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
3b82bfb5
P
Paddle-Lite
项目概览
PaddlePaddle
/
Paddle-Lite
通知
332
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
3b82bfb5
编写于
6年前
作者:
L
liuruilong
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
optimize batchnormal op kernel
上级
b37c8fef
develop
_release/v2.6.2
gh-pages
github/fork/AnBaolei1984/baolei/bitmain
github/fork/AnBaolei1984/bitmain
github/fork/Cambricon/develop
github/fork/DannyIsFunny/Android5
github/fork/DannyIsFunny/Update_MemOpt
github/fork/DannyIsFunny/fix_pow
github/fork/DannyIsFunny/fix_v26_windows
github/fork/GaoWei8/reduce_sum
github/fork/GaoWei8/reduce_sum_test_con
github/fork/LDOUBLEV/ocr
github/fork/Leonardo-Ding/dwh_dev
github/fork/MaxwellDing/develop
github/fork/MyPandaShaoxiang/fpga_patch
github/fork/MyPandaShaoxiang/int8
github/fork/MyPandaShaoxiang/nlp_correct
github/fork/MyPandaShaoxiang/opencl_valid
github/fork/MyPandaShaoxiang/release/v2.3
github/fork/NHZlX/more_jeston_support
github/fork/PaddleLite-EB/merge1.4
github/fork/PaddleLite-EB/new_dev
github/fork/Wangzheee/matrix_nms_op
github/fork/Xreki/step_rnn/opt_ddim_lite
github/fork/cathwong/patch-1
github/fork/cclauss/patch-1
github/fork/chenjiaoAngel/cherry_pic
github/fork/chenjiaoAngel/conv_dw_5x5
github/fork/chenjiaoAngel/conv_dw_5x5s2
github/fork/edimetia3d/arm_update_elementwise_op
github/fork/edimetia3d/host_deformable_conv
github/fork/edimetia3d/matrix_nms_host
github/fork/edimetia3d/update_pow_op
github/fork/edimetia3d/update_yolo_box
github/fork/haozech/develop
github/fork/haozech/infershape_chz
github/fork/haozech/parl-develop
github/fork/jackzhang235/develop
github/fork/jameswu2014/develop
github/fork/jiansowa/jiansowa/img_nna
github/fork/jiweibo/stream_manage
github/fork/juncaipeng/add_cast
github/fork/lijianshe02/lite-x86
github/fork/qili93/update_sup_model_v26
github/fork/qjing666/develop
github/fork/qnqinan/develop
github/fork/qnqinan/track-develop
github/fork/sangoly/python_compa
github/fork/smilejames/develop
github/fork/sunsetlh/sunsetlh/xpu_multi_test
github/fork/wangqunbaidu/develop
github/fork/weihaoji/whj_27
github/fork/weihaoji/xpu_res2net_fusion
github/fork/weihaoji/xpu_weihaoji_dev
github/fork/xiebaiyuan/fix_leak_opencl
github/fork/xiebaiyuan/opencl_depthwised1
github/fork/xiebaiyuan/opencl_softmax
github/fork/yanghongtian/yanghongtian/add_ascend310_target_place
github/fork/yiicy/computelib
github/fork/yongqiangma/bm_card
github/fork/yongqiangma/calib
github/fork/yongqiangma/copytocpu
github/fork/yongqiangma/gpu
github/fork/yongqiangma/pass
github/fork/yongqiangma/pool
github/fork/yongqiangma/priorbox
github/fork/yongqiangma/shape
github/fork/yongqiangma/split_c
github/fork/yongqiangma/trans
github/fork/yongqiangma/trans2
github/fork/yongqiangma/workspace
github/fork/ysh329/add-cl-kernel-member-for-opencl
github/fork/ysh329/add-get-output
github/fork/ysh329/cherry-pick-precision-profiler-enhance
github/fork/ysh329/fix-opencl-concat
github/fork/ysh329/support-int64-copy-from-to-cpu
github/fork/zhaoyang-star/enable_prifile_in_tiny_publish
github/fork/zhaoyang-star/fix_openc_demo
github/fork/zhaoyang-star/patch-1
github/fork/zhupengyang/opt
release/v2.0.0
release/v2.0.0-beta1
release/v2.0.0-beta2
release/v2.0.0-rc
release/v2.1.0
release/v2.2.0
release/v2.3
release/v2.6
release/v2.6.0
release/v2.7
revert-4368-hongming/test_v26
2.0.0-beta
1.5.0
1.1.1
1.1.0
v2.7-beta
v2.6.3-beta2
v2.6.3-beta1
v2.6.2
v2.6.1
v2.6.0
v2.3.0
v2.2.0
v2.1.0
v2.0.0
v2.0.0-rc
v2.0.0-beta1
v2.0.0-beta1-prerel
v1.0
v0.9
release/1.4
1 合并请求
!772
Metal
变更
10
隐藏空白更改
内联
并排
Showing
10 changed file
with
318 addition
and
126 deletion
+318
-126
src/operators/kernel/arm/batchnorm_kernel.cpp
src/operators/kernel/arm/batchnorm_kernel.cpp
+178
-38
src/operators/kernel/arm/relu_kernel.cpp
src/operators/kernel/arm/relu_kernel.cpp
+61
-61
test/CMakeLists.txt
test/CMakeLists.txt
+2
-0
test/common/test_lib_size.h
test/common/test_lib_size.h
+7
-3
test/framework/test_load.cpp
test/framework/test_load.cpp
+2
-4
test/net/test_googlenet.cpp
test/net/test_googlenet.cpp
+1
-1
test/operators/test_batchnorm_op.cpp
test/operators/test_batchnorm_op.cpp
+21
-19
tools/push2android.sh
tools/push2android.sh
+32
-0
tools/run.sh
tools/run.sh
+7
-0
tools/scripts/run_on_android.sh
tools/scripts/run_on_android.sh
+7
-0
未找到文件。
src/operators/kernel/arm/batchnorm_kernel.cpp
浏览文件 @
3b82bfb5
...
