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
Crayon鑫
Paddle
提交
8b15ac82
P
Paddle
项目概览
Crayon鑫
/
Paddle
与 Fork 源项目一致
Fork自
PaddlePaddle / Paddle
通知
1
Star
1
Fork
0
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
1
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
Paddle
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
1
Issue
1
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
8b15ac82
编写于
9月 01, 2017
作者:
L
Liu Yiqun
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Move the definition of hl_cpu_gru_forward and hl_cpu_gru_backward to function/GruFunctor.h.
上级
c54c7d91
变更
4
隐藏空白更改
内联
并排
Showing
4 changed file
with
181 addition
and
170 deletion
+181
-170
paddle/cuda/include/hl_cpu_gru.cuh
paddle/cuda/include/hl_cpu_gru.cuh
+0
-134
paddle/function/GruFunctor.h
paddle/function/GruFunctor.h
+160
-0
paddle/gserver/layers/GruCompute.cpp
paddle/gserver/layers/GruCompute.cpp
+17
-15
paddle/scripts/docker/build_android.sh
paddle/scripts/docker/build_android.sh
+4
-21
未找到文件。
paddle/cuda/include/hl_cpu_gru.cuh
浏览文件 @
8b15ac82
...
@@ -18,14 +18,6 @@ limitations under the License. */
...
@@ -18,14 +18,6 @@ limitations under the License. */
#ifndef __NVCC__
#ifndef __NVCC__
#include "paddle/math/MathFunctions.h"
// #ifndef PADDLE_TYPE_DOUBLE
// #define CBLAS_GEMM paddle::gemm<float>
// #else
// #define CBLAS_GEMM paddle::gemm<double>
// #endif
template
<
class
OpResetOutput
>
template
<
class
OpResetOutput
>
void
hl_naive_gru_forward_reset_output
(
OpResetOutput
opResetOutput
,
void
hl_naive_gru_forward_reset_output
(
OpResetOutput
opResetOutput
,
real
*
gateValue
,
real
*
gateValue
,
...
@@ -210,51 +202,6 @@ inline void forward_final_output(OpFinalOutput opFinalOutput,
...
@@ -210,51 +202,6 @@ inline void forward_final_output(OpFinalOutput opFinalOutput,
}
}
}
}
template
<
class
OpResetOutput
,
class
OpFinalOutput
>
void
hl_cpu_gru_forward
(
OpResetOutput
opResetOutput
,
OpFinalOutput
opFinalOutput
,
hl_gru_value
value
,
int
frameSize
,
int
batchSize
,
hl_activation_mode_t
active_node
,
hl_activation_mode_t
active_gate
)
{
if
(
value
.
prevOutValue
)
{
// CBLAS_GEMM(CblasNoTrans,
// CblasNoTrans,
// batchSize,
// 2 * frameSize,
// frameSize,
// 1,
// value.prevOutValue,
// frameSize,
// value.gateWeight,
// frameSize * 2,
// 1,
// value.gateValue,
// frameSize * 3);
}
forward_reset_output
(
opResetOutput
,
value
,
frameSize
,
batchSize
,
active_gate
);
if
(
value
.
prevOutValue
)
{
// CBLAS_GEMM(CblasNoTrans,
// CblasNoTrans,
// batchSize,
// frameSize,
// frameSize,
// 1,
// value.resetOutputValue,
// frameSize,
// value.stateWeight,
// frameSize,
// 1,
// value.gateValue + frameSize * 2,
// frameSize * 3);
}
forward_final_output
(
opFinalOutput
,
value
,
frameSize
,
batchSize
,
active_node
);
}
template
<
class
OpStateGrad
>
template
<
class
OpStateGrad
>
void
hl_naive_gru_backward_state_grad
(
OpStateGrad
opStateGrad
,
void
hl_naive_gru_backward_state_grad
(
OpStateGrad
opStateGrad
,
real
*
gateValue
,
real
*
gateValue
,
...
@@ -524,87 +471,6 @@ inline void backward_reset_grad(OpResetGrad opResetGrad,
...
