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
a1ad3a63
P
Paddle
项目概览
PaddlePaddle
/
Paddle
大约 1 年 前同步成功
通知
2297
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 >>
未验证
提交
a1ad3a63
编写于
12月 08, 2021
作者:
S
sneaxiy
提交者:
GitHub
12月 08, 2021
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Fix CUDA Graph H2D bug by restore host memory (#37774)
* fix CUDA Graph H2D bug again * fix no return bug
上级
9a2d327c
变更
2
隐藏空白更改
内联
并排
Showing
2 changed file
with
38 addition
and
28 deletion
+38
-28
paddle/fluid/operators/math/concat_and_split.cu
paddle/fluid/operators/math/concat_and_split.cu
+21
-28
paddle/fluid/platform/cuda_graph_with_memory_pool.h
paddle/fluid/platform/cuda_graph_with_memory_pool.h
+17
-0
未找到文件。
paddle/fluid/operators/math/concat_and_split.cu
浏览文件 @
a1ad3a63
...
...
@@ -287,13 +287,11 @@ class ConcatFunctor<platform::CUDADeviceContext, T> {
const
T
**
dev_ins_data
=
nullptr
;
if
(
!
has_same_shape
||
in_num
<
2
||
in_num
>
4
)
{
tmp_dev_ins_data
=
memory
::
Alloc
(
context
,
in_num
*
sizeof
(
T
*
));
{
platform
::
SkipCUDAGraphCaptureGuard
guard
;
memory
::
Copy
(
BOOST_GET_CONST
(
platform
::
CUDAPlace
,
context
.
GetPlace
()),
tmp_dev_ins_data
->
ptr
(),
platform
::
CPUPlace
(),
static_cast
<
void
*>
(
inputs_data
),
in_num
*
sizeof
(
T
*
),
context
.
stream
());
}
auto
*
restored
=
platform
::
RestoreHostMemIfCapturingCUDAGraph
(
inputs_data
,
in_num
);
memory
::
Copy
(
BOOST_GET_CONST
(
platform
::
CUDAPlace
,
context
.
GetPlace
()),
tmp_dev_ins_data
->
ptr
(),
platform
::
CPUPlace
(),
restored
,
in_num
*
sizeof
(
T
*
),
context
.
stream
());
dev_ins_data
=
reinterpret_cast
<
const
T
**>
(
tmp_dev_ins_data
->
ptr
());
}
...
...
@@ -317,13 +315,12 @@ class ConcatFunctor<platform::CUDADeviceContext, T> {
}
else
{
auto
tmp_dev_ins_col_data
=
memory
::
Alloc
(
context
,
inputs_col_num
*
sizeof
(
int64_t
));
{
platform
::
SkipCUDAGraphCaptureGuard
guard
;
memory
::
Copy
(
BOOST_GET_CONST
(
platform
::
CUDAPlace
,
context
.
GetPlace
()),
tmp_dev_ins_col_data
->
ptr
(),
platform
::
CPUPlace
(),
static_cast
<
void
*>
(
inputs_col
),
inputs_col_num
*
sizeof
(
int64_t
),
context
.
stream
());
}
auto
*
restored
=
platform
::
RestoreHostMemIfCapturingCUDAGraph
(
inputs_col
,
inputs_col_num
);
memory
::
Copy
(
BOOST_GET_CONST
(
platform
::
CUDAPlace
,
context
.
GetPlace
()),
tmp_dev_ins_col_data
->
ptr
(),
platform
::
CPUPlace
(),
restored
,
inputs_col_num
*
sizeof
(
int64_t
),
context
.
stream
());
int64_t
*
dev_ins_col_data
=
static_cast
<
int64_t
*>
(
tmp_dev_ins_col_data
->
ptr
());
...
...
