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
bc921927
P
Paddle
项目概览
PaddlePaddle
/
Paddle
大约 1 年 前同步成功
通知
2292
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 >>
提交
bc921927
编写于
2月 08, 2019
作者:
D
Dun Liang
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Fix Pr #15296
test=develop
上级
db8fcf6b
变更
6
隐藏空白更改
内联
并排
Showing
6 changed file
with
24 addition
and
6 deletion
+24
-6
paddle/fluid/memory/allocation/legacy_allocator.cc
paddle/fluid/memory/allocation/legacy_allocator.cc
+1
-1
paddle/fluid/memory/allocation/pinned_allocator.cc
paddle/fluid/memory/allocation/pinned_allocator.cc
+1
-1
paddle/fluid/memory/allocation/pinned_allocator.h
paddle/fluid/memory/allocation/pinned_allocator.h
+1
-1
paddle/fluid/memory/detail/system_allocator.cc
paddle/fluid/memory/detail/system_allocator.cc
+2
-2
paddle/fluid/operators/reader/buffered_reader.cc
paddle/fluid/operators/reader/buffered_reader.cc
+17
-1
paddle/fluid/operators/reader/buffered_reader.h
paddle/fluid/operators/reader/buffered_reader.h
+2
-0
未找到文件。
paddle/fluid/memory/allocation/legacy_allocator.cc
浏览文件 @
bc921927
...
@@ -257,7 +257,7 @@ void *Alloc<platform::CUDAPinnedPlace>(const platform::CUDAPinnedPlace &place,
...
@@ -257,7 +257,7 @@ void *Alloc<platform::CUDAPinnedPlace>(const platform::CUDAPinnedPlace &place,
void
*
ptr
=
buddy_allocator
->
Alloc
(
size
);
void
*
ptr
=
buddy_allocator
->
Alloc
(
size
);
if
(
ptr
==
nullptr
)
{
if
(
ptr
==
nullptr
)
{
LOG
(
WARNING
)
<<
"cuda
MallocHost
Cannot allocate "
<<
size
LOG
(
WARNING
)
<<
"cuda
HostAlloc
Cannot allocate "
<<
size
<<
" bytes in CUDAPinnedPlace"
;
<<
" bytes in CUDAPinnedPlace"
;
}
}
if
(
FLAGS_init_allocated_mem
)
{
if
(
FLAGS_init_allocated_mem
)
{
...
...
paddle/fluid/memory/allocation/pinned_allocator.cc
浏览文件 @
bc921927
...
@@ -32,7 +32,7 @@ Allocation *CPUPinnedAllocator::AllocateImpl(size_t size,
...
@@ -32,7 +32,7 @@ Allocation *CPUPinnedAllocator::AllocateImpl(size_t size,
// "CPUPinnedAllocator should be used for Cross-Device Communication");
// "CPUPinnedAllocator should be used for Cross-Device Communication");
void
*
ptr
;
void
*
ptr
;
PADDLE_ENFORCE
(
cuda
MallocHost
(
&
ptr
,
siz
e
));
PADDLE_ENFORCE
(
cuda
HostAlloc
(
&
ptr
,
size
,
cudaHostAllocPortabl
e
));
return
new
CPUPinnedAllocation
(
ptr
,
size
);
return
new
CPUPinnedAllocation
(
ptr
,
size
);
}
}
}
// namespace allocation
}
// namespace allocation
...
...
paddle/fluid/memory/allocation/pinned_allocator.h
浏览文件 @
bc921927
...
@@ -19,7 +19,7 @@ namespace paddle {
...
@@ -19,7 +19,7 @@ namespace paddle {
namespace
memory
{
namespace
memory
{
namespace
allocation
{
namespace
allocation
{
// Allocator uses `cuda
MallocHost
`
// Allocator uses `cuda
HostAlloc
`
class
CPUPinnedAllocation
:
public
Allocation
{
class
CPUPinnedAllocation
:
public
Allocation
{
public:
public:
CPUPinnedAllocation
(
void
*
ptr
,
size_t
size
)
CPUPinnedAllocation
(
void
*
ptr
,
size_t
size
)
...
...
paddle/fluid/memory/detail/system_allocator.cc
浏览文件 @
bc921927
...
@@ -173,14 +173,14 @@ void* CUDAPinnedAllocator::Alloc(size_t* index, size_t size) {
...
@@ -173,14 +173,14 @@ void* CUDAPinnedAllocator::Alloc(size_t* index, size_t size) {
void
*
p
;
void
*
p
;
// PINNED memory is visible to all CUDA contexts.
// PINNED memory is visible to all CUDA contexts.
cudaError_t
result
=
cuda
MallocHost
(
&
p
,
siz
e
);
cudaError_t
result
=
cuda
HostAlloc
(
&
p
,
size
,
cudaHostAllocPortabl
e
);
if
(
result
==
cudaSuccess
)
{
if
(
result
==
cudaSuccess
)
{
*
index
=
1
;
// PINNED memory
*
index
=
1
;
// PINNED memory
cuda_pinnd_alloc_size_
+=
size
;
cuda_pinnd_alloc_size_
+=
size
;
return
p
;
return
p
;
}
else
{
}
else
{
LOG
(
WARNING
)
<<
"cuda
MallocHost
failed."
;
LOG
(
WARNING
)
<<
"cuda
HostAlloc
failed."
;
return
nullptr
;
return
nullptr
;
}
}
...
...
paddle/fluid/operators/reader/buffered_reader.cc
浏览文件 @
bc921927
...