@@ -524,87 +471,6 @@ inline void backward_reset_grad(OpResetGrad opResetGrad,
}
}
}
}
}
}
template
<
class
OpStateGrad
,
class
OpResetGrad
>
void
hl_cpu_gru_backward
(
OpStateGrad
opStateGrad
,
OpResetGrad
opResetGrad
,
hl_gru_value
value
,
hl_gru_grad
grad
,
int
frameSize
,
int
batchSize
,
hl_activation_mode_t
active_node
,
hl_activation_mode_t
active_gate
)
{
backward_state_grad
(
opStateGrad
,
value
,
grad
,
frameSize
,
batchSize
,
active_node
);
if
(
value
.
prevOutValue
&&
grad
.
prevOutGrad
)
{
// CBLAS_GEMM(CblasNoTrans,
// CblasTrans,
// batchSize,
// frameSize,
// frameSize,
// 1,
// grad.gateGrad + frameSize * 2,
// frameSize * 3,
// value.stateWeight,
// frameSize,
// 0,
// grad.resetOutputGrad,
// frameSize);
if
(
grad
.
stateWeightGrad
)
{
// CBLAS_GEMM(CblasTrans,
// CblasNoTrans,
// frameSize,
// frameSize,
// batchSize,
// 1,
// value.resetOutputValue,
// frameSize,
// grad.gateGrad + frameSize * 2,
// frameSize * 3,
// 1,
// grad.stateWeightGrad,
// frameSize);
}
}
backward_reset_grad
(
opResetGrad
,
value
,
grad
,
frameSize
,
batchSize
,
active_gate
);
if
(
grad
.
prevOutGrad
&&
value
.
prevOutValue
)
{
// CBLAS_GEMM(CblasNoTrans,
// CblasTrans,
// batchSize,
// frameSize,
// frameSize * 2,
// 1,
// grad.gateGrad,
// frameSize * 3,
// value.gateWeight,
// frameSize * 2,
// 1,
// grad.prevOutGrad,
// frameSize);
if
(
grad
.
gateWeightGrad
)
{
// CBLAS_GEMM(CblasTrans,
// CblasNoTrans,
// frameSize,
// frameSize * 2,
// batchSize,
// 1,
// value.prevOutValue,
// frameSize,
// grad.gateGrad,
// frameSize * 3,
// 1,
// grad.gateWeightGrad,
// frameSize * 2);
}
}
}
#endif
#endif
#endif // HL_CPU_GRU_CUH_
#endif // HL_CPU_GRU_CUH_
paddle/function/GruFunctor.h
0 → 100644
浏览文件 @
8b15ac82
/* 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. */
#pragma once
#include "GemmFunctor.h"
#include "GruFunctor.h"
#include "hl_cpu_gru.cuh"
namespace
paddle
{
template
<
DeviceType
Device
,
class
T
>
struct
GruFunctor
{
template
<
class
OpResetOutput
,
class
OpFinalOutput
>
static
void
compute
(
OpResetOutput
opResetOutput
,
OpFinalOutput
opFinalOutput
,
hl_gru_value
value
,
int
frameSize
,
int
batchSize
,
hl_activation_mode_t
active_node
,
hl_activation_mode_t
active_gate
)
{
#ifndef __NVCC__
if
(
value
.
prevOutValue
)
{
BlasGemm
<
Device
,
T
>::
compute
(
false
,
false
,
batchSize
,
2
*
frameSize
,
frameSize
,
1
,
value
.
prevOutValue
,
frameSize
,
value
.
gateWeight
,
frameSize
*
2
,
1
,
value
.
gateValue
,
frameSize
*
3
);
}
forward_reset_output
(
opResetOutput
,
value
,
frameSize
,
batchSize
,
active_gate
);
if
(
value
.
prevOutValue
)
{
BlasGemm
<
Device
,
T
>::
compute
(
false
,
false
,
batchSize
,
frameSize
,
frameSize
,
1
,
value
.
resetOutputValue
,
frameSize
,
value
.
stateWeight
,
frameSize
,
1
,
value
.