...
@@ -23,7 +23,6 @@ namespace operators {
template
<
>
void
BatchNormKernel
<
CPU
,
float
>::
Compute
(
const
BatchNormParam
&
param
)
const
{
/// todo: test.
const
Tensor
*
input_x
=
param
.
InputX
();
auto
input_x_ptr
=
input_x
->
data
<
float
>
();
const
auto
&
x_dims
=
input_x
->
dims
();
...
...
@@ -46,50 +45,191 @@ void BatchNormKernel<CPU, float>::Compute(const BatchNormParam ¶m) const {
auto
scale_ptr
=
scale
->
data
<
float
>
();
auto
bias_ptr
=
bias
->
data
<
float
>
();
Tensor
inv_std
;
auto
inv_std_ptr
=
inv_std
.
mutable_data
<
float
>
(
make_ddim
({
C
}));
if
(
C
!=
variance
->
numel
())
{
DLOG
<<
"C must equal to variance.numel()"
;
}
assert
(
C
==
variance
->
numel
());
/// std = (var + epsilon).sqrt();
/// inv_std = 1 / std;
for
(
int
i
=
0
;
i
<
C
;
i
++
)
{
inv_std_ptr
[
i
]
=
1
/
static_cast
<
float
>
(
pow
((
variance_ptr
[
i
]
+
epsilon
),
0.5
));
}
// Tensor inv_std;
// auto inv_std_ptr = inv_std.mutable_data<float>(make_ddim({C}));
PADDLE_MOBILE_ENFORCE
(
C
==
variance
->
numel
(),
"C must equal to variance.numel()"
);
int
HXW
=
H
*
W
;
if
(
HXW
>
32
)
{
int
NXC
=
N
*
C
;
float
*
inv_std_ptr
=
new
float
[
NXC
*
4
];
float
*
volatile
new_scale_ptr
=
new
float
[
NXC
*
4
];
float
*
volatile
new_bias_ptr
=
new
float
[
NXC
*
4
];
/// std = (var + epsilon).sqrt();
/// inv_std = 1 / std;
for
(
int
i
=
0
;
i
<
C
*
4
;
i
+=
4
)
{
inv_std_ptr
[
i
]
=
1
/
static_cast
<
float
>
(
pow
((
variance_ptr
[
i
/
4
]
+
epsilon
),
0.5
));
inv_std_ptr
[
i
+
1
]
=
inv_std_ptr
[
i
];
inv_std_ptr
[
i
+
2
]
=
inv_std_ptr
[
i
];
inv_std_ptr
[
i
+
3
]
=
inv_std_ptr
[
i
];
new_scale_ptr
[
i
]
=
inv_std_ptr
[
i
]
*
scale_ptr
[
i
/
4
];
new_scale_ptr
[
i
+
1
]
=
new_scale_ptr
[
i
];
new_scale_ptr
[
i
+
2
]
=
new_scale_ptr
[
i
];
new_scale_ptr
[
i
+
3
]
=
new_scale_ptr
[
i
];
new_bias_ptr
[
i
]
=
bias_ptr
[
i
/
4
]
-
mean_ptr
[
i
/
4
]
*
inv_std_ptr
[
i
]
*
scale_ptr
[
i
/
4
];
new_bias_ptr
[
i
+
1
]
=
new_bias_ptr
[
i
];
new_bias_ptr
[
i
+
2
]
=
new_bias_ptr
[
i
];
new_bias_ptr
[
i
+
3
]
=
new_bias_ptr
[
i
];
}
for
(
int
j
=
C
*
4
;
j
<
NXC
*
4
;
++
j
)
{
new_scale_ptr
[
j
]
=
new_scale_ptr
[
j
-
C
*
4
];
new_bias_ptr
[
j
]
=
new_bias_ptr
[
j
-
C
*
4
];
}
asm
volatile
(
"subs %[N], %[N], #1
\n\t
"
"blt end_n_%=
\n\t
"
"loop_n_%=:
\n\t
"
"subs %[C], %[C], #1
\n\t
"
"blt end_c_%=
\n\t
"
"loop_c_%=:
\n\t
"
"vld1.32 {q9}, [%[new_scale_ptr]]!
\n\t
"
"vld1.32 {q10}, [%[new_bias_ptr]]!
\n\t
"
"mov r6, %[HXW]
\n\t
"
"subs r6, r6, #32
\n\t
"
"blt end_hw_%=
\n\t
"
"loop_hw_%=:
\n\t
"
"vld1.32 {q1, q2}, [%[input_x_ptr]]!
