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
612ba41a
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 >>
提交
612ba41a
编写于
9月 13, 2018
作者:
T
tensor-tang
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
add simple lstm compute
上级
83035e9d
变更
5
隐藏空白更改
内联
并排
Showing
5 changed file
with
123 addition
and
2 deletion
+123
-2
paddle/fluid/operators/CMakeLists.txt
paddle/fluid/operators/CMakeLists.txt
+1
-0
paddle/fluid/operators/fusion_lstm_op.cc
paddle/fluid/operators/fusion_lstm_op.cc
+13
-2
paddle/fluid/operators/math/CMakeLists.txt
paddle/fluid/operators/math/CMakeLists.txt
+2
-0
paddle/fluid/operators/math/cpu_lstm_compute.cc
paddle/fluid/operators/math/cpu_lstm_compute.cc
+57
-0
paddle/fluid/operators/math/cpu_lstm_compute.h
paddle/fluid/operators/math/cpu_lstm_compute.h
+50
-0
未找到文件。
paddle/fluid/operators/CMakeLists.txt
浏览文件 @
612ba41a
...
...
@@ -296,6 +296,7 @@ op_library(flatten_op DEPS reshape_op)
op_library
(
sequence_pad_op DEPS sequence_padding
)
op_library
(
unstack_op DEPS stack_op
)
op_library
(
fake_quantize_op DEPS memory
)
op_library
(
fusion_lstm_op DEPS cpu_lstm_compute
)
if
(
WITH_GPU
)
op_library
(
conv_op DEPS vol2col depthwise_conv im2col
)
...
...
paddle/fluid/operators/fusion_lstm_op.cc
浏览文件 @
612ba41a
...
...
@@ -15,6 +15,7 @@ limitations under the License. */
#include "paddle/fluid/operators/fusion_lstm_op.h"
#include <string>
#include "paddle/fluid/operators/math/blas.h"
#include "paddle/fluid/operators/math/cpu_lstm_compute.h"
#include "paddle/fluid/operators/math/cpu_vec.h"
#include "paddle/fluid/operators/math/fc_compute.h"
#include "paddle/fluid/operators/math/sequence2batch.h"
...
...
@@ -269,7 +270,6 @@ class FuisonLSTMKernel : public framework::OpKernel<T> {
blas.GEMM(CblasNoTrans, CblasNoTrans, bs, D4, D, static_cast<T>(1), prev, D, \
wh_data, D4, static_cast<T>(1), out, D4)
// gates: W_ch, W_ih, W_fh, W_oh
#define GET_Ct(ct_1, gates, ct) \
/* C_t = C_t-1 * fgated + cand_gated * igated*/
\
act_cand(D, gates, gates); \
...
...
@@ -395,11 +395,22 @@ class FuisonLSTMKernel : public framework::OpKernel<T> {
}
}
}
else
{
// TODO(TJ): unly workaround, clean me
std
::
function
<
void
(
const
T
*
,
const
T
*
,
T
*
,
T
*
)
>
compute_ctht
;
if
(
platform
::
jit
::
MayIUse
(
platform
::
jit
::
avx
)
&&
act_gate_str
==
"sigmoid"
&&
act_cand_str
==
"tanh"
&&
act_cell_str
==
"tanh"
&&
D
==
8
)
{
compute_ctht
=
math
::
lstm_compute_ctht
<
T
>
;
}
else
{
compute_ctht
=
[
&
](
const
T
*
gates
,
const
T
*
ct_1
,
T
*
ct
,
T
*
ht
)
{
COMPUTE_CtHt
(
gates
,
ct_1
,
ct
,
ht
);
}
}
for
(
int
i
=
0
;
i
<
N
;
++
i
)
{
PROCESS_H0C0
for
(
int
step
=
tstart
;
step
<
seq_len
;
++
step
)
{
GEMM_WH_ADDON
(
1
,
prev_h_data
,
xx_data
);
COMPUTE_CtH
t
(
xx_data
,
prev_c_data
,
c_out_data
,
h_out_data
);
compute_cth
t
(
xx_data
,
prev_c_data
,
c_out_data
,
h_out_data
);
MOVE_ONE_STEP
;
}
}
...
...
paddle/fluid/operators/math/CMakeLists.txt
浏览文件 @
612ba41a
...
...
