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
2b10d322
P
PaddleDetection
项目概览
PaddlePaddle
/
PaddleDetection
大约 1 年 前同步成功
通知
695
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
提交
2b10d322
编写于
9月 20, 2017
作者:
Z
zchen0211
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
lstm kernels
上级
6dad76c3
变更
1
隐藏空白更改
内联
并排
Showing
1 changed file
with
173 addition
and
0 deletion
+173
-0
paddle/operators/lstm_unit_op.cu
paddle/operators/lstm_unit_op.cu
+173
-0
未找到文件。
paddle/operators/lstm_unit_op.cu
0 → 100644
浏览文件 @
2b10d322
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
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 "paddle/framework/op_registry.h"
#include "paddle/operators/cross_entropy_op.h"
#include "paddle/platform/assert.h"
#include "paddle/platform/hostdevice.h"
namespace
paddle
{
namespace
operators
{
#define CUDA_1D_KERNEL_LOOP(i, n) \
for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < (n); \
i += blockDim.x * gridDim.x)
template
<
typename
Dtype
>
__device__
Dtype
cuda_sigmoid
(
const
Dtype
x
)
{
return
Dtype
(
1
)
/
(
Dtype
(
1
)
+
exp
(
-
x
));
}
template
<
typename
Dtype
>
__device__
Dtype
cuda_tanh
(
const
Dtype
x
)
{
return
Dtype
(
1
-
exp
(
-
2.
*
x
))
/
(
Dtype
(
1
)
+
exp
(
-
2.
*
x
));
}
template
<
typename
T
>
__global__
void
LSTMUnitKernel
(
const
int
nthreads
,
const
int
dim
,
const
int
t
,
const
T
*
C_prev
,
const
T
*
X
,
T
*
C
,
T
*
H
,
const
T
forget_bias
)
{
CUDA_1D_KERNEL_LOOP
(
index
,
nthreads
)
{
const
int
n
=
index
/
dim
;
const
int
d
=
index
%
dim
;
const
T
*
X_offset
=
X
+
4
*
dim
*
n
;
const
T
i
=
cuda_sigmoid
(
X_offset
[
d
]);
const
T
f
=
cuda_sigmoid
(
X_offset
[
1
*
dim
+
d
]
+
forget_bias
);
const
T
o
=
cuda_sigmoid
(
X_offset
[
2
*
dim
+
d
]);
const
T
g
=
cuda_tanh
(
X_offset
[
3
*
dim
+
d
]);
const
T
c_prev
=
C_prev
[
index
];
const
T
c
=
f
*
c_prev
+
i
*
g
;
C
[
index
]
=
c
;
const
T
tanh_c
=
cuda_tanh
(
c
);
H
[
index
]
=
o
*
tanh_c
;
}
}
template
<
typename
T
>
__global__
void
LSTMUnitGradientKernel
(
const
int
nthreads
,
const
int
dim
,
const
T
*
C_prev
,
const
T
*
X
,
const
T
*
C
,
const
T
*
H
,
const
T
*
C_diff
,
const
T
*
H_diff
,
T
*
C_prev_diff
,
T
*
X_diff
,
const
T
forget_bias
)
{
CUDA_1D_KERNEL_LOOP
(
index
,
nthreads
)
{
const
int
n
=
index
/
dim
;
const
int
d
=
index
%
dim
;
const
T
*
X_offset
=
X
+
4
*
dim
*
n
;
T
*
c_prev_diff
=
C_prev_diff
+
index
;
T
*
X_diff_offset
=
X_diff
+
4
*
dim
*
n
;
T
*
i_diff
=
X_diff_offset
+
d
;
T
*
f_diff
=
X_diff_offset
+
1
*
dim
+
d
;
T
*
o_diff
=
X_diff_offset
+
2
*
dim
+
d
;
T
*
g_diff
=
X_diff_offset
+
3
*
dim
+
d
;
const
T
i
=
cuda_sigmoid
(
X_offset
[
d
]);
const
T
f
=
cuda_sigmoid
(
X_offset
[
1
*
dim
+
d
]
+
forget_bias
);
const
T
o
=
cuda_sigmoid
(
X_offset
[
2
*
dim
+
d
]);
const
T
g
=
cuda_tanh
(
X_offset
[
3
*
dim
+
d
]);
const
T
c_prev
=
C_prev
[
index
];
const
T
c
=
C
[
index
];
const
T
tanh_c
=
cuda_tanh
(
c
);
const
T
c_term_diff
=
C_diff
[
index
]
+
H_diff
[
index
]
*
o
*
(
1
-
tanh_c
*
tanh_c
);
*
c_prev_diff
=
c_term_diff
*
f
;
*
i_diff
=
c_term_diff
*
g
*
i
*
(
1
-
i
);
*
f_diff
=
c_term_diff
*
c_prev
*
f
*
(
1
-
f
);
*
o_diff
=
H_diff
[
index
]
*
tanh_c
*
o
*
(
1
-
o
);
*
g_diff
=
c_term_diff
*
i
*
(
1
-
g
*
g
);
}
}
template
<
typename
T
,
typename
AttrType
=
T
>
class
LstmUnitOpCUDAKernel
:
public
framework
::
OpKernel
{
public:
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
PADDLE_ENFORCE
(
platform
::
is_gpu_place
(
ctx
.
