>::min();
for
(
int
i
=
wstart
;
i
<
wend
;
i
++
)
{
tmp
=
std
::
max
(
tmp
,
std
::
max
(
dr0
[
i
],
dr1
[
i
]));
tmp
=
std
::
max
(
tmp
,
dr2
[
i
]);
}
*
(
dr_out
++
)
=
tmp
;
}
#endif
}
}
r0
=
r2
;
r0
=
r2
;
...
...
lite/backends/opencl/cl_kernel/buffer/sigmoid_kernel.cl
0 → 100644
浏览文件 @
2b32484a
/*
Copyright
(
c
)
2018
PaddlePaddle
Authors.
All
Rights
Reserved.
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
<cl_common.h>
__kernel
void
sigmoid
(
__global
const
CL_DTYPE*
x_data,
const
int
count,
__global
CL_DTYPE*
out_data
)
{
const
int
index
=
get_global_id
(
0
)
;
if
(
index
<
count
)
{
out_data[index]
=
1
/
(
1
+
exp
(
-x_data[index]
))
;
}
}
lite/core/mir/fusion/fc_fuser.cc
浏览文件 @
2b32484a
...
@@ -42,7 +42,6 @@ void FcFuser::BuildPattern() {
...
@@ -42,7 +42,6 @@ void FcFuser::BuildPattern() {
add
->
AsIntermediate
();
add
->
AsIntermediate
();
if
(
act_type_
!=
""
)
{
if
(
act_type_
!=
""
)
{
std
::
cout
<<
"act_type_:"
<<
act_type_
<<
std
::
endl
;
auto
*
add_out
=
VarNode
(
"add_out"
);
auto
*
add_out
=
VarNode
(
"add_out"
);
auto
*
activation
=
OpNode
(
act_type_
,
act_type_
);
auto
*
activation
=
OpNode
(
act_type_
,
act_type_
);
std
::
vector
<
PMNode
*>
act_inputs
{
add_out
};
std
::
vector
<
PMNode
*>
act_inputs
{
add_out
};
...
@@ -51,7 +50,6 @@ void FcFuser::BuildPattern() {
...
@@ -51,7 +50,6 @@ void FcFuser::BuildPattern() {
add_out
->
AsIntermediate
();
add_out
->
AsIntermediate
();
activation
->
AsIntermediate
();
activation
->
AsIntermediate
();
}
else
{
}
else
{
std
::
cout
<<
"act_type_: empty"
<<
std
::
endl
;
add_inputs
>>
*
add
>>
*
Out
;
add_inputs
>>
*
add
>>
*
Out
;
}
}
}
}
...
...
lite/core/program.cc
100755 → 100644
浏览文件 @
2b32484a
文件模式从 100755 更改为 100644
lite/core/tensor.h
100755 → 100644
浏览文件 @
2b32484a
文件模式从 100755 更改为 100644
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