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
a96d013c
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
a96d013c
编写于
9月 27, 2019
作者:
N
NazgulLee
提交者:
Yanzhan Yang
9月 27, 2019
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
optimize instance norm. test=develop (#2120)
上级
4139dc2c
变更
5
隐藏空白更改
内联
并排
Showing
5 changed file
with
142 addition
and
115 deletion
+142
-115
mobile/src/operators/kernel/cl/cl-kernel-func/instancenorm_func.cpp
.../operators/kernel/cl/cl-kernel-func/instancenorm_func.cpp
+77
-0
mobile/src/operators/kernel/cl/cl-kernel-func/instancenorm_func.h
...rc/operators/kernel/cl/cl-kernel-func/instancenorm_func.h
+27
-0
mobile/src/operators/kernel/cl/cl_kernel/instancenorm_kernel.cl
.../src/operators/kernel/cl/cl_kernel/instancenorm_kernel.cl
+12
-5
mobile/src/operators/kernel/cl/instancenorm_kernel.cpp
mobile/src/operators/kernel/cl/instancenorm_kernel.cpp
+14
-56
mobile/src/operators/kernel/cl/instancenorm_relu_kernel.cpp
mobile/src/operators/kernel/cl/instancenorm_relu_kernel.cpp
+12
-54
未找到文件。
mobile/src/operators/kernel/cl/cl-kernel-func/instancenorm_func.cpp
0 → 100644
浏览文件 @
a96d013c
/* 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 "operators/kernel/cl/cl-kernel-func/instancenorm_func.h"
#include <algorithm>
namespace
paddle_mobile
{
namespace
operators
{
void
InstanceNorm
(
framework
::
CLHelper
*
cl_helper
,
const
InstanceNormParam
<
GPU_CL
>
&
param
)
{
auto
kernel
=
cl_helper
->
KernelAt
(
0
);
auto
&
dims
=
param
.
Out
()
->
dims
();
const
int
n
=
dims
[
0
];
const
int
c_group
=
(
dims
[
1
]
+
3
)
/
4
;
const
int
h
=
dims
[
2
];
const
int
w
=
dims
[
3
];
auto
epsilon
=
param
.
Epsilon
();
auto
input
=
param
.
InputX
()
->
GetCLImage
();
auto
out
=
param
.
Out
()
->
GetCLImage
();
// DLOG << "Epsilon: " << epsilon;
auto
local_work_size_info
=
cl_helper
->
LocalWorkSizeInfo
();
//
// DLOG << local_work_size_info.max_work_group_size;
// DLOG << local_work_size_info.max_work_item_size0;
// DLOG << local_work_size_info.max_work_item_size1;
// DLOG << local_work_size_info.max_work_item_size2;
int
maxTotal
=
std
::
min
(
static_cast
<
int
>
(
local_work_size_info
.
max_work_group_size
),
256
);
int
local_work_size1
=
std
::
min
(
static_cast
<
int
>
(
local_work_size_info
.