\n\t
"
"vld1.32 {q3, q4}, [%[input_x_ptr]]!
\n\t
"
"vld1.32 {q5, q6}, [%[input_x_ptr]]!
\n\t
"
"vld1.32 {q7, q8}, [%[input_x_ptr]]!
\n\t
"
Tensor
new_scale
;
auto
new_scale_ptr
=
new_scale
.
mutable_data
<
float
>
(
make_ddim
({
C
}));
Tensor
new_bias
;
auto
new_bias_ptr
=
new_bias
.
mutable_data
<
float
>
(
make_ddim
({
C
}));
/// ((x - est_mean) * (inv_var) * scale + bias equal to
/// (x * inv_var * scale) + (bias - est_mean * inv_var * scale)
for
(
int
i
=
0
;
i
<
C
;
i
++
)
{
new_scale_ptr
[
i
]
=
inv_std_ptr
[
i
]
*
scale_ptr
[
i
];
new_bias_ptr
[
i
]
=
bias_ptr
[
i
]
-
mean_ptr
[
i
]
*
inv_std_ptr
[
i
]
*
scale_ptr
[
i
];
{
for
(
int
n
=
0
;
n
<
N
;
n
++
)
{
for
(
int
h
=
0
;
h
<
H
;
h
++
)
{
int
tmp_index
=
n
*
stride0
+
i
*
stride1
+
h
*
stride2
;
for
(
int
w
=
0
;
w
<
W
;
w
++
)
{
int
index
=
tmp_index
+
w
;
out_ptr
[
index
]
=
input_x_ptr
[
index
]
*
new_scale_ptr
[
i
]
+
new_bias_ptr
[
i
];
"vmul.f32 q1, q1, q9
\n\t
"
"vmul.f32 q2, q2, q9
\n\t
"
"vmul.f32 q3, q3, q9
\n\t
"
"vmul.f32 q4, q4, q9
\n\t
"
"vmul.f32 q5, q5, q9
\n\t
"
"vmul.f32 q6, q6, q9
\n\t
"
"vmul.f32 q7, q7, q9
\n\t
"
"vmul.f32 q8, q8, q9
\n\t
"
"vadd.f32 q1, q1, q10
\n\t
"
"vadd.f32 q2, q2, q10
\n\t
"
"vadd.f32 q3, q3, q10
\n\t
"
"vadd.f32 q4, q4, q10
\n\t
"
"vadd.f32 q5, q5, q10
\n\t
"
"vadd.f32 q6, q6, q10
\n\t
"
"vadd.f32 q7, q7, q10
\n\t
"
"vadd.f32 q8, q8, q10
\n\t
"
"vst1.32 {q1, q2}, [%[out_ptr]]!
\n\t
"
"vst1.32 {q3, q4}, [%[out_ptr]]!
\n\t
"
"vst1.32 {q5, q6}, [%[out_ptr]]!
\n\t
"
"vst1.32 {q7, q8}, [%[out_ptr]]!
\n\t
"
"subs r6, r6, #32
\n\t
"
"bge loop_hw_%=
\n\t
"
"end_hw_%=:
\n\t
"
"cmp r6, #0
\n\t
"
"bge end_remainder_%=
\n\t
"
"mov r5, #4
\n\t
"
"mul r6, r6, r5
\n\t
"
"add %[input_x_ptr], %[input_x_ptr], r6
\n\t
"
"vld1.32 {q1, q2}, [%[input_x_ptr]]!
\n\t
"
"vld1.32 {q3, q4}, [%[input_x_ptr]]!
\n\t
"
"vld1.32 {q5, q6}, [%[input_x_ptr]]!
\n\t
"
"vld1.32 {q7, q8}, [%[input_x_ptr]]!
\n\t
"
"vmul.f32 q1, q1, q9
\n\t
"
"vmul.f32 q2, q2, q9
\n\t
"
"vmul.f32 q3, q3, q9
\n\t
"
"vmul.f32 q4, q4, q9
\n\t
"
"vmul.f32 q5, q5, q9
\n\t
"
"vmul.f32 q6, q6, q9
\n\t
"
"vmul.f32 q7, q7, q9
\n\t
"
"vmul.f32 q8, q8, q9
\n\t
"
"vadd.f32 q1, q1, q10
\n\t
"
"vadd.f32 q2, q2, q10
\n\t
"
"vadd.f32 q3, q3, q10
\n\t
"
"vadd.f32 q4, q4, q10
\n\t
"
"vadd.f32 q5, q5, q10
\n\t
"
"vadd.f32 q6, q6, q10
\n\t
"
"vadd.f32 q7, q7, q10
\n\t
"
"vadd.f32 q8, q8, q10
\n\t
"
"add %[out_ptr], %[out_ptr], r6
\n\t
"
"vst1.32 {q1, q2}, [%[out_ptr]]!
\n\t
"
"vst1.32 {q3, q4}, [%[out_ptr]]!
\n\t
"
"vst1.32 {q5, q6}, [%[out_ptr]]!
\n\t
"
"vst1.32 {q7, q8}, [%[out_ptr]]!