GetPlace
()),
"It must use GPUPlace."
);
auto
*
x_tensor
=
ctx
.
Input
<
framework
::
Tensor
>
(
"X"
);
auto
*
c_prev_tensor
=
ctx
.
Input
<
framework
::
Tensor
>
(
"C_prev"
);
auto
*
c_tensor
=
ctx
.
Output
<
framework
::
Tensor
>
(
"C"
);
auto
*
h_tensor
=
ctx
.
Output
<
framework
::
Tensor
>
(
"H"
);
auto
forget_bias
=
static_cast
<
T
>
(
ctx
.
Attr
<
AttrType
>
(
"forget_bias"
));
int
b_size
=
c_tensor
->
dims
()[
0
];
int
D
=
c_tensor
->
dims
()[
1
];
const
T
*
X
=
x_tensor
->
data
<
T
>
();
const
T
*
C_prev
=
c_prev_tensor
->
data
<
T
>
();
T
*
C
=
c_tensor
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
T
*
H
=
h_tensor
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
int
block
=
512
;
int
n
=
b_size
*
D
;
int
grid
=
(
n
+
block
-
1
)
/
block
;
LSTMUnitKernel
<
T
><<<
grid
,
block
>>>
(
n
,
D
,
C_prev
,
X
,
C
,
H
,
forget_bias
);
}
};
template
<
typename
T
,
typename
AttrType
=
T
>
class
LstmUnitGradOpCUDAKernel
:
public
framework
::
OpKernel
{
public:
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
PADDLE_ENFORCE
(
platform
::
is_gpu_place
(
ctx
.
GetPlace
()),
"It must use GPUPlace."
);
auto
x_tensor
=
ctx
.
Input
<
Tensor
>
(
"X"
);
auto
c_prev_tensor
=
ctx
.
Input
<
Tensor
>
(
"C_prev"
);
auto
c_tensor
=
ctx
.
Input
<
Tensor
>
(
"C"
);
auto
h_tensor
=
ctx
.
Input
<
Tensor
>
(
"H"
);
auto
hdiff_tensor
=
ctx
.
Input
<
Tensor
>
(
framework
::
GradVarName
(
"H"
));
auto
cdiff_tensor
=
ctx
.
Input
<
Tensor
>
(
framework
::
GradVarName
(
"C"
));
auto
xdiff_tensor
=
ctx
.
Output
<
Tensor
>
(
framework
::
GradVarName
(
"X"
));
auto
c_prev_diff_tensor
=
ctx
.
Output
<
Tensor
>
(
framework
::
GradVarName
(
"C_prev"
));
auto
*
X
=
x_tensor
->
data
<
T
>
();
auto
*
C_prev
=
c_prev_tensor
->
data
<
T
>
();
auto
*
C
=
c_tensor
->
data
<
T
>
();
auto
*
H
=
h_tensor
->
data
<
T
>
();
auto
*
H_diff
=
hdiff_tensor
->
data
<
T
>
();
auto
*
C_diff
=
cdiff_tensor
->
data
<
T
>
();
auto
*
C_prev_diff
=
c_prev_diff_tensor
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
auto
*
X_diff
=
xdiff_tensor
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
int
N
=
c_tensor
->
dims
()[
0
];
int
D
=
c_tensor
->
dims
()[
1
];
auto
forget_bias
=
static_cast
<
T
>
(
ctx
.
Attr
<
AttrType
>
(
"forget_bias"
));
int
block
=
512
;
int
n
=
N
*
D
;
int
grid
=
(
n
+
block
-
1
)
/
block
;
LSTMUnitGradientKernel
<
T
><<<
N
*
D
,
block
>>>
(
n
,
D
,
C_prev
,
X
,
C
,
H
,
C_diff
,
H_diff
,
C_prev_diff
,
X_diff
,
T
forget_bias
)
}
};
}
// namespace operators
}
// namespace paddle
namespace
ops
=
paddle
::
operators
;
REGISTER_OP_GPU_KERNEL
(
lstm_unit
,
ops
::
LstmUnitOpCUDAKernel
<
float
>
);
REGISTER_OP_GPU_KERNEL
(
lstm_unit_grad
,
ops
::
LstmUnitGradOpCUDAKernel
<
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