max_work_item_size1
),
std
::
min
(
256
,
w
));
int
local_work_size2
=
1
;
const
size_t
work_size
[
3
]
=
{(
size_t
)(
n
*
c_group
),
(
size_t
)
local_work_size1
,
(
size_t
)
local_work_size2
};
const
size_t
local_work_size
[
3
]
=
{(
size_t
)
1
,
(
size_t
)
local_work_size1
,
(
size_t
)
local_work_size2
};
// DLOG << "work_size" << work_size[0] << " " << work_size[1] << " "
// << work_size[2];
// DLOG << "local_work_size" << local_work_size[0] << " " <<
// local_work_size[1]
// << " " << local_work_size[2];
cl_int
status
;
clSetKernelArg
(
kernel
,
0
,
sizeof
(
cl_int
),
&
w
);
CL_CHECK_ERRORS
(
status
);
clSetKernelArg
(
kernel
,
1
,
sizeof
(
cl_int
),
&
h
);
CL_CHECK_ERRORS
(
status
);
clSetKernelArg
(
kernel
,
2
,
sizeof
(
cl_int
),
&
c_group
);
CL_CHECK_ERRORS
(
status
);
clSetKernelArg
(
kernel
,
3
,
sizeof
(
cl_int
),
&
local_work_size1
);
CL_CHECK_ERRORS
(
status
);
clSetKernelArg
(
kernel
,
4
,
sizeof
(
cl_int
),
&
local_work_size2
);
CL_CHECK_ERRORS
(
status
);
clSetKernelArg
(
kernel
,
5
,
sizeof
(
cl_float
),
&
epsilon
);
CL_CHECK_ERRORS
(
status
);
clSetKernelArg
(
kernel
,
6
,
sizeof
(
cl_mem
),
&
input
);
CL_CHECK_ERRORS
(
status
);
clSetKernelArg
(
kernel
,
7
,
sizeof
(
cl_mem
),
&
out
);
CL_CHECK_ERRORS
(
status
);
clEnqueueNDRangeKernel
(
cl_helper
->
CLCommandQueue
(),
kernel
,
3
,
NULL
,
work_size
,
local_work_size
,
0
,
NULL
,
NULL
);
}
}
// namespace operators
}
// namespace paddle_mobile
mobile/src/operators/kernel/cl/cl-kernel-func/instancenorm_func.h
0 → 100644
浏览文件 @
a96d013c
/* 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. */
#if defined(INSTANCENORM_OP) || defined(FUSION_INSTANCENORM_RELU_OP)
#pragma once
#include "framework/cl/cl_helper.h"
#include "operators/op_param.h"
namespace
paddle_mobile
{
namespace
operators
{
void
InstanceNorm
(
framework
::
CLHelper
*
cl_helper
,
const
InstanceNormParam
<
GPU_CL
>
&
param
);
}
}
// namespace paddle_mobile
#endif
mobile/src/operators/kernel/cl/cl_kernel/instancenorm_kernel.cl
浏览文件 @
a96d013c
...
...
@@ -32,13 +32,19 @@ __kernel void instancenorm(__private const int in_width,
const
sampler_t
sampler
=
CLK_NORMALIZED_COORDS_TRUE
| CLK_ADDRESS_CLAMP |
CLK_FILTER_NEAREST
;
#
ifdef
LOCAL_MEM_128
__local
float4
shared_mem[128]
;
#
elif
defined
(
LOCAL_MEM_64
)
__local
float4
shared_mem[64]
;
#
else
__local
float4
shared_mem[256]
;
#
endif
int
xOffset
=
c
*
in_width
;
int
yOffset
=
n
*
in_height
;
float4
sum
=
0.0f
;
for
(
int
xIndex
=
w
; xIndex < in_width; xIndex += local_work_size_x) {
for
(
int
yIndex
=
h
; yIndex < in_height; yIndex += local_work_size_y) {
sum
+=
read_imagef
(
input,
sampler,
(
int2
)(
mad24
(
c,
in_width,
xIndex
)
,
mad24
(
n,
in_height,
yIndex
)
))
;
sum
+=
read_imagef
(
input,
sampler,
(
int2
)(
xOffset
+
xIndex,
yOffset
+
yIndex
))
;
}
}
shared_mem[local_id]
=
sum
;
...
...
@@ -73,7 +79,8 @@ __kernel void instancenorm(__private const int in_width,
sum
=
0.0f
;
for
(
int
xIndex
=
w
; xIndex < in_width; xIndex += local_work_size_x) {
for
(
int
yIndex
=
h
; yIndex < in_height; yIndex += local_work_size_y) {
sum
+=
pow
(
read_imagef
(
input,
sampler,
(
int2
)(
mad24
(
c,
in_width,
xIndex
)
,
mad24
(
n,
in_height,
yIndex
)))
-
mean_val,
2
)
;
float4
temp
=
read_imagef
(
input,
sampler,
(
int2
)(
xOffset
+
xIndex,
yOffset
+
yIndex
))
-
mean_val
;
sum
+=
temp
*
temp
;
}
}
shared_mem[local_id]
=
sum
;
...