\n\t
"
"end_remainder_%=:
\n\t
"
"subs %[C], %[C], #1
\n\t
"
"bge loop_c_%=
\n\t
"
"end_c_%=:
\n\t
"
"subs %[N], %[N], #1
\n\t
"
"bge loop_n_%=
\n\t
"
"end_n_%=:
\n\t
"
:
:
[
input_x_ptr
]
"r"
(
input_x_ptr
),
[
out_ptr
]
"r"
(
out_ptr
),
[
new_scale_ptr
]
"r"
(
new_scale_ptr
),
[
new_bias_ptr
]
"r"
(
new_bias_ptr
),
[
N
]
"r"
(
N
),
[
C
]
"r"
(
C
),
[
HXW
]
"r"
(
HXW
)
:
"memory"
,
"q0"
,
"q1"
,
"q2"
,
"q3"
,
"q4"
,
"q5"
,
"q6"
,
"q7"
,
"q8"
,
"q9"
,
"q10"
,
"r5"
,
"r6"
);
delete
[]
inv_std_ptr
;
delete
[]
new_scale_ptr
;
delete
[]
new_bias_ptr
;
}
else
{
float
*
inv_std_ptr
=
new
float
[
C
];
for
(
int
i
=
0
;
i
<
C
;
i
++
)
{
inv_std_ptr
[
i
]
=
1
/
static_cast
<
float
>
(
pow
((
variance_ptr
[
i
]
+
epsilon
),
0.5
));
}
Tensor
new_scale
;
auto
new_scale_ptr
=
new_scale
.
mutable_data
<
float
>
(
make_ddim
({
C
}));
Tensor
new_bias
;
auto
new_bias_ptr
=
new_bias
.
mutable_data
<
float
>
(
make_ddim
({
C
}));
/// ((x - est_mean) * (inv_var) * scale + bias equal to
/// (x * inv_var * scale) + (bias - est_mean * inv_var * scale)
for
(
int
i
=
0
;
i
<
C
;
i
++
)
{
new_scale_ptr
[
i
]
=
inv_std_ptr
[
i
]
*
scale_ptr
[
i
];
new_bias_ptr
[
i
]
=
bias_ptr
[
i
]
-
mean_ptr
[
i
]
*
inv_std_ptr
[
i
]
*
scale_ptr
[
i
];
{
for
(
int
n
=
0
;
n
<
N
;
n
++
)
{
for
(
int
h
=
0
;
h
<
H
;
h
++
)
{
int
tmp_index
=
n
*
stride0
+
i
*
stride1
+
h
*
stride2
;
for
(
int
w
=
0
;
w
<
W
;
w
++
)
{
int
index
=
tmp_index
+
w
;
out_ptr
[
index
]
=
input_x_ptr
[
index
]
*
new_scale_ptr
[
i
]
+
new_bias_ptr
[
i
];
}
}
}
}
}
delete
[]
inv_std_ptr
;
// DLOG << "input[2,5,1,0](input[102]) ,channel 5 :";
// DLOG << "input_x_ptr : " << input_x_ptr[102];
// DLOG << "variance : " << variance_ptr[5];
// DLOG << "inv_std_ptr : " << inv_std_ptr[5];
// DLOG << "new_scale_ptr : " << new_scale_ptr[5];
// DLOG << "new_bias_ptr : " << new_bias_ptr[5];
// DLOG << "out_ptr : " << out_ptr[102];
}
DLOG
<<
"input[2,5,1,0](input[102]) ,channel 5 :"
;
DLOG
<<
"input_x_ptr : "
<<
input_x_ptr
[
102
];
DLOG
<<
"variance : "
<<
variance_ptr
[
5
];
DLOG
<<
"inv_std_ptr : "
<<
inv_std_ptr
[
5
];
DLOG
<<
"new_scale_ptr : "
<<
new_scale_ptr
[
5
];
DLOG
<<
"new_bias_ptr : "
<<
new_bias_ptr
[
5
];
DLOG
<<
"out_ptr : "
<<
out_ptr
[
102
];
}
}
// namespace operators
}
// namespace paddle_mobile
...
...
This diff is collapsed.
Click to expand it.
src/operators/kernel/arm/relu_kernel.cpp
浏览文件 @
3b82bfb5
...
...
@@ -38,70 +38,70 @@ void ReluKernel<CPU, float>::Compute(const ReluParam ¶m) const {
auto
*
out_ptr
=
out
->
mutable_data
<
float
>
();
int
numel
=
input_x
->
numel
();
if
(
numel
>
32
)
{
asm
volatile
(
"pld [%[input_x_ptr], #0]
\n\t
"
"vmov.f32 q8, #0.0
\n\t
"
"subs %[num], %[num], #32
\n\t
"
"blt end_num_%=
\n\t
"
"loop_num_%=:
\n\t
"
"pld [%[input_x_ptr], #1024]
\n\t
"
"vld1.32 {q0, q1}, [%[input_x_ptr]]!
\n\t
"
"vld1.32 {q2, q3}, [%[input_x_ptr]]!
\n\t
"
"vld1.32 {q4, q5}, [%[input_x_ptr]]!
\n\t
"
"vld1.32 {q6, q7}, [%[input_x_ptr]]!
