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
夏有凉风,冬有雪
Paddle
提交
faf827ba
P
Paddle
项目概览
夏有凉风,冬有雪
/
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
体验新版 GitCode,发现更多精彩内容 >>
提交
faf827ba
编写于
9月 11, 2017
作者:
T
Tao Luo
提交者:
GitHub
9月 11, 2017
浏览文件
操作
浏览文件
下载
差异文件
Merge pull request #4008 from tensor-tang/refine
Refine MKLDNNMatrix and MKLDNNLayer
上级
7bd51712
f40d5f58
变更
5
隐藏空白更改
内联
并排
Showing
5 changed file
with
60 addition
and
69 deletion
+60
-69
paddle/gserver/layers/MKLDNNFcLayer.cpp
paddle/gserver/layers/MKLDNNFcLayer.cpp
+16
-25
paddle/gserver/layers/MKLDNNFcLayer.h
paddle/gserver/layers/MKLDNNFcLayer.h
+2
-2
paddle/gserver/layers/MKLDNNLayer.h
paddle/gserver/layers/MKLDNNLayer.h
+10
-29
paddle/math/MKLDNNMatrix.cpp
paddle/math/MKLDNNMatrix.cpp
+4
-6
paddle/math/MKLDNNMatrix.h
paddle/math/MKLDNNMatrix.h
+28
-7
未找到文件。
paddle/gserver/layers/MKLDNNFcLayer.cpp
浏览文件 @
faf827ba
...
...
@@ -77,24 +77,6 @@ void MKLDNNFcLayer::convertWeightsToPaddle() {
wgtVal_
->
reorderDataTo
(
wgtVal_
,
dstFmt
,
targetDim
);
}
void
MKLDNNFcLayer
::
convertOutputToOtherDevice
()
{
copyOutputInfoToOtherDevice
();
// find other cpu device and reorder output to cpu device
int
cnt
=
0
;
for
(
size_t
i
=
0
;
i
<
outputOtherDevice_
.
size
();
i
++
)
{
if
(
outputOtherDevice_
[
i
].
deviceId
==
CPU_DEVICE
)
{
// fc cpu output value do not need convert
// just share point
outputOtherDevice_
[
i
].
value
=
output_
.
value
;
++
cnt
;
}
}
if
(
cnt
>
1
)
{
LOG
(
WARNING
)
<<
"should not have more than one CPU devie"
;
}
}
void
MKLDNNFcLayer
::
reshape
()
{
const
Argument
&
input
=
getInput
(
0
,
getPrev
(
0
)
->
getDeviceId
());
int
batchSize
=
input
.
getBatchSize
();
...
...
@@ -155,7 +137,10 @@ void MKLDNNFcLayer::resetFwd() {
// change original output value to mkldnn output value
output_
.
value
=
std
::
dynamic_pointer_cast
<
Matrix
>
(
outVal_
);
if
(
!
outputIsOnlyMKLDNN
())
{
convertOutputToOtherDevice
();
copyOutputInfoToOtherDevice
();
// fc cpu output value do not need create convert
// just share point
getOutput
(
CPU_DEVICE
).
value
->
setData
(
output_
.
value
->
getData
());
}
// create forward handle
...
...
@@ -235,13 +220,12 @@ void MKLDNNFcLayer::resetBwd() {
pipelineBwd_
.
push_back
(
*
bwdWgt_
);
/// backward data
device
=
inputIsOnlyMKLDNN
()
?
MKLDNN_DEVICE
:
CPU_DEVICE
;
const
MatrixPtr
&
in
=
getInputGrad
(
0
,
device
);
const
MatrixPtr
&
in
=
inputLayers_
[
0
]
->
getOutput
().
grad
;
if
(
in
==
nullptr
)
{
return
;
}
if
(
getInput
(
0
,
device
).
getAllCount
()
>
1
)
{
// TODO(TJ): use outputMaps_ ways when merge outgrad done
if
(
getInput
(
0
,
MKLDNN_DEVICE
).
getAllCount
()
>
1
)
{
// TODO(TJ): use outputMaps_ ways
to get the inGrad_
when merge outgrad done
}
else
{
inGrad_
=
MKLDNNMatrix
::
create
(
in
,
inVal_
->
getPrimitiveDesc
());
}
...