...
@@ -107,7 +114,7 @@ __kernel void instancenorm(__private const int in_width,
for
(
int
xIndex
=
w
; xIndex < in_width; xIndex += local_work_size_x) {
for
(
int
yIndex
=
h
; yIndex < in_height; yIndex += local_work_size_y) {
int2
intout_pos
=
(
int2
)(
mad24
(
c,
in_width,
xIndex
)
,
mad24
(
n,
in_height,
yIndex
)
)
;
int2
intout_pos
=
(
int2
)(
xOffset
+
xIndex,
yOffset
+
yIndex
)
;
float4
in_val
=
read_imagef
(
input,
sampler,
intout_pos
)
;
half4
out_val
=
convert_half4
((
in_val
-
mean_val
)
*
s
)
;
#
ifdef
RELU
...
...
mobile/src/operators/kernel/cl/instancenorm_kernel.cpp
浏览文件 @
a96d013c
...
...
@@ -16,74 +16,32 @@ limitations under the License. */
#include "operators/kernel/instancenorm_kernel.h"
#include <cmath>
#include "operators/kernel/cl/cl-kernel-func/instancenorm_func.h"
namespace
paddle_mobile
{
namespace
operators
{
template
<
>
bool
InstanceNormKernel
<
GPU_CL
,
float
>::
Init
(
InstanceNormParam
<
GPU_CL
>
*
param
)
{
this
->
cl_helper_
.
AddKernel
(
"instancenorm"
,
"instancenorm_kernel.cl"
);
auto
&
dims
=
param
->
Out
()
->
dims
();
const
int
h
=
dims
[
2
];
std
::
string
build_options
=
""
;
if
(
h
==
128
)
{
build_options
=
"-DLOCAL_MEM_128"
;
}
else
if
(
h
==
64
)
{
build_options
=
"-DLOCAL_MEM_64"
;
}
else
if
(
h
>
256
)
{
PADDLE_MOBILE_THROW_EXCEPTION
(
"instance norm unsupported input height"
);
}
this
->
cl_helper_
.
AddKernel
(
"instancenorm"
,
"instancenorm_kernel.cl"
,
build_options
);
return
true
;
}
template
<
>
void
InstanceNormKernel
<
GPU_CL
,
float
>::
Compute
(
const
InstanceNormParam
<
GPU_CL
>
&
param
)
{
auto
kernel
=
this
->
cl_helper_
.
KernelAt
(
0
);
auto
&
dims
=
param
.
Out
()
->
dims
();
const
int
n
=
dims
[
0
];
const
int
c_group
=
(
dims
[
1
]
+
3
)
/
4
;
const
int
h
=
dims
[
2
];
const
int
w
=
dims
[
3
];
auto
epsilon
=
param
.
Epsilon
();
auto
input
=
param
.
InputX
()
->
GetCLImage
();
auto
out
=
param
.
Out
()
->
GetCLImage
();
DLOG
<<
"Epsilon: "
<<
epsilon
;
auto
local_work_size_info
=
this
->
cl_helper_
.