gateValue
+
frameSize
*
2
,
frameSize
*
3
);
}
forward_final_output
(
opFinalOutput
,
value
,
frameSize
,
batchSize
,
active_node
);
#endif
}
};
template
<
DeviceType
Device
,
class
T
>
struct
GruGradFunctor
{
template
<
class
OpStateGrad
,
class
OpResetGrad
>
static
void
compute
(
OpStateGrad
opStateGrad
,
OpResetGrad
opResetGrad
,
hl_gru_value
value
,
hl_gru_grad
grad
,
int
frameSize
,
int
batchSize
,
hl_activation_mode_t
active_node
,
hl_activation_mode_t
active_gate
)
{
#ifndef __NVCC__
backward_state_grad
(
opStateGrad
,
value
,
grad
,
frameSize
,
batchSize
,
active_node
);
if
(
value
.
prevOutValue
&&
grad
.
prevOutGrad
)
{
BlasGemm
<
Device
,
T
>::
compute
(
false
,
true
,
batchSize
,
frameSize
,
frameSize
,
1
,
grad
.
gateGrad
+
frameSize
*
2
,
frameSize
*
3
,
value
.
stateWeight
,
frameSize
,
0
,
grad
.
resetOutputGrad
,
frameSize
);
if
(
grad
.
stateWeightGrad
)
{
BlasGemm
<
Device
,
T
>::
compute
(
true
,
false
,
frameSize
,
frameSize
,
batchSize
,
1
,
value
.
resetOutputValue
,
frameSize
,
grad
.
gateGrad
+
frameSize
*
2
,
frameSize
*
3
,
1
,
grad
.
stateWeightGrad
,
frameSize
);
}
}
backward_reset_grad
(
opResetGrad
,
value
,
grad
,
frameSize
,
batchSize
,
active_gate
);
if
(
grad
.
prevOutGrad
&&
value
.
prevOutValue
)
{
BlasGemm
<
Device
,
T
>::
compute
(
false
,
true
,
batchSize
,
frameSize
,
frameSize
*
2
,
1
,
grad
.
gateGrad
,
frameSize
*
3
,
value
.
gateWeight
,
frameSize
*
2
,
1
,
grad
.
prevOutGrad
,
frameSize
);
if
(
grad
.
gateWeightGrad
)
{
BlasGemm
<
Device
,
T
>::
compute
(
true
,
false
,
frameSize
,
frameSize
*
2
,
batchSize
,
1
,
value
.
prevOutValue
,
frameSize
,
grad
.
gateGrad
,
frameSize
*
3
,
1
,
grad
.
gateWeightGrad
,
frameSize
*
2
);
}
}
#endif
}
};
}
// namespace paddle
paddle/gserver/layers/GruCompute.cpp
浏览文件 @
8b15ac82
...
@@ -14,6 +14,7 @@ limitations under the License. */
...
@@ -14,6 +14,7 @@ limitations under the License. */
#include "GruCompute.h"
#include "GruCompute.h"
#include "hl_recurrent_apply.cuh"
#include "hl_recurrent_apply.cuh"
#include "paddle/function/GruFunctor.h"
#include "paddle/utils/Util.h"
#include "paddle/utils/Util.h"
namespace
paddle
{
namespace
paddle
{
...
@@ -25,13 +26,13 @@ void GruCompute::init(LayerConfig &config) {
...
@@ -25,13 +26,13 @@ void GruCompute::init(LayerConfig &config) {
template
<
>
template
<
>
void
GruCompute
::
forward
<
0
>
(
hl_gru_value
value
,
int
frameSize
,
int
batchSize
)
{
void
GruCompute
::
forward
<
0
>
(
hl_gru_value
value
,
int
frameSize
,
int
batchSize
)
{
hl_cpu_gru_forward
(
hppl
::
forward
::
gru_resetOutput
(),
GruFunctor
<
DEVICE_TYPE_CPU
,
real
>::
compute
(
hppl
::
forward
::
gru_resetOutput
(),
hppl
::
forward
::
gru_finalOutput
(),
hppl
::
forward
::
gru_finalOutput
(),
value
,
value
,
frameSize
,
frameSize
,
batchSize
,
batchSize
,
activeNode_
,
activeNode_
,
activeGate_
);
activeGate_
);
}
}
template
<
>
template
<
>
...
@@ -39,14 +40,15 @@ void GruCompute::backward<0>(hl_gru_value value,
...