@@ -29,6 +29,7 @@ BufferedReader::~BufferedReader() {
...
@@ -29,6 +29,7 @@ BufferedReader::~BufferedReader() {
if
(
platform
::
is_gpu_place
(
place_
))
{
if
(
platform
::
is_gpu_place
(
place_
))
{
platform
::
SetDeviceId
(
boost
::
get
<
platform
::
CUDAPlace
>
(
place_
).
device
);
platform
::
SetDeviceId
(
boost
::
get
<
platform
::
CUDAPlace
>
(
place_
).
device
);
PADDLE_ENFORCE
(
cudaStreamDestroy
(
stream
));
PADDLE_ENFORCE
(
cudaStreamDestroy
(
stream
));
for
(
auto
&
event
:
events
)
PADDLE_ENFORCE
(
cudaEventDestroy
(
event
));
}
}
#endif
#endif
}
}
...
@@ -43,7 +44,14 @@ BufferedReader::BufferedReader(
...
@@ -43,7 +44,14 @@ BufferedReader::BufferedReader(
#ifdef PADDLE_WITH_CUDA
#ifdef PADDLE_WITH_CUDA
if
(
platform
::
is_gpu_place
(
place_
))
{
if
(
platform
::
is_gpu_place
(
place_
))
{
platform
::
SetDeviceId
(
boost
::
get
<
platform
::
CUDAPlace
>
(
place_
).
device
);
platform
::
SetDeviceId
(
boost
::
get
<
platform
::
CUDAPlace
>
(
place_
).
device
);
PADDLE_ENFORCE
(
cudaStreamCreate
(
&
stream
));
compute_stream
=
((
platform
::
CUDADeviceContext
*
)(
platform
::
DeviceContextPool
::
Instance
()
.
Get
(
place_
)))
->
stream
();
events
.
resize
(
buffer_size
);
for
(
auto
&
event
:
events
)
PADDLE_ENFORCE
(
cudaEventCreateWithFlags
(
&
event
,
cudaEventDisableTiming
));
PADDLE_ENFORCE
(
cudaStreamCreateWithFlags
(
&
stream
,
cudaStreamNonBlocking
));
}
}
#endif
#endif
cpu_buffer_
.
resize
(
buffer_size
);
cpu_buffer_
.
resize
(
buffer_size
);
...
@@ -59,6 +67,12 @@ void BufferedReader::ReadTillBufferFullAsync() {
...
@@ -59,6 +67,12 @@ void BufferedReader::ReadTillBufferFullAsync() {
}
}
void
BufferedReader
::
ReadAsync
(
size_t
i
)
{
void
BufferedReader
::
ReadAsync
(
size_t
i
)
{
#ifdef PADDLE_WITH_CUDA
if
(
platform
::
is_gpu_place
(
place_
))
{
platform
::
SetDeviceId
(
boost
::
get
<
platform
::
CUDAPlace
>
(
place_
).
device
);
PADDLE_ENFORCE
(
cudaEventRecord
(
events
[
i
],
compute_stream
));
}
#endif
position_
.
emplace
(
thread_pool_
.
enqueue
([
this
,
i
]()
->
size_t
{
position_
.
emplace
(
thread_pool_
.
enqueue
([
this
,
i
]()
->
size_t
{
TensorVec
&
cpu
=
cpu_buffer_
[
i
];
TensorVec
&
cpu
=
cpu_buffer_
[
i
];
reader_
->
ReadNext
(
&
cpu
);
reader_
->
ReadNext
(
&
cpu
);
...
@@ -71,6 +85,8 @@ void BufferedReader::ReadAsync(size_t i) {
...
@@ -71,6 +85,8 @@ void BufferedReader::ReadAsync(size_t i) {
// NOTE(liangdun): using async copy instead of TensorCopySync
// NOTE(liangdun): using async copy instead of TensorCopySync
// TensorCopySync would block other stream
// TensorCopySync would block other stream
if
(
platform
::
is_gpu_place
(
place_
))
{
if
(
platform
::
is_gpu_place
(
place_
))
{
platform
::
SetDeviceId
(
boost
::
get
<
platform
::
CUDAPlace
>
(
place_
).
device
);
PADDLE_ENFORCE
(
cudaStreamWaitEvent
(
stream
,
events
[
i
],
0
));
TensorVec
&
gpu
=
gpu_buffer_
[
i
];
TensorVec
&
gpu
=
gpu_buffer_
[
i
];
gpu
.
resize
(
cpu
.
size
());
gpu
.
resize
(
cpu
.
size
());
for
(
size_t
i
=
0
;
i
<
cpu
.
size
();
++
i
)
{
for
(
size_t
i
=
0
;
i
<
cpu
.
size
();
++
i
)
{
...
...
paddle/fluid/operators/reader/buffered_reader.h
浏览文件 @
bc921927
...
@@ -64,6 +64,8 @@ class BufferedReader : public framework::DecoratedReader {
...
@@ -64,6 +64,8 @@ class BufferedReader : public framework::DecoratedReader {
size_t
prev_pos_
{
-
1UL
};
size_t
prev_pos_
{
-
1UL
};
#ifdef PADDLE_WITH_CUDA
#ifdef PADDLE_WITH_CUDA
cudaStream_t
stream
;
cudaStream_t
stream
;
cudaStream_t
compute_stream
;
std
::
vector
<
cudaEvent_t
>
events
;
#endif
#endif
};
};
...
...
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