LocalWorkSizeInfo
();
DLOG
<<
local_work_size_info
.
max_work_group_size
;
DLOG
<<
local_work_size_info
.
max_work_item_size0
;
DLOG
<<
local_work_size_info
.
max_work_item_size1
;
DLOG
<<
local_work_size_info
.
max_work_item_size2
;
int
local_work_size1
=
std
::
min
(
static_cast
<
int
>
(
local_work_size_info
.
max_work_item_size1
),
std
::
min
(
256
,
w
));
int
local_work_size2
=
1
;
const
size_t
work_size
[
3
]
=
{(
size_t
)(
n
*
c_group
),
(
size_t
)
local_work_size1
,
(
size_t
)
local_work_size2
};
const
size_t
local_work_size
[
3
]
=
{(
size_t
)
1
,
(
size_t
)
local_work_size1
,
(
size_t
)
local_work_size2
};
DLOG
<<
"work_size"
<<
work_size
[
0
]
<<
" "
<<
work_size
[
1
]
<<
" "
<<
work_size
[
2
];
DLOG
<<
"local_work_size"
<<
local_work_size
[
0
]
<<
" "
<<
local_work_size
[
1
]
<<
" "
<<
local_work_size
[
2
];
cl_int
status
;
status
=
clSetKernelArg
(
kernel
,
0
,
sizeof
(
cl_int
),
&
w
);
CL_CHECK_ERRORS
(
status
);
status
=
clSetKernelArg
(
kernel
,
1
,
sizeof
(
cl_int
),
&
h
);
CL_CHECK_ERRORS
(
status
);
status
=
clSetKernelArg
(
kernel
,
2
,
sizeof
(
cl_int
),
&
c_group
);
CL_CHECK_ERRORS
(
status
);
status
=
clSetKernelArg
(
kernel
,
3
,
sizeof
(
cl_int
),
&
local_work_size1
);
CL_CHECK_ERRORS
(
status
);
status
=
clSetKernelArg
(
kernel
,
4
,
sizeof
(
cl_int
),
&
local_work_size2
);
CL_CHECK_ERRORS
(
status
);
status
=
clSetKernelArg
(
kernel
,
5
,
sizeof
(
cl_float
),
&
epsilon
);
CL_CHECK_ERRORS
(
status
);
status
=
clSetKernelArg
(
kernel
,
6
,
sizeof
(
cl_mem
),
&
input
);
CL_CHECK_ERRORS
(
status
);
status
=
clSetKernelArg
(
kernel
,
7
,
sizeof
(
cl_mem
),
&
out
);
CL_CHECK_ERRORS
(
status
);
status
=
clEnqueueNDRangeKernel
(
this
->
cl_helper_
.
CLCommandQueue
(),
kernel
,
3
,
NULL
,
work_size
,
local_work_size
,
0
,
NULL
,
NULL
);
CL_CHECK_ERRORS
(
status
);
InstanceNorm
(
&
this
->
cl_helper_
,
param
);
}
template
class
InstanceNormKernel
<
GPU_CL
,
float
>;
...
...
mobile/src/operators/kernel/cl/instancenorm_relu_kernel.cpp
浏览文件 @
a96d013c
...
...
@@ -16,6 +16,7 @@ limitations under the License. */
#include "operators/kernel/instancenorm_relu_kernel.h"
#include <cmath>
#include "operators/kernel/cl/cl-kernel-func/instancenorm_func.h"
namespace
paddle_mobile
{
namespace
operators
{
...
...
@@ -23,7 +24,16 @@ namespace operators {
template
<
>
bool
InstanceNormReluKernel
<
GPU_CL
,
float
>::
Init
(
InstanceNormParam
<
GPU_CL
>
*
param
)
{
const
std
::
string
build_options
=
"-DRELU"
;
auto
&
dims
=
param
->
Out
()
->
dims
();
const
int
h
=
dims
[
2
];
std
::
string
build_options
=
"-DRELU"
;
if
(
h
==
128
)
{
build_options
+=
" -DLOCAL_MEM_128"
;
}
else
if
(
h
==
64
)
{
build_options
+=
" -DLOCAL_MEM_64"
;
}
else
if
(
h
>
256
)
{
PADDLE_MOBILE_THROW_EXCEPTION
(
"instance norm unsupported input height"
);
}
this
->
cl_helper_
.
AddKernel
(
"instancenorm"
,
"instancenorm_kernel.cl"
,
build_options
);
return
true
;
...
...
