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
PaddleDetection
提交
2d9508f8
P
PaddleDetection
项目概览
PaddlePaddle
/
PaddleDetection
12 个月 前同步成功
通知
692
Star
11112
Fork
2696
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
184
列表
看板
标记
里程碑
合并请求
40
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
184
Issue
184
列表
看板
标记
里程碑
合并请求
40
合并请求
40
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
体验新版 GitCode,发现更多精彩内容 >>
未验证
提交
2d9508f8
编写于
8月 13, 2018
作者:
S
sneaxiy
提交者:
GitHub
8月 13, 2018
浏览文件
操作
浏览文件
下载
差异文件
Merge pull request #12554 from sneaxiy/refine_elementwise_add
Refine elementwise_add op
上级
6b45c5a1
5d698589
变更
2
隐藏空白更改
内联
并排
Showing
2 changed file
with
87 addition
and
8 deletion
+87
-8
paddle/fluid/operators/elementwise_add_op.cu
paddle/fluid/operators/elementwise_add_op.cu
+54
-0
paddle/fluid/operators/elementwise_add_op.h
paddle/fluid/operators/elementwise_add_op.h
+33
-8
未找到文件。
paddle/fluid/operators/elementwise_add_op.cu
浏览文件 @
2d9508f8
...
...
@@ -16,6 +16,60 @@ limitations under the License. */
#include "paddle/fluid/operators/elementwise_add_op.h"
#include "paddle/fluid/platform/float16.h"
namespace
paddle
{
namespace
operators
{
template
<
typename
T
>
__global__
void
ElementwiseAddCUDAKernel
(
const
T
*
x
,
const
T
*
y
,
T
*
z
,
int
n
,
int
post
,
int
size
)
{
int
idx_x
=
threadIdx
.
x
+
blockIdx
.
x
*
blockDim
.
x
;
if
(
idx_x
<
size
)
{
int
idx_y
=
idx_x
/
post
-
(
idx_x
/
(
n
*
post
))
*
n
;
z
[
idx_x
]
=
x
[
idx_x
]
+
y
[
idx_y
];
}
}
template
<
typename
T
>
class
ElementwiseAddKernel
<
platform
::
CUDADeviceContext
,
T
>
:
public
framework
::
OpKernel
<
T
>
{
public:
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
using
Tensor
=
framework
::
Tensor
;
const
auto
x
=
ctx
.
Input
<
Tensor
>
(
"X"
);
const
auto
y
=
ctx
.
Input
<
Tensor
>
(
"Y"
);
auto
z
=
ctx
.
Output
<
Tensor
>
(
"Out"
);
auto
*
z_data
=
z
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
auto
&
device
=
*
(
ctx
.
cuda_device_context
().
eigen_device
());
const
framework
::
DDim
&
x_dim
=
x
->
dims
();
framework
::
DDim
y_dim
=
y
->
dims
();
int
size
=
x
->
numel
();
if
(
x_dim
==
y_dim
)
{
auto
dim
=
framework
::
make_ddim
({
size
});
auto
z_eigen
=
framework
::
EigenTensor
<
T
,
1
>::
From
(
*
z
,
dim
);
auto
x_eigen
=
framework
::
EigenTensor
<
T
,
1
>::
From
(
*
x
,
dim
);
auto
y_eigen
=
framework
::
EigenTensor
<
T
,
1
>::
From
(
*
y
,
dim
);
z_eigen
.
device
(
device
)
=
x_eigen
+
y_eigen
;
}
else
{
int
axis
=
ctx
.