\n\t
"
"vmax.f32 q0, q0, q8
\n\t
"
"vmax.f32 q1, q1, q8
\n\t
"
"vmax.f32 q2, q2, q8
\n\t
"
"vmax.f32 q3, q3, q8
\n\t
"
"vmax.f32 q4, q4, q8
\n\t
"
"vmax.f32 q5, q5, q8
\n\t
"
"vmax.f32 q6, q6, q8
\n\t
"
"vmax.f32 q7, q7, q8
\n\t
"
"vst1.32 {q0, q1}, [%[out_ptr]]!
\n\t
"
"vst1.32 {q2, q3}, [%[out_ptr]]!
\n\t
"
"vst1.32 {q4, q5}, [%[out_ptr]]!
\n\t
"
"vst1.32 {q6, q7}, [%[out_ptr]]!
\n\t
"
"subs %[num], %[num], #32
\n\t
"
"bge loop_num_%=
\n\t
"
"end_num_%=:
\n\t
"
"cmp %[num], #0
\n\t
"
"bge end_%=
\n\t
"
"mov r6, #4
\n\t
"
"mul r5, %[num], r6
\n\t
"
"add %[input_x_ptr], %[input_x_ptr], r5
\n\t
"
"vld1.32 {q0, q1}, [%[input_x_ptr]]!
\n\t
"
"vld1.32 {q2, q3}, [%[input_x_ptr]]!
\n\t
"
"vld1.32 {q4, q5}, [%[input_x_ptr]]!
\n\t
"
"vld1.32 {q6, q7}, [%[input_x_ptr]]!
\n\t
"
"vmax.f32 q0, q0, q8
\n\t
"
"vmax.f32 q1, q1, q8
\n\t
"
"vmax.f32 q2, q2, q8
\n\t
"
"vmax.f32 q3, q3, q8
\n\t
"
"vmax.f32 q4, q4, q8
\n\t
"
"vmax.f32 q5, q5, q8
\n\t
"
"vmax.f32 q6, q6, q8
\n\t
"
"vmax.f32 q7, q7, q8
\n\t
"
"add %[out_ptr], %[out_ptr], r5
\n\t
"
"vst1.32 {q0, q1}, [%[out_ptr]]!
\n\t
"
"vst1.32 {q2, q3}, [%[out_ptr]]!
\n\t
"
"vst1.32 {q4, q5}, [%[out_ptr]]!
\n\t
"
"vst1.32 {q6, q7}, [%[out_ptr]]!
\n\t
"
"end_%=:
\n\t
"
:
:
[
out_ptr
]
"r"
(
out_ptr
),
[
input_x_ptr
]
"r"
(
input_x_ptr
),
[
num
]
"r"
(
numel
)
:
"memory"
,
"q0"
,
"q1"
,
"q2"
,
"q3"
,
"q4"
,
"q5"
,
"q6"
,
"q7"
,
"q8"
,
"r5"
,
"r6"
);
}
else
{
// if (numel > 64
) {
//
asm volatile(
//
"pld [%[input_x_ptr], #0] \n\t"
//
"vmov.f32 q8, #0.0 \n\t"
//
"subs %[num], %[num], #32 \n\t"
//
"blt end_num_%= \n\t"
//
"loop_num_%=: \n\t"
//
"pld [%[input_x_ptr], #1024] \n\t"
//
//
"vld1.32 {q0, q1}, [%[input_x_ptr]]! \n\t"
//
"vld1.32 {q2, q3}, [%[input_x_ptr]]! \n\t"
//
"vld1.32 {q4, q5}, [%[input_x_ptr]]! \n\t"
//
"vld1.32 {q6, q7}, [%[input_x_ptr]]! \n\t"
//
//
"vmax.f32 q0, q0, q8 \n\t"
//
"vmax.f32 q1, q1, q8 \n\t"
//
"vmax.f32 q2, q2, q8 \n\t"
//
"vmax.f32 q3, q3, q8 \n\t"
//
"vmax.f32 q4, q4, q8 \n\t"
//
"vmax.f32 q5, q5, q8 \n\t"
//
"vmax.f32 q6, q6, q8 \n\t"
//
"vmax.f32 q7, q7, q8 \n\t"
//
//
"vst1.32 {q0, q1}, [%[out_ptr]]! \n\t"
//
"vst1.32 {q2, q3}, [%[out_ptr]]! \n\t"
//
"vst1.32 {q4, q5}, [%[out_ptr]]! \n\t"
//
"vst1.32 {q6, q7}, [%[out_ptr]]! \n\t"
//
//
"subs %[num], %[num], #32 \n\t"
//
"bge loop_num_%= \n\t"
//
"end_num_%=: \n\t"
//
"cmp %[num], #0 \n\t"
//
"bge end_%= \n\t"
//
"mov r6, #4 \n\t"
//
"mul r5, %[num], r6 \n\t"
//
"add %[input_x_ptr], %[input_x_ptr], r5 \n\t"
//
"vld1.32 {q0, q1}, [%[input_x_ptr]]! \n\t"
//
"vld1.32 {q2, q3}, [%[input_x_ptr]]! \n\t"
//
"vld1.32 {q4, q5}, [%[input_x_ptr]]! \n\t"
//
"vld1.32 {q6, q7}, [%[input_x_ptr]]! \n\t"
//
"vmax.f32 q0, q0, q8 \n\t"
//
"vmax.f32 q1, q1, q8 \n\t"
//
"vmax.f32 q2, q2, q8 \n\t"
//
"vmax.f32 q3, q3, q8 \n\t"
//
"vmax.f32 q4, q4, q8 \n\t"
//
"vmax.f32 q5, q5, q8 \n\t"
//
"vmax.f32 q6, q6, q8 \n\t"
//
"vmax.f32 q7, q7, q8 \n\t"
//
"add %[out_ptr], %[out_ptr], r5 \n\t"
//
"vst1.32 {q0, q1}, [%[out_ptr]]! \n\t"
//
"vst1.32 {q2, q3}, [%[out_ptr]]! \n\t"
//
"vst1.32 {q4, q5}, [%[out_ptr]]! \n\t"
//
"vst1.32 {q6, q7}, [%[out_ptr]]! \n\t"
//
"end_%=: \n\t"
//
:
//
:
//
[out_ptr] "r"(out_ptr), [input_x_ptr] "r"(input_x_ptr), [num] "r"(numel)
//
: "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "r5",
//
"r6");
//
} else {
ReluFunctor
<
float
>
func_
;
math
::
Transform
trans
;
trans
(
input_x_ptr
,
input_x_ptr
+
numel
,
out_ptr
,
func_
);
}
//
}
}
}
// namespace operators
}
// namespace paddle_mobile
...