@@ -422,13 +419,11 @@ class SplitFunctor<platform::CUDADeviceContext, T> {
T
**
dev_out_gpu_data
=
nullptr
;
if
(
!
has_same_shape
||
o_num
<
2
||
o_num
>
4
)
{
tmp_dev_outs_data
=
memory
::
Alloc
(
context
,
o_num
*
sizeof
(
T
*
));
{
platform
::
SkipCUDAGraphCaptureGuard
guard
;
memory
::
Copy
(
BOOST_GET_CONST
(
platform
::
CUDAPlace
,
context
.
GetPlace
()),
tmp_dev_outs_data
->
ptr
(),
platform
::
CPUPlace
(),
reinterpret_cast
<
void
*>
(
outputs_data
),
o_num
*
sizeof
(
T
*
),
context
.
stream
());
}
auto
*
restored
=
platform
::
RestoreHostMemIfCapturingCUDAGraph
(
outputs_data
,
o_num
);
memory
::
Copy
(
BOOST_GET_CONST
(
platform
::
CUDAPlace
,
context
.
GetPlace
()),
tmp_dev_outs_data
->
ptr
(),
platform
::
CPUPlace
(),
restored
,
o_num
*
sizeof
(
T
*
),
context
.
stream
());
dev_out_gpu_data
=
reinterpret_cast
<
T
**>
(
tmp_dev_outs_data
->
ptr
());
}
...
...
@@ -452,13 +447,11 @@ class SplitFunctor<platform::CUDADeviceContext, T> {
}
else
{
auto
tmp_dev_ins_col_data
=
memory
::
Alloc
(
context
,
outputs_cols_num
*
sizeof
(
int64_t
));
{
platform
::
SkipCUDAGraphCaptureGuard
guard
;
memory
::
Copy
(
BOOST_GET_CONST
(
platform
::
CUDAPlace
,
context
.
GetPlace
()),
tmp_dev_ins_col_data
->
ptr
(),
platform
::
CPUPlace
(),
reinterpret_cast
<
void
*>
(
outputs_cols
),
outputs_cols_num
*
sizeof
(
int64_t
),
context
.
stream
());
}
auto
*
restored
=
platform
::
RestoreHostMemIfCapturingCUDAGraph
(
outputs_cols
,
outputs_cols_num
);
memory
::
Copy
(
BOOST_GET_CONST
(
platform
::
CUDAPlace
,
context
.
GetPlace
()),
tmp_dev_ins_col_data
->
ptr
(),
platform
::
CPUPlace
(),
restored
,
outputs_cols_num
*
sizeof
(
int64_t
),
context
.
stream
());
int64_t
*
dev_outs_col_data
=
reinterpret_cast
<
int64_t
*>
(
tmp_dev_ins_col_data
->
ptr
());
...
...
paddle/fluid/platform/cuda_graph_with_memory_pool.h
浏览文件 @
a1ad3a63
...
...
@@ -60,6 +60,23 @@ inline void AddResetCallbackIfCapturingCUDAGraph(Callback &&callback) {
callback
();
}
template
<
typename
T
>
inline
T
*
RestoreHostMemIfCapturingCUDAGraph
(
T
*
host_mem
,
size_t
size
)
{
static_assert
(
std
::
is_trivial
<
T
>::
value
,
"T must be trivial type"
);
static_assert
(
!
std
::
is_same
<
T
,
void
>::
value
,
"T cannot be void"
);
#ifdef PADDLE_WITH_CUDA
if
(
UNLIKELY
(
IsCUDAGraphCapturing
()))
{
size_t
nbytes
=
size
*
sizeof
(
T
);
void
*
new_host_mem
=
new
uint8_t
[
nbytes
];
std
::
memcpy
(
new_host_mem
,
host_mem
,
nbytes
);
AddResetCallbackIfCapturingCUDAGraph
(
[
new_host_mem
]
{
delete
[]
reinterpret_cast
<
uint8_t
*>
(
new_host_mem
);
});
return
reinterpret_cast
<
T
*>
(
new_host_mem
);
}
#endif
return
host_mem
;
}
class
SkipCUDAGraphCaptureGuard
{
DISABLE_COPY_AND_ASSIGN
(
SkipCUDAGraphCaptureGuard
);
...
...
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