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
s920243400
PaddleDetection
提交
ad3b3d9d
P
PaddleDetection
项目概览
s920243400
/
PaddleDetection
与 Fork 源项目一致
Fork自
PaddlePaddle / PaddleDetection
通知
2
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
PaddleDetection
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
提交
ad3b3d9d
编写于
3月 21, 2018
作者:
W
wangyang59
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
ported old paddle gpu bilinear_interp
上级
67ce5864
变更
2
显示空白变更内容
内联
并排
Showing
2 changed file
with
121 addition
and
8 deletion
+121
-8
paddle/fluid/operators/bilinear_interp_op.cu
paddle/fluid/operators/bilinear_interp_op.cu
+16
-8
paddle/fluid/operators/bilinear_interp_op.cu.h
paddle/fluid/operators/bilinear_interp_op.cu.h
+105
-0
未找到文件。
paddle/fluid/operators/bilinear_interp_op.cu
浏览文件 @
ad3b3d9d
...
@@ -9,7 +9,7 @@
...
@@ -9,7 +9,7 @@
See the License for the specific language governing permissions and
See the License for the specific language governing permissions and
limitations under the License. */
limitations under the License. */
#include "
hl_cnn
.h"
#include "
paddle/fluid/operators/bilinear_interp_op.cu
.h"
#include "paddle/fluid/operators/bilinear_interp_op.h"
#include "paddle/fluid/operators/bilinear_interp_op.h"
namespace
paddle
{
namespace
paddle
{
...
@@ -44,9 +44,13 @@ class BilinearInterpOpCUDAKernel : public framework::OpKernel<T> {
...
@@ -44,9 +44,13 @@ class BilinearInterpOpCUDAKernel : public framework::OpKernel<T> {
if
(
in_h
==
out_h
&&
in_w
==
out_w
)
{
if
(
in_h
==
out_h
&&
in_w
==
out_w
)
{
memcpy
(
output
,
input
,
input_t
->
numel
()
*
sizeof
(
T
));
memcpy
(
output
,
input
,
input_t
->
numel
()
*
sizeof
(
T
));
}
else
{
}
else
{
hl_bilinear_forward
(
input
,
in_h
,
in_w
,
batch_size
,
in_chw
,
output
,
out_h
,
int
threadNum
=
batch_size
*
out_chw
;
out_w
,
batch_size
,
out_chw
,
channels
,
ratio_h
,
int
blocks
=
(
threadNum
+
1024
-
1
)
/
1024
;
ratio_w
);
KeBilinearInterpFw
<
T
><<<
blocks
,
1024
,
0
,
ctx
.
cuda_device_context
().
stream
()
>>>
(
input
,
in_h
,
in_w
,
batch_size
,
in_chw
,
output
,
out_h
,
out_w
,
batch_size
,
out_chw
,
channels
,
ratio_h
,
ratio_w
);
}
}
}
}
};
};
...
@@ -78,9 +82,13 @@ class BilinearInterpGradOpCUDAKernel : public framework::OpKernel<T> {
...
@@ -78,9 +82,13 @@ class BilinearInterpGradOpCUDAKernel : public framework::OpKernel<T> {
if
(
in_h
==
out_h
&&
in_w
==
out_w
)
{
if
(
in_h
==
out_h
&&
in_w
==
out_w
)
{
memcpy
(
d_input
,
d_output
,
d_input_t
->
numel
()
*
sizeof
(
T
));
memcpy
(
d_input
,
d_output
,
d_input_t
->
numel
()
*
sizeof
(
T
));
}
else
{
}
else
{
hl_bilinear_backward
(
d_input
,
in_h
,
in_w
,
batch_size
,
in_chw
,
d_output
,
int
threadNum
=
batch_size
*
out_chw
;
out_h
,
out_w
,
batch_size
,
out_chw
,
channels
,
ratio_h
,
int
blocks
=
(
threadNum
+
1024
-
1
)
/
1024
;
ratio_w
);
KeBilinearInterpBw
<
T
><<<
blocks
,
1024
,
0
,
ctx
.
cuda_device_context
().
stream
()
>>>
(
d_input
,
in_h
,
in_w
,
batch_size
,
in_chw
,
d_output
,
out_h
,
out_w
,
batch_size
,
out_chw
,
channels
,
ratio_h
,
ratio_w
);
}
}
}
}
};
};
...
