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
Paddle
提交
1199aa68
P
Paddle
项目概览
PaddlePaddle
/
Paddle
大约 1 年 前同步成功
通知
2298
Star
20931
Fork
5422
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
1423
列表
看板
标记
里程碑
合并请求
543
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
1,423
Issue
1,423
列表
看板
标记
里程碑
合并请求
543
合并请求
543
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
1199aa68
编写于
10月 16, 2017
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
fix bug: clear grad and always share data when output has cpu
and add activation in unit tests
上级
dee7f813
变更
4
隐藏空白更改
内联
并排
Showing
4 changed file
with
31 addition
and
7 deletion
+31
-7
paddle/gserver/layers/MKLDNNConvLayer.cpp
paddle/gserver/layers/MKLDNNConvLayer.cpp
+5
-3
paddle/gserver/layers/MKLDNNLayer.h
paddle/gserver/layers/MKLDNNLayer.h
+18
-1
paddle/gserver/layers/MKLDNNPoolLayer.cpp
paddle/gserver/layers/MKLDNNPoolLayer.cpp
+5
-3
paddle/gserver/tests/test_MKLDNN.cpp
paddle/gserver/tests/test_MKLDNN.cpp
+3
-0
未找到文件。
paddle/gserver/layers/MKLDNNConvLayer.cpp
浏览文件 @
1199aa68
...
...
@@ -313,6 +313,7 @@ void MKLDNNConvLayer::resetOutValue(
cvtOutVal_
=
MKLDNNMatrix
::
createReorder
(
out
,
cpuOutVal_
);
CHECK
(
cvtOutVal_
)
<<
"should not be empty"
;
}
else
{
cpuOut
->
setData
(
output_
.
value
->
getData
());
cpuOutVal_
=
out
;
}
// when output is cpu device, change the mkldnn output value and make them
...
...
@@ -456,17 +457,18 @@ void MKLDNNConvLayer::resetOutGrad(
MKLDNNLayer
::
resetOutGrad
(
out
,
outVal_
->
getPrimitiveDesc
());
}
else
{
const
MatrixPtr
&
cpuOut
=
getOutput
(
CPU_DEVICE
).
grad
;
// always share the same grad data of CPU output
// then the activation can get the right grad from output_.grad
output_
.
grad
->
setData
(
cpuOut
->
getData
());
// same PrimitiveDesc with cpuInVal_
CHECK
(
cpuOutVal_
);
cpuOutGrad_
=
MKLDNNMatrix
::
create
(
cpuOut
,
cpuOutVal_
->
getPrimitiveDesc
());
// create reorder if primitive desc does not match
if
(
cpuOutGrad_
->
getPrimitiveDesc
()
!=
outVal_
->
getPrimitiveDesc
())
{
out
=
MKLDNNMatrix
::
create
(
output_
.
grad
,
outVal_
->
getPrimitiveDesc
());
out
=
MKLDNNMatrix
::
create
(
nullptr
,
outVal_
->
getPrimitiveDesc
());
cvtOutGrad_
=
MKLDNNMatrix
::
createReorder
(
cpuOutGrad_
,
out
);
CHECK
(
cvtOutGrad_
);
}
else
{
// share the same data of CPU output
output_
.
grad
->
setData
(
cpuOut
->
getData
());
out
=
cpuOutGrad_
;
}
}
...
...
paddle/gserver/layers/MKLDNNLayer.h
浏览文件 @
1199aa68
...
...
@@ -46,6 +46,9 @@ protected:
// backward also need reset after reset forward handle
bool
needResetBwd_
;
// is output only mkldnn
bool
outputOnlyMKLDNN_
;
// mkldnn engine, stream and primivtives
mkldnn
::
engine
engine_
;
std
::
shared_ptr
<
MKLDNNStream
>
stream_
;
...
...
@@ -141,6 +144,9 @@ public:
updateInputData
();
}
if
(
!
outputOnlyMKLDNN_
)
{
clearGrads
();
}
stream_
->
submit
(
pipelineFwd_
);
}
...
...
@@ -389,7 +395,8 @@ protected:
CHECK_EQ
(
outputOtherDevice_
[
i
].
deviceId
,
CPU_DEVICE
)
<<
"Only support other device is CPU yet"
;
}
return
outputOtherDevice_
.
size
()
==
0
;
outputOnlyMKLDNN_
=
outputOtherDevice_
.
size
()
==
0
;
return
outputOnlyMKLDNN_
;
}
/**
...
...