@@ -32,59 +42,7 @@ bool InstanceNormReluKernel<GPU_CL, float>::Init(
template
<
>
void
InstanceNormReluKernel
<
GPU_CL
,
float
>::
Compute
(
const
InstanceNormParam
<
GPU_CL
>
&
param
)
{
auto
kernel
=
this
->
cl_helper_
.
KernelAt
(
0
);
auto
&
dims
=
param
.
Out
()
->
dims
();
const
int
n
=
dims
[
0
];
const
int
c_group
=
(
dims
[
1
]
+
3
)
/
4
;
const
int
h
=
dims
[
2
];
const
int
w
=
dims
[
3
];
auto
epsilon
=
param
.
Epsilon
();
auto
input
=
param
.
InputX
()
->
GetCLImage
();
auto
out
=
param
.
Out
()
->
GetCLImage
();
DLOG
<<
"Epsilon: "
<<
epsilon
;
auto
local_work_size_info
=
this
->
cl_helper_
.
LocalWorkSizeInfo
();
DLOG
<<
local_work_size_info
.
max_work_group_size
;
DLOG
<<
local_work_size_info
.
max_work_item_size0
;
DLOG
<<
local_work_size_info
.
max_work_item_size1
;
DLOG
<<
local_work_size_info
.
max_work_item_size2
;
int
local_work_size1
=
std
::
min
(
static_cast
<
int
>
(
local_work_size_info
.
max_work_item_size1
),
std
::
min
(
256
,
w
));
int
local_work_size2
=
1
;
const
size_t
work_size
[
3
]
=
{(
size_t
)(
n
*
c_group
),
(
size_t
)
local_work_size1
,
(
size_t
)
local_work_size2
};
const
size_t
local_work_size
[
3
]
=
{(
size_t
)
1
,
(
size_t
)
local_work_size1
,
(
size_t
)
local_work_size2
};
DLOG
<<
"work_size"
<<
work_size
[
0
]
<<
" "
<<
work_size
[
1
]
<<
" "
<<
work_size
[
2
];
DLOG
<<
"local_work_size"
<<
local_work_size
[
0
]
<<
" "
<<
local_work_size
[
1
]
<<
" "
<<
local_work_size
[
2
];
cl_int
status
;
clSetKernelArg
(
kernel
,
0
,
sizeof
(
cl_int
),
&
w
);
CL_CHECK_ERRORS
(
status
);
clSetKernelArg
(
kernel
,
1
,
sizeof
(
cl_int
),
&
h
);
CL_CHECK_ERRORS
(
status
);
clSetKernelArg
(
kernel
,
2
,
sizeof
(
cl_int
),
&
c_group
);
CL_CHECK_ERRORS
(
status
);
clSetKernelArg
(
kernel
,
3
,
sizeof
(
cl_int
),
&
local_work_size1
);
CL_CHECK_ERRORS
(
status
);
clSetKernelArg
(
kernel
,
4
,
sizeof
(
cl_int
),
&
local_work_size2
);
CL_CHECK_ERRORS
(
status
);
clSetKernelArg
(
kernel
,
5
,
sizeof
(
cl_float
),
&
epsilon
);
CL_CHECK_ERRORS
(
status
);
clSetKernelArg
(
kernel
,
6
,
sizeof
(
cl_mem
),
&
input
);
CL_CHECK_ERRORS
(
status
);
clSetKernelArg
(
kernel
,
7
,
sizeof
(
cl_mem
),
&
out
);
CL_CHECK_ERRORS
(
status
);
clEnqueueNDRangeKernel
(
this
->
cl_helper_
.
CLCommandQueue
(),
kernel
,
3
,
NULL
,
work_size
,
local_work_size
,
0
,
NULL
,
NULL
);
InstanceNorm
(
&
this
->
cl_helper_
,
param
);
}
template
class
InstanceNormReluKernel
<
GPU_CL
,
float
>;
...
...
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