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
BaiXuePrincess
Paddle
提交
0c24b3f9
P
Paddle
项目概览
BaiXuePrincess
/
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
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
提交
0c24b3f9
编写于
4月 25, 2018
作者:
Y
Yu Yang
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Clean memcpy async
上级
bfbbe19f
变更
2
隐藏空白更改
内联
并排
Showing
2 changed file
with
8 addition
and
33 deletion
+8
-33
paddle/fluid/framework/details/fetch_op_handle.cc
paddle/fluid/framework/details/fetch_op_handle.cc
+0
-1
paddle/fluid/pybind/tensor_py.h
paddle/fluid/pybind/tensor_py.h
+8
-32
未找到文件。
paddle/fluid/framework/details/fetch_op_handle.cc
浏览文件 @
0c24b3f9
...
@@ -67,7 +67,6 @@ void FetchOpHandle::RunImpl() {
...
@@ -67,7 +67,6 @@ void FetchOpHandle::RunImpl() {
if
(
platform
::
is_gpu_place
(
t
.
place
()))
{
if
(
platform
::
is_gpu_place
(
t
.
place
()))
{
#ifdef PADDLE_WITH_CUDA
#ifdef PADDLE_WITH_CUDA
TensorCopy
(
t
,
cpu
,
*
dev_ctxes_
[
t
.
place
()],
&
tensors_
[
i
],
true
);
TensorCopy
(
t
,
cpu
,
*
dev_ctxes_
[
t
.
place
()],
&
tensors_
[
i
],
true
);
dev_ctxes_
.
at
(
t
.
place
())
->
Wait
();
#endif
#endif
}
else
{
}
else
{
tensors_
[
i
].
ShareDataWith
(
t
);
tensors_
[
i
].
ShareDataWith
(
t
);
...
...
paddle/fluid/pybind/tensor_py.h
浏览文件 @
0c24b3f9
...
@@ -63,15 +63,9 @@ struct CastToPyBufferImpl<true, I, ARGS...> {
...
@@ -63,15 +63,9 @@ struct CastToPyBufferImpl<true, I, ARGS...> {
auto
*
dst_ptr
=
static_cast
<
void
*>
(
dst_tensor
.
mutable_data
<
CUR_TYPE
>
(
auto
*
dst_ptr
=
static_cast
<
void
*>
(
dst_tensor
.
mutable_data
<
CUR_TYPE
>
(
tensor
.
dims
(),
platform
::
CPUPlace
()));
tensor
.
dims
(),
platform
::
CPUPlace
()));
platform
::
DeviceContextPool
&
pool
=
paddle
::
platform
::
GpuMemcpySync
(
dst_ptr
,
src_ptr
,
platform
::
DeviceContextPool
::
Instance
();
sizeof
(
CUR_TYPE
)
*
tensor
.
numel
(),
auto
dev_ctx
=
static_cast
<
const
platform
::
CUDADeviceContext
*>
(
cudaMemcpyDeviceToHost
);
pool
.
Get
(
tensor
.
place
()));
paddle
::
platform
::
GpuMemcpyAsync
(
dst_ptr
,
src_ptr
,
sizeof
(
CUR_TYPE
)
*
tensor
.
numel
(),
cudaMemcpyDeviceToHost
,
dev_ctx
->
stream
());
dev_ctx
->
Wait
();
#else
#else
PADDLE_THROW
(
"'CUDAPlace' is not supported in CPU only device."
);
PADDLE_THROW
(
"'CUDAPlace' is not supported in CPU only device."
);
#endif
#endif
...
@@ -184,17 +178,8 @@ void PyCUDATensorSetFromArray(
...
@@ -184,17 +178,8 @@ void PyCUDATensorSetFromArray(
self
->
Resize
(
framework
::
make_ddim
(
dims
));
self
->
Resize
(
framework
::
make_ddim
(
dims
));
auto
*
dst
=
self
->
mutable_data
<
T
>
(
place
);
auto
*
dst
=
self
->
mutable_data
<
T
>
(
place
);
paddle
::
platform
::
GpuMemcpySync
(
dst
,
array
.
data
(),
sizeof
(
T
)
*
array
.
size
(),
platform
::
DeviceContextPool
&
pool
=
platform
::
DeviceContextPool
::
Instance
();
cudaMemcpyHostToDevice
);
auto
dev_ctx
=
static_cast
<
const
platform
::
CUDADeviceContext
*>
(
pool
.
Get
(
place
));
paddle
::
platform
::
GpuMemcpyAsync
(
dst
,
array
.
data
(),
sizeof
(
T
)
*
array
.
size
(),
cudaMemcpyHostToDevice
,
dev_ctx
->
stream
());
// NOTE: For safety, here wait the copy complete.
// It because the CPU array.data() could be destroyed after this method.
// If we make this method async, it could be copied data from a memory buffer
// that has been freed.
dev_ctx
->
Wait
();
}
}
template
<
>
template
<
>
...
@@ -214,18 +199,9 @@ void PyCUDATensorSetFromArray(
...
@@ -214,18 +199,9 @@ void PyCUDATensorSetFromArray(
self
->
Resize
(
framework
::
make_ddim
(
dims
));
self
->
Resize
(
framework
::
make_ddim
(
dims
));
auto
*
dst
=
self
->
mutable_data
<
platform
::
float16
>
(
place
);
auto
*
dst
=
self
->
mutable_data
<
platform
::
float16
>
(
place
);
paddle
::
platform
::
GpuMemcpySync
(
dst
,
array
.
data
(),
platform
::
DeviceContextPool
&
pool
=
platform
::
DeviceContextPool
::
Instance
();
sizeof
(
uint16_t
)
*
array
.
size
(),
auto
dev_ctx
=
cudaMemcpyHostToDevice
);
static_cast
<
const
platform
::
CUDADeviceContext
*>
(
pool
.
Get
(
place
));
paddle
::
platform
::
GpuMemcpyAsync
(
dst
,
array
.
data
(),
sizeof
(
uint16_t
)
*
array
.
size
(),
cudaMemcpyHostToDevice
,
dev_ctx
->
stream
());
// NOTE: For safety, here wait the copy complete.
// It because the CPU array.data() could be destroyed after this method.
// If we make this method async, it could be copied data from a memory buffer
// that has been freed.
dev_ctx
->
Wait
();
}
}
template
<
typename
T
>
template
<
typename
T
>
...
...
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