...
@@ -258,13 +242,21 @@ void MKLDNNFcLayer::resetBwd() {
pipelineBwd_
.
push_back
(
*
bwdData_
);
}
void
MKLDNNFcLayer
::
updateInputData
()
{
if
(
inputLayers_
[
0
]
->
getType
()
!=
"data"
)
{
return
;
}
real
*
iData
=
getInputValue
(
0
,
CPU_DEVICE
)
->
getData
();
inVal_
->
setData
(
iData
);
}
void
MKLDNNFcLayer
::
forward
(
PassType
passType
)
{
Layer
::
forward
(
passType
);
reshape
();
{
REGISTER_TIMER_INFO
(
"mkldnn_FwdTimer"
,
getName
().
c_str
());
syncInputValue
();
updateInputData
();
// just submit forward pipeline
stream_
->
submit
(
pipelineFwd_
);
...
...
@@ -286,7 +278,6 @@ void MKLDNNFcLayer::backward(const UpdateCallback& callback) {
REGISTER_TIMER_INFO
(
"mkldnn_bwdTimer"
,
getName
().
c_str
());
resetBwd
();
syncOutputGrad
();
// just sumbmit backward pipeline
stream_
->
submit
(
pipelineBwd_
);
}
...
...
paddle/gserver/layers/MKLDNNFcLayer.h
浏览文件 @
faf827ba
...
...
@@ -53,6 +53,8 @@ public:
void
backward
(
const
UpdateCallback
&
callback
)
override
;
void
updateInputData
()
override
;
protected:
/**
* reshape the input image sizes
...
...
@@ -72,8 +74,6 @@ protected:
* only would be called when needed
*/
void
resetBwd
();
void
convertOutputToOtherDevice
()
override
;
};
}
// namespace paddle
paddle/gserver/layers/MKLDNNLayer.h
浏览文件 @
faf827ba
...
...
@@ -114,10 +114,10 @@ public:
virtual
void
convertWeightsToPaddle
()
{}
/**
*
convert MKLDNN output to other devic
e.
*
only support CPU device yet
*
Update input value data when input layer is "data" typ
e.
*
Since the input value data address might be changed.
*/
virtual
void
convertOutputToOtherDevice
()
{}
virtual
void
updateInputData
()
{}
/**
* print info about sizes
...
...
@@ -155,6 +155,7 @@ protected:
* copy base info and do not copy data value
*/
void
copyOutputInfoToOtherDevice
()
{
int
cnt
=
0
;
for
(
size_t
i
=
0
;
i
<
outputOtherDevice_
.
size
();
i
++
)
{
outputOtherDevice_
[
i
].
setFrameHeight
(
output_
.
getFrameHeight
());
outputOtherDevice_
[
i
].
setFrameWidth
(
output_
.
getFrameWidth
());
...
...
@@ -163,6 +164,12 @@ protected:
outputOtherDevice_
[
i
].
subSequenceStartPositions
=
output_
.
subSequenceStartPositions
;
outputOtherDevice_
[
i
].
cpuSequenceDims
=
output_
.
cpuSequenceDims
;
if
(
outputOtherDevice_
[
i
].
deviceId
==
CPU_DEVICE
)
{
++
cnt
;
}
}
if
(
cnt
>
1
)
{
LOG
(
WARNING
)
<<
"should not have more than one CPU devie"
;
}
}
...
...
@@ -193,32 +200,6 @@ protected:
return
outputOtherDevice_
.
size
()
==
0
;
}
/**
* Sync input value data
*/
void
syncInputValue
()
{
if
(
inputIsOnlyMKLDNN
())
{
return
;
}
real
*
iData
=
getInputValue
(
0
,
CPU_DEVICE
)
->
getData
();
// update input data
// since it might be changed if this is after data layer
inVal_
->
updateData
(
iData
);
}
/**
* Sync output grad data
*/
void
syncOutputGrad
()
{
if
(
outputIsOnlyMKLDNN
())
{
return
;
}
// update diff
real
*
oDiff
=
getOutput
(
CPU_DEVICE
).
grad
->
getData
();
outGrad_
->
updateData
(
oDiff
);
}
/**
* Set deviceId of this layer.