@@ -398,6 +405,16 @@ protected:
void
setDevice
(
int
id
)
{
deviceId_
=
id
;
}
private:
/**
* clear all grad
*/
void
clearGrads
()
{
output_
.
grad
->
zeroMem
();
for
(
size_t
i
=
0
;
i
<
outputOtherDevice_
.
size
();
i
++
)
{
outputOtherDevice_
[
i
].
grad
->
zeroMem
();
}
}
/**
* Set deviceId of the params used in this layer.
*/
...
...
paddle/gserver/layers/MKLDNNPoolLayer.cpp
浏览文件 @
1199aa68
...
...
@@ -146,6 +146,7 @@ void MKLDNNPoolLayer::resetOutValue(MKLDNNMatrixPtr& out) {
cvtOutVal_
=
MKLDNNMatrix
::
createReorder
(
out
,
cpuOutVal_
);
CHECK
(
cvtOutVal_
)
<<
"should not be emptry"
;
}
else
{
cpuOut
->
setData
(
output_
.
value
->
getData
());
cpuOutVal_
=
out
;
}
output_
.
value
=
std
::
dynamic_pointer_cast
<
Matrix
>
(
cpuOutVal_
);
...
...
@@ -213,15 +214,16 @@ void MKLDNNPoolLayer::resetOutGrad(MKLDNNMatrixPtr& out) {
MKLDNNLayer
::
resetOutGrad
(
out
,
outVal_
->
getPrimitiveDesc
());
}
else
{
const
MatrixPtr
&
cpuOut
=
getOutput
(
CPU_DEVICE
).
grad
;
// always share the same grad data of CPU output
// then the activation can get the right grad from output_.grad
output_
.
grad
->
setData
(
cpuOut
->
getData
());
cpuOutGrad_
=
MKLDNNMatrix
::
create
(
cpuOut
,
memory
::
dims
{
bs_
,
oc_
,
oh_
,
ow_
},
format
::
nchw
,
engine_
);
if
(
cpuOutGrad_
->
getPrimitiveDesc
()
!=
outVal_
->
getPrimitiveDesc
())
{
out
=
MKLDNNMatrix
::
create
(
output_
.
grad
,
outVal_
->
getPrimitiveDesc
());
out
=
MKLDNNMatrix
::
create
(
nullptr
,
outVal_
->
getPrimitiveDesc
());
cvtOutGrad_
=
MKLDNNMatrix
::
createReorder
(
cpuOutGrad_
,
out
);
CHECK
(
cvtOutGrad_
)
<<
"should not be emptry"
;
}
else
{
// share the same data of CPU output
output_
.
grad
->
setData
(
cpuOut
->
getData
());
out
=
cpuOutGrad_
;
}
}
...
...
paddle/gserver/tests/test_MKLDNN.cpp
浏览文件 @
1199aa68
...
...
@@ -46,6 +46,7 @@ struct testFcDesc {
static
void
getMKLDNNFcConfig
(
TestConfig
&
cfg
,
const
testFcDesc
&
pm
)
{
cfg
.
layerConfig
.
set_type
(
"mkldnn_fc"
);
cfg
.
layerConfig
.
set_active_type
(
"sigmoid"
);
cfg
.
layerConfig
.
set_size
(
pm
.
oc
);
cfg
.
inputDefs
.
push_back
(
{
INPUT_DATA
,
...
...
@@ -86,6 +87,7 @@ struct testConvDesc {
static
void
getMKLDNNConvConfig
(
TestConfig
&
cfg
,
const
testConvDesc
&
pm
)
{
cfg
.
layerConfig
.
set_type
(
"mkldnn_conv"
);
cfg
.
layerConfig
.
set_active_type
(
"relu"
);
cfg
.
layerConfig
.
set_num_filters
(
pm
.
oc
);
cfg
.
layerConfig
.
set_size
(
pm
.
oc
*
pm
.
oh
*
pm
.
ow
);
cfg
.
layerConfig
.
set_shared_biases
(
true
);
...
...
@@ -158,6 +160,7 @@ struct testPoolDesc {
static
void
getMKLDNNPoolConfig
(
TestConfig
&
cfg
,
const
testPoolDesc
&
pm
)
{
cfg
.
layerConfig
.
set_type
(
"mkldnn_pool"
);
cfg
.
layerConfig
.
set_active_type
(
"relu"
);
cfg
.
layerConfig
.
set_size
(
pm
.
ic
*
pm
.
oh
*
pm
.
ow
);
cfg
.
inputDefs
.
push_back
(
{
INPUT_DATA
,
...
...
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