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
urhero
Paddle
提交
7b54f168
P
Paddle
项目概览
urhero
/
Paddle
与 Fork 源项目一致
Fork自
PaddlePaddle / Paddle
通知
1
Star
0
Fork
0
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
0
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
0
Issue
0
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
7b54f168
编写于
6月 29, 2018
作者:
C
chengduo
提交者:
GitHub
6月 29, 2018
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Follow comment (#11845)
上级
b0378963
变更
2
隐藏空白更改
内联
并排
Showing
2 changed file
with
17 addition
and
15 deletion
+17
-15
paddle/fluid/framework/tensor_util.cc
paddle/fluid/framework/tensor_util.cc
+2
-15
paddle/fluid/framework/tensor_util.h
paddle/fluid/framework/tensor_util.h
+15
-0
未找到文件。
paddle/fluid/framework/tensor_util.cc
浏览文件 @
7b54f168
...
...
@@ -73,18 +73,12 @@ void TensorCopy(const Tensor& src, const platform::Place& dst_place,
memory
::
Copy
(
dst_gpu_place
,
dst_ptr
,
src_gpu_place
,
src_ptr
,
size
,
stream
);
}
else
{
// NOTE(zcd): Because TensorCopy is an async operation, when the src_place
// and dst_place are two different GPU, to ensure that the operation can
// be carried out correctly, we should make ctx wait.
// If ctx_place and src_place are the same, we should add ctx.Wait()
// after memory::Copy; if ctx_place and dst_place are the same, we should
// add ctx.Wait() before memory::Copy.
if
(
platform
::
is_same_place
(
ctx_place
,
src_place
))
{
memory
::
Copy
(
dst_gpu_place
,
dst_ptr
,
src_gpu_place
,
src_ptr
,
size
,
stream
);
ctx
.
Wait
();
platform
::
DeviceContextPool
::
Instance
().
Get
(
src
.
place
())
->
Wait
();
}
else
if
(
platform
::
is_same_place
(
ctx_place
,
dst_place
))
{
ctx
.
Wait
();
platform
::
DeviceContextPool
::
Instance
().
Get
(
src
.
place
())
->
Wait
();
memory
::
Copy
(
dst_gpu_place
,
dst_ptr
,
src_gpu_place
,
src_ptr
,
size
,
stream
);
}
else
{
...
...
@@ -97,13 +91,6 @@ void TensorCopy(const Tensor& src, const platform::Place& dst_place,
void
TensorCopy
(
const
Tensor
&
src
,
const
platform
::
Place
&
dst_place
,
Tensor
*
dst
)
{
// NOTE(zcd): If the src.place() and dst_place are two different GPU,
// the copy operation is carried out on the dst_place's stream. This is
// very important, because TensorCopy is an async operator, and in most
// case, once this copy operator returns, dst is to be used in dst_place's
// stream, if this copy operation is carried out on the src_place's stream,
// when dst is used in dst_place's stream the copy operation may be
// not completed.
platform
::
DeviceContextPool
&
pool
=
platform
::
DeviceContextPool
::
Instance
();
const
platform
::
DeviceContext
*
dev_ctx
;
if
(
platform
::
is_gpu_place
(
dst_place
))
{
...
...
paddle/fluid/framework/tensor_util.h
浏览文件 @
7b54f168
...
...
@@ -23,10 +23,25 @@ limitations under the License. */
namespace
paddle
{
namespace
framework
{
// NOTE(zcd): Because TensorCopy is an async operation, when the src_place
// and dst_place are two different GPU, to ensure that the operation can
// be carried out correctly, there is a src_ctx wait operation in TensorCopy.
// If ctx_place and src_place are the same, src_ctx.Wait() is added
// after memory::Copy; if ctx_place and dst_place are the same,
// src_ctx.Wait() is added before memory::Copy.
void
TensorCopy
(
const
Tensor
&
src
,
const
platform
::
Place
&
dst_place
,
const
platform
::
DeviceContext
&
ctx
,
Tensor
*
dst
);
// NOTE(zcd): If the src.place() and dst_place are two different GPU,
// the copy operation is carried out on the dst_place's stream. This is
// very important, because TensorCopy is an async operator, and in most
// case, once this copy operator returns, dst is to be used in dst_place's
// stream, if this copy operation is carried out on the src_place's stream,
// when dst is used in dst_place's stream the copy operation may be
// not completed.
void
TensorCopy
(
const
Tensor
&
src
,
const
platform
::
Place
&
dst_place
,
Tensor
*
dst
);
void
TensorCopySync
(
const
Tensor
&
src
,
const
platform
::
Place
&
dst_place
,
Tensor
*
dst
);
...
...
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