...
This diff is collapsed.
Click to expand it.
test/CMakeLists.txt
浏览文件 @
3b82bfb5
...
...
@@ -137,4 +137,6 @@ else ()
ADD_EXECUTABLE
(
test-depthwise-conv-op operators/test_depthwise_conv_op.cpp test_helper.h test_include.h executor_for_test.h
)
target_link_libraries
(
test-depthwise-conv-op paddle-mobile
)
#add_library(test-lib-size SHARED common/test_lib_size.h common/test_lib_size.cpp)
endif
()
This diff is collapsed.
Click to expand it.
test/common/test_lib_size.h
浏览文件 @
3b82bfb5
...
...
@@ -20,6 +20,8 @@ limitations under the License. */
#define PADDLE_MOBILE_TEST_LIB_SIZE_H
#include <vector>
#include <pthread.h>
#include <thread>
//#include <list>
//#include <tuple>
//#include <typeinfo>
...
...
@@ -33,7 +35,7 @@ limitations under the License. */
//#include <iostream>
//#include <sstream>
#include <memory>
//
#include <memory>
//#include <stdio.h>
//#include <cstring>
...
...
@@ -44,8 +46,10 @@ void foo() {
// std::cout << "12345" << std::endl;
std
::
vector
<
int
>
vec
=
{
1
,
2
,
3
,
4
,
5
};
vec
.
push_back
(
2
);
// std::find(vec.begin(), vec.end(), 1);
pthread_mutex_init
(
NULL
,
NULL
);
pthread_attr_destroy
(
NULL
);
// std::find(vec.begin(), vec.end(), 1);
// std::list<int> l;
...
...
@@ -70,7 +74,7 @@ void foo() {
// int z = 10;
// }
std
::
shared_ptr
<
int
>
s1
=
std
::
make_shared
<
int
>
();
//
std::shared_ptr<int> s1 = std::make_shared<int>();
// std::stringstream ss;
// ss << "12345";
...
...
This diff is collapsed.
Click to expand it.
test/framework/test_load.cpp
浏览文件 @
3b82bfb5
...
...
@@ -19,11 +19,9 @@ int main() {
paddle_mobile
::
Loader
<
paddle_mobile
::
CPU
>
loader
;
// ../../../test/models/googlenet
// ../../../test/models/mobilenet
auto
program
=
loader
.
Load
(
g_
googlenet
,
true
,
tru
e
);
auto
program
=
loader
.
Load
(
g_
mobilenet_ssd
,
false
,
fals
e
);
// loader.Load(g_googlenet_combine + "/model", g_googlenet_combine +
// "/params",
// true);
// "/params", true);
program
.
originProgram
->
Description
(
"program desc: "
);
return
0
;
}
This diff is collapsed.
Click to expand it.
test/net/test_googlenet.cpp
浏览文件 @
3b82bfb5
...
...
@@ -18,7 +18,7 @@ limitations under the License. */
int
main
()
{
paddle_mobile
::
Loader
<
paddle_mobile
::
CPU
>
loader
;
bool
optimize
=
fals
e
;
bool
optimize
=
tru
e
;
auto
time1
=
time
();
auto
program
=
loader
.
Load
(
g_googlenet
,
optimize
);
// auto program = loader.Load(g_googlenet_combine + "/model",
...
...
This diff is collapsed.
Click to expand it.
test/operators/test_batchnorm_op.cpp
浏览文件 @
3b82bfb5
...
...
@@ -41,7 +41,7 @@ class TestBatchNormOp {
for
(
int
j
=
0
;
j
<
ops
.
size
();
++
j
)
{
std
::
shared_ptr
<
OpDesc
>
op
=
ops
[
j
];
if
(
op
->
Type
()
==
"batch_norm"
&&
op
->
Input
(
"X"
)[
0
]
==
"conv2d_
0
.tmp_0"
)
{
op
->
Input
(
"X"
)[
0
]
==
"conv2d_
5
.tmp_0"
)
{
DLOG
<<
" mul attr size: "
<<
op
->
GetAttrMap
().
size
();
DLOG
<<
" inputs size: "
<<
op
->
GetInputs
().
size
();
DLOG
<<
" outputs size: "
<<
op
->
GetOutputs
().
size
();
...