@@ -45,6 +45,8 @@ math_library(im2col)
if
(
NOT WIN32
)
# windows do not support avx functions yet.
math_library
(
gru_compute DEPS activation_functions math_function
)
math_library
(
lstm_compute DEPS activation_functions
)
# TODO(TJ): ugly workaround, clean me
cc_library
(
cpu_lstm_compute SRCS cpu_lstm_compute.cc DEPS activation_functions cblas
)
endif
(
NOT WIN32
)
cc_library
(
blas SRCS blas.cc DEPS cblas framework_proto device_context
)
...
...
paddle/fluid/operators/math/cpu_lstm_compute.cc
0 → 100644
浏览文件 @
612ba41a
/* 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 "paddle/fluid/operators/math/cpu_lstm_compute.h"
#ifdef __AVX__
#include <immintrin.h>
#endif
namespace
paddle
{
namespace
operators
{
namespace
math
{
#ifdef __AVX__
// TODO(TJ): ugly workaround, clean me
namespace
detail
{
namespace
forward
{
namespace
avx
{}
// namespace avx
}
// namespace forward
}
// namespace detail
template
<
>
void
lstm_compute_ctht
<
float
>
(
const
float
*
gates
,
const
float
*
ct_1
,
float
*
ct
,
float
*
ht
)
{
namespace
act
=
detail
::
forward
::
avx
;
// gates: W_ch, W_ih, W_fh, W_oh
__m256
c
,
i
,
f
,
o
;
c
=
_mm256_loadu_ps
(
gates
);
i
=
_mm256_loadu_ps
(
gates
+
8
);
f
=
_mm256_loadu_ps
(
gates
+
16
);
o
=
_mm256_loadu_ps
(
gates
+
24
);
/* C_t = C_t-1 * fgated + cand_gated * igated*/
c
=
_mm256_mul_ps
(
act
::
Tanh
(
c
),
act
::
Sigmoid
(
i
));
i
=
_mm256_loadu_ps
(
ct_1
);
f
=
_mm256_mul_ps
(
i
,
act
::
Sigmoid
(
f
));
f
=
_mm256_add_ps
(
c
,
f
);
_mm256_storeu_ps
(
ct
,
f
);
/* H_t = act_cell(C_t) * ogated */
o
=
_mm256_mul_ps
(
act
::
Tanh
(
f
),
act
::
Sigmoid
(
o
));
_mm256_storeu_ps
(
ht
,
o
);
}
#endif
}
// namespace math
}
// namespace operators
}
// namespace paddle
paddle/fluid/operators/math/cpu_lstm_compute.h
0 → 100644
浏览文件 @
612ba41a
/* 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. */
#pragma once
#include <string>
#include "paddle/fluid/operators/math/cpu_vec.h"
#include "paddle/fluid/platform/cpu_info.h"
namespace
paddle
{
namespace
operators
{
namespace
math
{
// TODO(TJ): ugly workaround, clean me
template
<
typename
T
>
void
lstm_compute_ctht
(
const
T
*
gates
,
const
T
*
ct_1
,
T
*
ct
,
T
*
ht
)
{
// gates: W_ch, W_ih, W_fh, W_oh
vec_sigmoid
<
T
,
platform
::
jit
::
avx
>
(
24
,
gates
+
8
,
gates
+
8
);
vec_tanh
<
T
,
platform
::
jit
::
avx
>
(
8
,
gates
,
gates
);
const
T
*
i
=
gates
+
8
,
*
f
=
gates
+
16
,
*
o
=
gates
+
24
;
for
(
int
d
=
0
;
d
<
8
;
++
d
)
{
// C_t = C_t-1 * fgated + cand_gated * igated
ct
[
d
]
=
ct_1
[
d
]
*
f
[
d
]
+
gates
[
d
]
*
i
[
d
];
// H_t = act_cell(C_t) * ogated
T
tmp
=
ct
[
d
]
*
2
;
tmp
=
static_cast
<
T
>
(
0
)
-
(
tmp
<
static_cast
<
T
>
(
SIGMOID_THRESHOLD_MIN
))
?
min
:
((
tmp
>
static_cast
<
T
>
(
SIGMOID_THRESHOLD_MAX
))
?
static_cast
<
T
>
(
SIGMOID_THRESHOLD_MAX
)
:
tmp
);
vec_exp
<
T
>
(
1
,
&
tmp
,
&
tmp
);
tmp
=
static_cast
<
T
>
(
2
)
/
(
static_cast
<
T
>
(
1
)
+
tmp
)
-
static_cast
<
T
>
(
1
);
ht
[
d
]
=
tmp
*
o
[
d
];
}
}
}
// namespace math
}
// 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