...
paddle/fluid/operators/bilinear_interp_op.cu.h
0 → 100644
浏览文件 @
ad3b3d9d
/* Copyright (c) 2016 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. */
#pragma once
#include "paddle/fluid/framework/tensor.h"
#include "paddle/fluid/platform/cuda_helper.h"
#include "paddle/fluid/platform/place.h"
namespace
paddle
{
namespace
operators
{
using
Tensor
=
framework
::
Tensor
;
template
<
typename
T
>
__global__
void
KeBilinearInterpFw
(
const
T
*
in
,
const
size_t
inImgH
,
const
size_t
inImgW
,
const
size_t
inputH
,
const
size_t
inputW
,
T
*
out
,
const
size_t
outImgH
,
const
size_t
outImgW
,
const
size_t
outputH
,
const
size_t
outputW
,
const
size_t
numChannels
,
const
T
ratioH
,
const
T
ratioW
)
{
int
nthreads
=
outputH
*
outputW
;
int
tid
=
blockIdx
.
x
*
blockDim
.
x
+
threadIdx
.
x
;
if
(
tid
<
nthreads
)
{
int
outIdH
=
tid
/
outputW
;
int
outIdW
=
tid
%
outputW
;
int
inImgSize
=
inputW
/
numChannels
;
int
outImgSize
=
outputW
/
numChannels
;
int
channelId
=
outIdW
/
outImgSize
;
int
outImgIdy
=
(
outIdW
%
outImgSize
)
/
outImgW
;
int
inImgIdy
=
ratioH
*
outImgIdy
;
int
hId
=
(
inImgIdy
<
inImgH
-
1
)
?
1
:
0
;
T
h1lambda
=
ratioH
*
outImgIdy
-
inImgIdy
;
T
h2lambda
=
1.
f
-
h1lambda
;
int
outImgIdx
=
tid
%
outImgW
;
int
inImgIdx
=
ratioW
*
outImgIdx
;
int
wId
=
(
inImgIdx
<
inImgW
-
1
)
?
1
:
0
;
T
w1lambda
=
ratioW
*
outImgIdx
-
inImgIdx
;
T
w2lambda
=
1.
f
-
w1lambda
;
const
T
*
inPos
=
&
in
[
outIdH
*
inputW
+
channelId
*
inImgSize
+
inImgIdy
*
inImgW
+
inImgIdx
];
// bilinear interpolation
out
[
outIdH
*
outputW
+
outIdW
]
=
h2lambda
*
(
w2lambda
*
inPos
[
0
]
+
w1lambda
*
inPos
[
wId
])
+
h1lambda
*
(
w2lambda
*
inPos
[
hId
*
inImgW
]
+
w1lambda
*
inPos
[
hId
*
inImgW
+
wId
]);
}
}
template
<
typename
T
>
__global__
void
KeBilinearInterpBw
(
T
*
in
,
const
size_t
inImgH
,
const
size_t
inImgW
,
const
size_t
inputH
,
const
size_t
inputW
,
const
T
*
out
,
const
size_t
outImgH
,
const
size_t
outImgW
,
const
size_t
outputH
,
const
size_t
outputW
,
const
size_t
numChannels
,
const
T
ratioH
,
const
T
ratioW
)
{
int
nthreads
=
outputH
*
outputW
;
int
tid
=
blockIdx
.
x
*
blockDim
.
x
+
threadIdx
.
x
;
if
(
tid
<
nthreads
)
{
int
outIdH
=
tid
/
outputW
;
int
outIdW
=
tid
%
outputW
;
int
inImgSize
=
inputW
/
numChannels
;
int
outImgSize
=
outputW
/
numChannels
;
int
channelId
=
outIdW
/
outImgSize
;
int
outImgIdy
=
(
outIdW
%
outImgSize
)
/
outImgW
;
int
inImgIdy
=
ratioH
*
outImgIdy
;
int
hId
=
(
inImgIdy
<
inImgH
-
1
)
?
1
:
0
;
T
h1lambda
=
ratioH
*
outImgIdy
-
inImgIdy
;
T
h2lambda
=
1.
f
-
h1lambda
;
int
outImgIdx
=
tid
%
outImgW
;
int
inImgIdx
=
ratioW
*
outImgIdx
;
int
wId
=
(
inImgIdx
<
inImgW
-
1
)
?
1
:
0
;
T
w1lambda
=
ratioW
*
outImgIdx
-
inImgIdx
;
T
w2lambda
=
1.
f
-
w1lambda
;
T
*
inPos
=
&
in
[
outIdH
*
inputW
+
channelId
*
inImgSize
+
inImgIdy
*
inImgW
+
inImgIdx
];
const
T
*
outPos
=
&
out
[
outIdH
*
outputW
+
outIdW
];
atomicAdd
(
&
inPos
[
0
],
h2lambda
*
w2lambda
*
outPos
[
0
]);
atomicAdd
(
&
inPos
[
wId
],
h2lambda
*
w1lambda
*
outPos
[
0
]);
atomicAdd
(
&
inPos
[
hId
*
inImgW
],
h1lambda
*
w2lambda
*
outPos
[
0
]);
atomicAdd
(
&
inPos
[
hId
*
inImgW
+
wId
],
h1lambda
*
w1lambda
*
outPos
[
0
]);
}
}
}
// namespace operators
}
// namespace paddle
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