...
@@ -67,29 +67,29 @@ class TestBatchNormOp {
const
Tensor
&
t5
)
{
// feed
auto
scope
=
program_
.
scope
;
Variable
*
x1_feed_value
=
scope
->
Var
(
"conv2d_
0
.tmp_0"
);
Variable
*
x1_feed_value
=
scope
->
Var
(
"conv2d_
5
.tmp_0"
);
auto
tensor_x1
=
x1_feed_value
->
GetMutable
<
LoDTensor
>
();
tensor_x1
->
ShareDataWith
(
t1
);
Variable
*
mean_feed_value
=
scope
->
Var
(
"batch_norm_0.w_1"
);
Variable
*
mean_feed_value
=
scope
->
Var
(
"batch_norm_
1
0.w_1"
);
auto
tensor_mean
=
mean_feed_value
->
GetMutable
<
LoDTensor
>
();
tensor_mean
->
ShareDataWith
(
t2
);
Variable
*
scale_feed_value
=
scope
->
Var
(
"batch_norm_0.w_0"
);
Variable
*
scale_feed_value
=
scope
->
Var
(
"batch_norm_
1
0.w_0"
);
auto
tensor_scale
=
scale_feed_value
->
GetMutable
<
LoDTensor
>
();
tensor_scale
->
ShareDataWith
(
t3
);
Variable
*
variance_feed_value
=
scope
->
Var
(
"batch_norm_0.w_2"
);
Variable
*
variance_feed_value
=
scope
->
Var
(
"batch_norm_
1
0.w_2"
);
auto
tensor_variance
=
variance_feed_value
->
GetMutable
<
LoDTensor
>
();
tensor_variance
->
ShareDataWith
(
t4
);
Variable
*
bias_feed_value
=
scope
->
Var
(
"batch_norm_0.b_0"
);
Variable
*
bias_feed_value
=
scope
->
Var
(
"batch_norm_
1
0.b_0"
);
auto
tensor_bias
=
bias_feed_value
->
GetMutable
<
LoDTensor
>
();
tensor_bias
->
ShareDataWith
(
t5
);
Variable
*
output
=
scope
->
Var
(
"batch_norm_0.tmp_2"
);
Variable
*
output
=
scope
->
Var
(
"batch_norm_
1
0.tmp_2"
);
auto
*
output_tensor
=
output
->
GetMutable
<
LoDTensor
>
();
output_tensor
->
mutable_data
<
float
>
({
4
,
10
,
2
,
2
});
output_tensor
->
mutable_data
<
float
>
({
1
,
256
,
38
,
38
});
// DLOG << typeid(output_tensor).name();
// DLOG << "output_tensor dims: " << output_tensor->dims();
...
...
@@ -128,30 +128,30 @@ int main() {
DLOG
<<
"----------**********----------"
;
DLOG
<<
"begin to run BatchNormOp Test"
;
paddle_mobile
::
Loader
<
paddle_mobile
::
CPU
>
loader
;
auto
program
=
loader
.
Load
(
std
::
string
(
g_
resnet
));
auto
program
=
loader
.
Load
(
std
::
string
(
g_
mobilenet_ssd
));
/// input x (4,10,2,2)
paddle_mobile
::
framework
::
Tensor
inputx1
;
SetupTensor
<
float
>
(
&
inputx1
,
{
4
,
10
,
2
,
2
},
static_cast
<
float
>
(
0
),
SetupTensor
<
float
>
(
&
inputx1
,
{
1
,
256
,
38
,
38
},
static_cast
<
float
>
(
0
),
static_cast
<
float
>
(
1
));
auto
*
inputx1_ptr
=
inputx1
.
data
<
float
>
();
paddle_mobile
::
framework
::
Tensor
mean
;
SetupTensor
<
float
>
(
&
mean
,
{
10
},
static_cast
<
float
>
(
0
),
static_cast
<
float
>
(
1
));
SetupTensor
<
float
>
(
&
mean
,
{
256
},
static_cast
<
float
>
(
0
),
static_cast
<
float
>
(
1
));
auto
*
mean_ptr
=
mean
.
data
<
float
>
();
paddle_mobile
::
framework
::
Tensor
scale
;
SetupTensor
<
float
>
(
&
scale
,
{
10
},
static_cast
<
float
>
(
0
),
SetupTensor
<
float
>
(
&
scale
,
{
256
},
static_cast
<
float
>
(
0
),
static_cast
<
float
>
(
1
));
auto
*
scale_ptr
=
scale
.
data
<
float
>
();
paddle_mobile
::
framework
::
Tensor
variance
;
SetupTensor
<
float
>
(
&
variance
,
{
10
},
static_cast
<
float
>
(
0
),
SetupTensor
<
float
>
(
&
variance
,
{
256
},
static_cast
<
float
>
(
0
),
static_cast
<
float
>
(
1
));
auto
*
variance_ptr
=
variance
.
data
<
float
>
();
paddle_mobile
::
framework
::
Tensor
bias
;
SetupTensor
<
float
>
(
&
bias
,
{
10
},
static_cast
<
float
>
(
0
),
static_cast
<
float
>
(
1
));
SetupTensor
<
float
>
(
&
bias
,
{
256
},
static_cast
<
float
>
(
0
),
static_cast
<
float
>
(
1
));
auto
*
bias_ptr
=
bias
.
data
<
float
>
();
paddle_mobile
::
framework
::
TestBatchNormOp
<
paddle_mobile
::
CPU
>
testBatchNormOp
(
...
