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
OPTHREE
Paddle
提交
f11934cb
P
Paddle
项目概览
OPTHREE
/
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
f11934cb
编写于
10月 30, 2018
作者:
T
Tomasz Patejko
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
MKLDNN conv residual data: residual data is reorder when formats are incorrect
上级
62a0fe08
变更
1
隐藏空白更改
内联
并排
Showing
1 changed file
with
30 addition
and
14 deletion
+30
-14
paddle/fluid/operators/conv_mkldnn_op.cc
paddle/fluid/operators/conv_mkldnn_op.cc
+30
-14
未找到文件。
paddle/fluid/operators/conv_mkldnn_op.cc
浏览文件 @
f11934cb
...
...
@@ -15,6 +15,8 @@
#include "paddle/fluid/operators/conv_op.h"
#include "paddle/fluid/platform/mkldnn_helper.h"
#include "paddle/fluid/framework/data_layout_transform.h"
namespace
paddle
{
namespace
operators
{
...
...
@@ -108,6 +110,11 @@ class ConvMKLDNNHandler : public platform::MKLDNNHandler {
"@data-weights_mem_p"
,
pipeline
);
}
std
::
shared_ptr
<
mkldnn
::
memory
>
AcquireResidualDataMemory
(
const
mkldnn
::
memory
::
desc
&
md
,
void
*
ptr
)
{
return
this
->
AcquireMemory
(
md
,
ptr
,
"@user_residual_data_mem_p"
);
}
std
::
shared_ptr
<
mkldnn
::
memory
>
AcquireDiffSrcMemoryFromDataPrimitive
(
void
*
ptr
)
{
return
this
->
AcquireMemoryFromPrimitive
(
...
...
@@ -386,7 +393,15 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
auto
user_weights_memory_p
=
handler
.
AcquireWeightsMemory
(
user_weights_md
,
to_void_cast
<
T
>
(
filter_data
));
T
*
output_data
=
nullptr
;
// create reorder primitive if the input format is not the preferred one
auto
src_memory_p
=
handler
.
AcquireSrcMemoryFromPrimitive
(
user_src_memory_p
,
pipeline
);
auto
weights_memory_p
=
handler
.
AcquireWeightsMemoryFromPrimitive
(
user_weights_memory_p
,
pipeline
,
is_test
);
auto
output_data
=
output
->
mutable_data
<
T
>
(
ctx
.
GetPlace
(),
handler
.
GetDstMemorySize
());
auto
dst_memory_p
=
handler
.
AcquireDstMemoryFromPrimitive
(
to_void_cast
<
T
>
(
output_data
));
if
(
fuse_residual_conn
)
{
auto
residual_param
=
ctx
.
Input
<
Tensor
>
(
"ResidualData"
);
...
...
@@ -399,21 +414,22 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
"Output and elementwise parameter need to have the "
"same dimension sizes"
);
output
->
ShareDataWith
(
*
residual_param
);
output_data
=
output
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
}
else
{
output_data
=
output
->
mutable_data
<
T
>
(
ctx
.
GetPlace
(),
handler
.
GetDstMemorySize
());
if
(
residual_param
->
format
()
!=
output
->
format
())
{
auto
residual_data_tz
=
paddle
::
framework
::
vectorize2int
(
residual_param
->
dims
());
auto
residual_data_type
=
paddle
::
framework
::
ToMKLDNNDataType
(
residual_param
->
type
());
auto
user_residual_md
=
platform
::
MKLDNNMemDesc
(
residual_data_tz
,
residual_data_type
,
residual_param
->
format
());
auto
user_residual_memory_p
=
handler
.
AcquireResidualDataMemory
(
user_residual_md
,
to_void_cast
<
T
>
(
residual_param_data
));
platform
::
Reorder
(
*
user_residual_memory_p
,
*
dst_memory_p
);
}
else
{
output
->
ShareDataWith
(
*
residual_param
);
}
}
// create reorder primitive if the input format is not the preferred one
auto
src_memory_p
=
handler
.
AcquireSrcMemoryFromPrimitive
(
user_src_memory_p
,
pipeline
);
auto
weights_memory_p
=
handler
.
AcquireWeightsMemoryFromPrimitive
(
user_weights_memory_p
,
pipeline
,
is_test
);
auto
dst_memory_p
=
handler
.
AcquireDstMemoryFromPrimitive
(
to_void_cast
<
T
>
(
output_data
));
// create convolution op primitive
std
::
shared_ptr
<
mkldnn
::
convolution_forward
>
conv_p
;
if
(
bias
)
{
...
...
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