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
5842ae67
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 >>
未验证
提交
5842ae67
编写于
3月 17, 2020
作者:
A
Adam
提交者:
GitHub
3月 17, 2020
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Revert "Change ShareDataWith() to TensorCopy() in conv_mkldnn (#22695)" (#22985)
上级
6d6cc4ca
变更
2
隐藏空白更改
内联
并排
Showing
2 changed file
with
7 addition
and
19 deletion
+7
-19
paddle/fluid/framework/tensor_util.cc
paddle/fluid/framework/tensor_util.cc
+0
-13
paddle/fluid/operators/mkldnn/conv_mkldnn_op.cc
paddle/fluid/operators/mkldnn/conv_mkldnn_op.cc
+7
-6
未找到文件。
paddle/fluid/framework/tensor_util.cc
浏览文件 @
5842ae67
...
...
@@ -48,19 +48,6 @@ void TensorCopy(const Tensor& src, const platform::Place& dst_place,
}
auto
size
=
src
.
numel
()
*
SizeOfType
(
src
.
type
());
#ifdef PADDLE_WITH_MKLDNN
if
(
src
.
layout
()
==
DataLayout
::
kMKLDNN
)
{
PADDLE_ENFORCE_EQ
(
src
.
memory_size
(),
dst
->
memory_size
(),
platform
::
errors
::
InvalidArgument
(
"When copying tensor with MKL-DNN data layout, "
"memory size of source tensor should be the same as memory size of "
"destination tensor. "
"But received src.memory_size = %d, dst.memory_size = %d."
,
src
.
memory_size
(),
dst
->
memory_size
()));
size
=
src
.
memory_size
();
}
#endif
if
(
platform
::
is_cpu_place
(
src_place
)
&&
platform
::
is_cpu_place
(
dst_place
))
{
memory
::
Copy
(
boost
::
get
<
platform
::
CPUPlace
>
(
dst_place
),
dst_ptr
,
...
...
paddle/fluid/operators/mkldnn/conv_mkldnn_op.cc
浏览文件 @
5842ae67
...
...
@@ -316,9 +316,11 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
dst_memory_p
=
handler
.
AcquireDstMemoryFromResidualDataMemory
(
user_residual_memory_p
,
to_void_cast
<
T
>
(
output_data
),
pipeline
);
}
else
{
auto
output_data
=
output
->
mutable_data
<
T
>
(
ctx
.
GetPlace
(),
residual_param
->
memory_size
());
framework
::
TensorCopy
(
*
residual_param
,
residual_param
->
place
(),
output
);
// Changing ShareDataWith to TensorCopy results in performance drop
// on ResNet architectures
// (https://github.com/PaddlePaddle/Paddle/issues/22964)
output
->
ShareDataWith
(
*
residual_param
);
auto
output_data
=
output
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
dst_memory_p
=
handler
.
AcquireDstMemoryFromPrimitive
(
to_void_cast
<
T
>
(
output_data
));
}
...
...
@@ -611,8 +613,7 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
ctx
,
output
,
residual_param
,
user_residual_md
,
handler
,
&
pipeline
);
}
else
{
framework
::
TensorCopy
(
*
residual_param
,
residual_param
->
place
(),
output
);
output
->
ShareDataWith
(
*
residual_param
);
dst_memory_p
=
platform
::
SetDstMemory
<
T_out
>
(
ctx
,
output
,
handler
);
}
need_s8_to_u8
=
...
...
@@ -683,7 +684,7 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
if
(
fuse_residual_conn
)
{
auto
residual_param
=
ctx
.
Input
<
Tensor
>
(
"ResidualData"
);
framework
::
TensorCopy
(
*
residual_param
,
residual_param
->
place
(),
output
);
output
->
ShareDataWith
(
*
residual_param
);
need_s8_to_u8
=
(
platform
::
MKLDNNGetDataType
<
T_out
>
()
==
memory
::
data_type
::
s8
)
&&
unsigned_output
;
...
...
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