*/
...
...
paddle/math/MKLDNNMatrix.cpp
浏览文件 @
faf827ba
...
...
@@ -33,14 +33,12 @@ MKLDNNMatrixPtr MKLDNNMatrix::create(MatrixPtr m, memory::primitive_desc pd) {
size_t
width
=
cnts
/
dims
[
0
];
m
=
Matrix
::
create
(
height
,
width
,
false
,
false
);
}
CHECK
(
m
)
<<
" Matrix should not be empty"
;
CpuMatrixPtr
cpuMatrix
=
std
::
dynamic_pointer_cast
<
CpuMatrix
>
(
m
);
CHECK
(
cpuMatrix
)
<<
"Only support create from CPU matrix yet"
;
CHECK_EQ
(
cnts
,
m
->
getElementCnt
())
<<
"Count size does not match"
;
return
std
::
make_shared
<
MKLDNNMatrix
>
(
m
->
getData
(),
m
->
getHeight
(),
m
->
getWidth
(),
pd
);
CHECK_EQ
(
cpuMatrix
->
getElementCnt
(),
cnts
)
<<
"Count size does not match"
;
return
std
::
make_shared
<
MKLDNNMatrix
>
(
cpuMatrix
,
pd
);
}
MKLDNNMatrixPtr
MKLDNNMatrix
::
create
(
MatrixPtr
m
,
...
...
@@ -138,7 +136,7 @@ void MKLDNNMatrix::downSpatial() {
mkldnn_primitive_create
(
&
result
,
pd
.
get
(),
nullptr
,
nullptr
),
"could not create a memory primitive"
);
reset
(
result
);
set_data_handle
(
getData
()
);
set_data_handle
(
data_
);
}
}
// namespace paddle
paddle/math/MKLDNNMatrix.h
浏览文件 @
faf827ba
...
...
@@ -30,11 +30,10 @@ typedef std::shared_ptr<MKLDNNMatrix> MKLDNNMatrixPtr;
*/
class
MKLDNNMatrix
:
public
CpuMatrix
,
public
mkldnn
::
memory
{
public:
MKLDNNMatrix
(
real
*
data
,
size_t
height
,
size_t
width
,
mkldnn
::
memory
::
primitive_desc
pd
)
:
CpuMatrix
(
data
,
height
,
width
,
false
),
mkldnn
::
memory
(
pd
,
data
)
{}
MKLDNNMatrix
(
CpuMatrixPtr
m
,
mkldnn
::
memory
::
primitive_desc
pd
)
:
CpuMatrix
(
m
->
getData
(),
m
->
getHeight
(),
m
->
getWidth
(),
false
),
mkldnn
::
memory
(
pd
,
m
->
getData
()),
m_
(
m
)
{}
~
MKLDNNMatrix
()
{}
...
...
@@ -81,11 +80,29 @@ public:
void
downSpatial
();
/**
*
Update
the memory data handle.
*
set
the memory data handle.
* Caution: This will not check the buffer size of the data,
* it should be coverd by user.
*/
void
updateData
(
void
*
data
)
{
set_data_handle
(
data
);
}
void
setData
(
real
*
data
)
{
set_data_handle
(
data
);
CpuMatrix
::
setData
(
data
);
m_
.
reset
();
}
/**
* override Matrix::getData
* check data before return
*/
real
*
getData
()
override
{
CHECK_EQ
((
void
*
)
data_
,
get_data_handle
());
return
data_
;
}
const
real
*
getData
()
const
override
{
CHECK_EQ
((
void
*
)
data_
,
get_data_handle
());
return
data_
;
}
/**
* Get primitive descriptor.
...
...
@@ -143,6 +160,10 @@ protected:
memory
::
format
srcFmt
,
memory
::
format
dstFmt
,
memory
::
dims
dm
);
private:
// save the CpuMatrixPtr in case the buffer released outside
CpuMatrixPtr
m_
;
};
}
// 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