@@ -39,14 +40,15 @@ void GruCompute::backward<0>(hl_gru_value value,
hl_gru_grad
grad
,
hl_gru_grad
grad
,
int
frameSize
,
int
frameSize
,
int
batchSize
)
{
int
batchSize
)
{
hl_cpu_gru_backward
(
hppl
::
backward
::
gru_stateGrad
(),
GruGradFunctor
<
DEVICE_TYPE_CPU
,
real
>::
compute
(
hppl
::
backward
::
gru_resetGrad
(),
hppl
::
backward
::
gru_stateGrad
(),
value
,
hppl
::
backward
::
gru_resetGrad
(),
grad
,
value
,
frameSize
,
grad
,
batchSize
,
frameSize
,
activeNode_
,
batchSize
,
activeGate_
);
activeNode_
,
activeGate_
);
}
}
}
// namespace paddle
}
// namespace paddle
paddle/scripts/docker/build_android.sh
浏览文件 @
8b15ac82
...
@@ -2,25 +2,8 @@
...
@@ -2,25 +2,8 @@
set
-xe
set
-xe
COMPILER
=
gcc
BUILD_ROOT
=
/paddle/build_android
USE_EIGEN
=
ON
DEST_ROOT
=
/paddle/install
if
[
$COMPILER
==
clang
]
;
then
SUFFIX
=
_clang
C_COMPILER
=
clang
CXX_COMPILER
=
clang++
else
SUFFIX
=
_gcc
C_COMPILER
=
gcc
CXX_COMPILER
=
g++
fi
if
[
$USE_EIGEN
==
ON
]
;
then
SUFFIX
=
${
SUFFIX
}
_eigen
else
SUFFIX
=
${
SUFFIX
}
_openblas
fi
BUILD_ROOT
=
/paddle/build_android
$SUFFIX
DEST_ROOT
=
/paddle/install
$SUFFIX
rm
-rf
$BUILD_ROOT
2>/dev/null
||
true
rm
-rf
$BUILD_ROOT
2>/dev/null
||
true
mkdir
-p
$BUILD_ROOT
mkdir
-p
$BUILD_ROOT
...
@@ -41,7 +24,7 @@ if [ $ANDROID_ABI == "armeabi-v7a" ]; then
...
@@ -41,7 +24,7 @@ if [ $ANDROID_ABI == "armeabi-v7a" ]; then
-DCMAKE_INSTALL_PREFIX
=
$DEST_ROOT
\
-DCMAKE_INSTALL_PREFIX
=
$DEST_ROOT
\
-DTHIRD_PARTY_PATH
=
$THIRD_PARTY_PATH
\
-DTHIRD_PARTY_PATH
=
$THIRD_PARTY_PATH
\
-DCMAKE_BUILD_TYPE
=
Release
\
-DCMAKE_BUILD_TYPE
=
Release
\
-DUSE_EIGEN_FOR_BLAS
=
${
USE_EIGEN
}
\
-DUSE_EIGEN_FOR_BLAS
=
ON
\
-DWITH_C_API
=
ON
\
-DWITH_C_API
=
ON
\
-DWITH_SWIG_PY
=
OFF
\
-DWITH_SWIG_PY
=
OFF
\
-DWITH_STYLE_CHECK
=
OFF
\
-DWITH_STYLE_CHECK
=
OFF
\
...
@@ -58,7 +41,7 @@ elif [ $ANDROID_ABI == "arm64-v8a" ]; then
...
@@ -58,7 +41,7 @@ elif [ $ANDROID_ABI == "arm64-v8a" ]; then
-DCMAKE_INSTALL_PREFIX
=
$DEST_ROOT
\
-DCMAKE_INSTALL_PREFIX
=
$DEST_ROOT
\
-DTHIRD_PARTY_PATH
=
$THIRD_PARTY_PATH
\
-DTHIRD_PARTY_PATH
=
$THIRD_PARTY_PATH
\
-DCMAKE_BUILD_TYPE
=
Release
\
-DCMAKE_BUILD_TYPE
=
Release
\
-DUSE_EIGEN_FOR_BLAS
=
${
USE_EIGEN
}
\
-DUSE_EIGEN_FOR_BLAS
=
OFF
\
-DWITH_C_API
=
ON
\
-DWITH_C_API
=
ON
\
-DWITH_SWIG_PY
=
OFF
\
-DWITH_SWIG_PY
=
OFF
\
..
..
...
...
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