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
15ad7ee4
P
Paddle
项目概览
PaddlePaddle
/
Paddle
大约 1 年 前同步成功
通知
2298
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
未验证
提交
15ad7ee4
编写于
12月 23, 2021
作者:
W
Wilber
提交者:
GitHub
12月 23, 2021
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Support external stream. (#38373)
* support external stream. * update * update * update
上级
b7bafee8
变更
7
隐藏空白更改
内联
并排
Showing
7 changed file
with
106 addition
and
1 deletion
+106
-1
paddle/fluid/inference/api/analysis_predictor.cc
paddle/fluid/inference/api/analysis_predictor.cc
+36
-0
paddle/fluid/inference/api/analysis_predictor.h
paddle/fluid/inference/api/analysis_predictor.h
+6
-0
paddle/fluid/inference/api/paddle_inference_api.h
paddle/fluid/inference/api/paddle_inference_api.h
+17
-0
paddle/fluid/platform/device_context.cc
paddle/fluid/platform/device_context.cc
+20
-0
paddle/fluid/platform/device_context.h
paddle/fluid/platform/device_context.h
+7
-0
paddle/fluid/platform/stream/cuda_stream.cc
paddle/fluid/platform/stream/cuda_stream.cc
+15
-1
paddle/fluid/platform/stream/cuda_stream.h
paddle/fluid/platform/stream/cuda_stream.h
+5
-0
未找到文件。
paddle/fluid/inference/api/analysis_predictor.cc
浏览文件 @
15ad7ee4
...
...
@@ -24,6 +24,7 @@
#include <utility>
#include <vector>
#include "paddle/fluid//platform/device/gpu/gpu_types.h"
#include "paddle/fluid/framework/feed_fetch_method.h"
#include "paddle/fluid/framework/feed_fetch_type.h"
#include "paddle/fluid/framework/ir/fuse_pass_base.h"
...
...
@@ -1043,6 +1044,20 @@ bool AnalysisPredictor::ZeroCopyRun() {
return
true
;
}
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
bool
AnalysisPredictor
::
ExpRunWithExternalStream
(
const
gpuStream_t
stream
)
{
if
(
stream
!=
nullptr
)
{
paddle
::
platform
::
DeviceContextPool
&
pool
=
paddle
::
platform
::
DeviceContextPool
::
Instance
();
auto
gpu_place
=
BOOST_GET_CONST
(
paddle
::
platform
::
CUDAPlace
,
place_
);
auto
*
dev_ctx
=
reinterpret_cast
<
paddle
::
platform
::
CUDADeviceContext
*>
(
pool
.
Get
(
gpu_place
));
dev_ctx
->
SetThreadLocalStream
(
stream
);
}
return
ZeroCopyRun
();
}
#endif
void
AnalysisPredictor
::
CollectShapeRangeInfo
()
{
// if use gpu, sync first.
if
(
config_
.
use_gpu
())
{
...
...
@@ -1567,4 +1582,25 @@ Predictor *PredictorPool::Retrive(size_t idx) {
return
preds_
[
idx
-
1
].
get
();
}
}
// namespace services
namespace
experimental
{
// Note: Can only be used under thread_local semantics.
bool
InternalUtils
::
RunWithExternalStream
(
paddle_infer
::
Predictor
*
p
,
cudaStream_t
stream
)
{
#ifdef PADDLE_WITH_CUDA
auto
pred
=
dynamic_cast
<
paddle
::
AnalysisPredictor
*>
(
p
->
predictor_
.
get
());
return
pred
->
ExpRunWithExternalStream
(
stream
);
#endif
return
false
;
}
bool
InternalUtils
::
RunWithExternalStream
(
paddle_infer
::
Predictor
*
p
,
hipStream_t
stream
)
{
#ifdef PADDLE_WITH_HIP
auto
pred
=
dynamic_cast
<
paddle
::
AnalysisPredictor
*>
(
p
->
predictor_
.
get
());
return
pred
->
ExpRunWithExternalStream
(
stream
);
#endif
return
false
;
}
}
// namespace experimental
}
// namespace paddle_infer
paddle/fluid/inference/api/analysis_predictor.h
浏览文件 @
15ad7ee4
...
...
@@ -25,6 +25,7 @@
#include "paddle/fluid/inference/api/details/reset_tensor_array.h"
#include "paddle/fluid/inference/api/helper.h"
#include "paddle/fluid/inference/api/paddle_inference_api.h"
#include "paddle/fluid/platform/device/gpu/gpu_types.h"
#include "paddle/fluid/platform/float16.h"
#include "paddle/fluid/string/printf.h"
#ifdef PADDLE_WITH_TESTING
...
...
@@ -172,6 +173,11 @@ class AnalysisPredictor : public PaddlePredictor {
///
bool
ZeroCopyRun
()
override
;
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
// Note: Can only be used under thread_local semantics.
bool
ExpRunWithExternalStream
(
const
gpuStream_t
stream
);
#endif
///
/// \brief Create feed fetch variables
///
...
...
paddle/fluid/inference/api/paddle_inference_api.h
浏览文件 @
15ad7ee4
...
...