Attr
<
int
>
(
"axis"
);
axis
=
(
axis
==
-
1
?
x_dim
.
size
()
-
y_dim
.
size
()
:
axis
);
y_dim
=
trim_trailing_singular_dims
(
y_dim
);
axis
=
(
y_dim
.
size
()
==
0
)
?
x_dim
.
size
()
:
axis
;
int
pre
,
n
,
post
;
get_mid_dims
(
x_dim
,
y_dim
,
axis
,
&
pre
,
&
n
,
&
post
);
int
threads
=
512
;
int
grids
=
(
size
+
threads
-
1
)
/
threads
;
auto
stream
=
ctx
.
cuda_device_context
().
stream
();
ElementwiseAddCUDAKernel
<
T
><<<
grids
,
threads
,
0
,
stream
>>>
(
x
->
data
<
T
>
(),
y
->
data
<
T
>
(),
z_data
,
n
,
post
,
size
);
}
}
};
}
// namespace operators
}
// namespace paddle
namespace
ops
=
paddle
::
operators
;
namespace
plat
=
paddle
::
platform
;
...
...
paddle/fluid/operators/elementwise_add_op.h
浏览文件 @
2d9508f8
...
...
@@ -144,16 +144,41 @@ class ElementwiseAddGradKernel : public framework::OpKernel<T> {
auto
*
dout
=
ctx
.
Input
<
Tensor
>
(
framework
::
GradVarName
(
"Out"
));
auto
*
dx
=
ctx
.
Output
<
Tensor
>
(
framework
::
GradVarName
(
"X"
));
auto
*
dy
=
ctx
.
Output
<
Tensor
>
(
framework
::
GradVarName
(
"Y"
));
// skip out, x, y
auto
*
out
=
dout
;
auto
*
x
=
dout
,
*
y
=
dout
;
if
(
platform
::
is_cpu_place
(
ctx
.
GetPlace
())
&&
dx
!=
nullptr
&&
dy
!=
nullptr
&&
(
dx
->
dims
()
==
dy
->
dims
()))
{
elementwise_add_grad
<
DeviceContext
,
T
>
(
ctx
,
x
,
y
,
out
,
dout
,
dx
,
dy
);
if
(
dx
!=
nullptr
)
{
// In fact, we can just share memory, but it may cause a bug of memory
// optimizer
// dx->ShareDataWith(*dout);
framework
::
TensorCopy
(
*
dout
,
ctx
.
GetPlace
(),
ctx
.
template
device_context
<
DeviceContext
>(),
dx
);
}
if
(
dy
==
nullptr
)
return
;
const
framework
::
DDim
&
x_dim
=
dout
->
dims
();
framework
::
DDim
y_dim
=
dy
->
dims
();
if
(
x_dim
==
y_dim
)
{
// dy->ShareDataWith(*dout);
framework
::
TensorCopy
(
*
dout
,
ctx
.
GetPlace
(),
ctx
.
template
device_context
<
DeviceContext
>(),
dy
);
}
else
{
default_elementwise_add_grad
<
DeviceContext
,
T
>
(
ctx
,
x
,
y
,
out
,
dout
,
dx
,
dy
);
dy
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
// Perform reduction to dout to calculate dy
int
axis
=
ctx
.
Attr
<
int
>
(
"axis"
);
axis
=
(
axis
==
-
1
?
x_dim
.
size
()
-
y_dim
.
size
()
:
axis
);
y_dim
=
trim_trailing_singular_dims
(
y_dim
);
axis
=
(
y_dim
.
size
()
==
0
)
?
x_dim
.
size
()
:
axis
;
auto
&
device
=
*
(
ctx
.
template
device_context
<
DeviceContext
>().
eigen_device
());
int
pre
,
n
,
post
;
get_mid_dims
(
x_dim
,
y_dim
,
axis
,
&
pre
,
&
n
,
&
post
);
auto
eigen_dout
=
framework
::
EigenTensor
<
T
,
3
>::
From
(
*
dout
,
framework
::
make_ddim
({
pre
,
n
,
post
}));
auto
eigen_dy
=
framework
::
EigenTensor
<
T
,
1
>::
From
(
*
dy
,
framework
::
make_ddim
({
n
}));
eigen_dy
.
device
(
device
)
=
eigen_dout
.
sum
(
framework
::
EigenDim
<
2
>::
From
(
framework
::
make_ddim
({
0
,
2
})));
}
}
};
...
...
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