...
@@ -161,11 +161,13 @@ int main() {
testBatchNormOp
.
predict_bn
(
inputx1
,
mean
,
scale
,
variance
,
bias
);
auto
*
output_bn_ptr
=
output_bn
->
data
<
float
>
();
/// [2, 5, 1, 0]
DLOG
<<
" ("
<<
inputx1_ptr
[
102
]
<<
" - "
<<
mean_ptr
[
5
]
<<
")/(("
<<
variance_ptr
[
5
]
<<
" + 0.00001"
<<
")^0.5)* "
<<
scale_ptr
[
5
]
<<
" + "
<<
bias_ptr
[
5
]
<<
" = "
;
DLOG
<<
output_bn_ptr
[
102
];
DLOG
<<
" ("
<<
inputx1_ptr
[
0
]
<<
" - "
<<
mean_ptr
[
0
]
<<
")/(("
<<
variance_ptr
[
0
]
<<
" + 0.00001"
<<
")^0.5)* "
<<
scale_ptr
[
0
]
<<
" + "
<<
bias_ptr
[
0
]
<<
" = "
;
DLOG
<<
output_bn_ptr
[
0
];
DLOG
<<
"input_ptr 0 : "
<<
inputx1_ptr
[
0
];
DLOG
<<
"output_ptr 0 : "
<<
output_bn_ptr
[
0
];
return
0
;
}
This diff is collapsed.
Click to expand it.
tools/push2android.sh
0 → 100644
浏览文件 @
3b82bfb5
#!/usr/bin/env sh
push_fn
()
{
MODELS_PATH
=
"../test/models/*"
MODELS_SRC
=
"../test/models"
IMAGE_PATH
=
"../test/images/*"
EXE_FILE
=
"../test/build/*"
EXE_DIR
=
"data/local/tmp/bin"
adb shell
mkdir
${
EXE_DIR
}
MODELS_DIR
=
"data/local/tmp/models"
adb shell
mkdir
${
MODELS_DIR
}
for
file
in
`
ls
${
MODELS_SRC
}
`
do
adb shell
mkdir
${
MODELS_DIR
}
"/"
${
file
}
done
IMAGES_DIR
=
"data/local/tmp/images"
adb shell
mkdir
${
IMAGES_DIR
}
LIB_PATH
=
"../build/release/arm-v7a/build/*"
adb push
${
EXE_FILE
}
${
EXE_DIR
}
adb push
${
LIB_PATH
}
${
EXE_DIR
}
if
[[
$1
!=
"npm"
]]
;
then
adb push
${
IMAGE_PATH
}
${
IMAGES_DIR
}
adb push
${
MODELS_PATH
}
${
MODELS_DIR
}
fi
}
if
[[
$1
==
"npm"
]]
;
then
push_fn
$1
else
push_fn
fi
This diff is collapsed.
Click to expand it.
tools/run.sh
浏览文件 @
3b82bfb5
...
...
@@ -24,8 +24,15 @@ adb shell mkdir ${IMAGES_DIR}
LIB_PATH
=
"../build/release/arm-v7a/build/*"
adb push
${
EXE_FILE
}
${
EXE_DIR
}
adb push
${
LIB_PATH
}
${
EXE_DIR
}
if
[[
$1
!=
"npm"
]]
;
then
adb push
${
IMAGE_PATH
}
${
IMAGES_DIR
}
adb push
${
MODELS_PATH
}
${
MODELS_DIR
}
fi
adb shell
"cd /data/local/tmp/bin; LD_LIBRARY_PATH=. ./
${
TESTUNIT
}
"
}
if
[[
$1
==
"npm"
]]
;
then
push_fn
$1
else
push_fn
fi
\ No newline at end of file
This diff is collapsed.
Click to expand it.
tools/scripts/run_on_android.sh
浏览文件 @
3b82bfb5
...
...
@@ -19,12 +19,19 @@ adb shell mkdir ${IMAGES_DIR}
LIB_PATH
=
"../../build/release/arm-v7a/build/*"
adb push
${
EXE_FILE
}
${
EXE_DIR
}
adb push
${
LIB_PATH
}
${
EXE_DIR
}
if
[[
$1
!=
"npm"
]]
;
then
adb push
${
IMAGE_PATH
}
${
IMAGES_DIR
}
adb push
${
MODELS_PATH
}
${
MODELS_DIR
}
fi
echo
"test-op or test-net below : "
adb shell
ls
/data/local/tmp/bin
echo
"**** choose OP or NET to test ****"
read
-p
"which to test : "
test_name
adb shell
"cd /data/local/tmp/bin; LD_LIBRARY_PATH=. ./
${
test_name
}
"
}
if
[[
$1
==
"npm"
]]
;
then
push_fn
$1
else
push_fn
fi
\ No newline at end of file
This diff is collapsed.
Click to expand it.
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
反馈
建议
客服
返回
顶部