@@ -41,11 +41,27 @@ limitations under the License. */
/// \since 2.0.0-beta
///
// forward declation
using
cudaStream_t
=
struct
CUstream_st
*
;
using
hipStream_t
=
struct
ihipStream_t
*
;
namespace
paddle_infer
{
using
PrecisionType
=
paddle
::
AnalysisConfig
::
Precision
;
using
Config
=
paddle
::
AnalysisConfig
;
class
Predictor
;
namespace
experimental
{
class
PD_INFER_DECL
InternalUtils
{
public:
// Note: Can only be used under thread_local semantics.
static
bool
RunWithExternalStream
(
paddle_infer
::
Predictor
*
pred
,
cudaStream_t
stream
);
static
bool
RunWithExternalStream
(
paddle_infer
::
Predictor
*
pred
,
hipStream_t
stream
);
};
}
// namespace experimental
///
/// \class Predictor
///
...
...
@@ -150,6 +166,7 @@ class PD_INFER_DECL Predictor {
private:
std
::
unique_ptr
<
paddle
::
PaddlePredictor
>
predictor_
;
friend
class
paddle_infer
::
experimental
::
InternalUtils
;
};
///
...
...
paddle/fluid/platform/device_context.cc
浏览文件 @
15ad7ee4
...
...
@@ -488,6 +488,26 @@ CUDAContext::CUDAContext(const CUDAPlace& place,
#endif
}
void
CUDAContext
::
SetStream
(
gpuStream_t
stream
)
{
if
(
stream_
->
raw_stream
()
!=
stream
)
{
CUDADeviceGuard
guard
(
place_
.
device
);
DestoryCuDNNContext
();
DestoryCuBlasContext
();
#ifndef PADDLE_WITH_HIP
DestoryCuSolverContext
();
#endif
stream_
->
SetStream
(
stream
);
InitEigenContext
();
InitCuBlasContext
();
InitCuDNNContext
();
#ifndef PADDLE_WITH_HIP
InitCuSolverContext
();
#endif
}
}
CUDAContext
::~
CUDAContext
()
{
CUDADeviceGuard
guard
(
place_
.
device
);
DestoryCuDNNContext
();
...
...
paddle/fluid/platform/device_context.h
浏览文件 @
15ad7ee4
...
...
@@ -334,6 +334,8 @@ class CUDAContext {
return
old_stream_ptr
;
}
void
SetStream
(
gpuStream_t
stream
);
const
gpuStream_t
&
RawStream
()
{
return
stream_
->
raw_stream
();
}
#ifdef PADDLE_WITH_HIP
...
...
@@ -616,6 +618,11 @@ class CUDADeviceContext : public DeviceContext {
return
thread_ctx_
.
at
(
this
);
}
// Note: Can only be used under thread_local semantics.
void
SetThreadLocalStream
(
const
gpuStream_t
stream
)
{
thread_ctx_
.
at
(
this
)
->
SetStream
(
stream
);
}
private:
CUDAPlace
place_
;
std
::
shared_ptr
<
CUDAContext
>
default_ctx_
;
...
...
paddle/fluid/platform/stream/cuda_stream.cc
浏览文件 @
15ad7ee4
...
...
@@ -56,7 +56,7 @@ void CUDAStream::Destroy() {
CUDADeviceGuard
guard
(
BOOST_GET_CONST
(
CUDAPlace
,
place_
).
device
);
Wait
();
WaitCallback
();
if
(
stream_
)
{
if
(
stream_
&&
owned_stream_
)
{
#ifdef PADDLE_WITH_HIP
PADDLE_ENFORCE_GPU_SUCCESS
(
hipStreamDestroy
(
stream_
));
#else
...
...
@@ -92,6 +92,20 @@ void CUDAStream::Wait() const {
PADDLE_ENFORCE_GPU_SUCCESS
(
e_sync
);
}
// Note: Can only be used under thread_local semantics.
void
CUDAStream
::
SetStream
(
gpuStream_t
stream
)
{
if
(
owned_stream_
&&
stream_
)
{
#ifdef PADDLE_WITH_HIP
PADDLE_ENFORCE_GPU_SUCCESS
(
hipStreamDestroy
(
stream_
));
#else
PADDLE_ENFORCE_GPU_SUCCESS
(
cudaStreamDestroy
(
stream_
));
#endif
}
owned_stream_
=
false
;
stream_
=
stream
;
callback_manager_
.
reset
(
new
StreamCallbackManager
<
gpuStream_t
>
(
stream_
));
}
CUDAStream
*
get_current_stream
(
int
deviceId
)
{
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
if
(
deviceId
==
-
1
)
{
...
...
paddle/fluid/platform/stream/cuda_stream.h
浏览文件 @
15ad7ee4
...
...
@@ -18,6 +18,7 @@ limitations under the License. */
#include <memory>
#include "paddle/fluid/platform/device/gpu/gpu_info.h"
#include "paddle/fluid/platform/device/gpu/gpu_types.h"
#include "paddle/fluid/platform/macros.h"
#include "paddle/fluid/platform/place.h"
#include "paddle/fluid/platform/stream_callback_manager.h"
...
...
@@ -130,8 +131,12 @@ class CUDAStream final {
const
Place
&
GetPlace
()
const
{
return
place_
;
}
// Note: Can only be used under thread_local semantics.
void
SetStream
(
gpuStream_t
stream
);
private:
Place
place_
;
bool
owned_stream_
{
true
};
#ifdef PADDLE_WITH_HIP
hipStream_t
stream_
{
nullptr
};
#else
...
...
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