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
afd5a96b
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
afd5a96b
编写于
10月 25, 2022
作者:
H
HongyuJia
提交者:
GitHub
10月 25, 2022
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
opt conv_transpose cudnn (#47294)
上级
3e7abca5
变更
3
显示空白变更内容
内联
并排
Showing
3 changed file
with
24 addition
and
49 deletion
+24
-49
paddle/fluid/operators/conv_transpose_op.cc
paddle/fluid/operators/conv_transpose_op.cc
+22
-47
paddle/fluid/platform/device/gpu/cuda/cudnn_helper.h
paddle/fluid/platform/device/gpu/cuda/cudnn_helper.h
+1
-1
paddle/fluid/platform/device/gpu/rocm/miopen_helper.h
paddle/fluid/platform/device/gpu/rocm/miopen_helper.h
+1
-1
未找到文件。
paddle/fluid/operators/conv_transpose_op.cc
浏览文件 @
afd5a96b
...
@@ -28,6 +28,9 @@ limitations under the License. */
...
@@ -28,6 +28,9 @@ limitations under the License. */
#ifdef PADDLE_WITH_MKLDNN
#ifdef PADDLE_WITH_MKLDNN
#include "paddle/fluid/platform/mkldnn_helper.h"
#include "paddle/fluid/platform/mkldnn_helper.h"
#endif
#endif
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
#include "paddle/fluid/platform/device/gpu/gpu_dnn.h"
#endif
namespace
paddle
{
namespace
paddle
{
namespace
operators
{
namespace
operators
{
...
@@ -38,16 +41,12 @@ framework::OpKernelType ConvTransposeOp::GetExpectedKernelType(
...
@@ -38,16 +41,12 @@ framework::OpKernelType ConvTransposeOp::GetExpectedKernelType(
const
framework
::
ExecutionContext
&
ctx
)
const
{
const
framework
::
ExecutionContext
&
ctx
)
const
{
auto
data_type
=
OperatorWithKernel
::
IndicateVarDataType
(
ctx
,
"Input"
);
auto
data_type
=
OperatorWithKernel
::
IndicateVarDataType
(
ctx
,
"Input"
);
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
if
(
platform
::
is_gpu_place
(
ctx
.
GetPlace
()))
{
if
(
platform
::
CanCUDNNBeUsed
(
ctx
))
{
auto
&
dev_ctx
=
ctx
.
template
device_context
<
phi
::
GPUContext
>();
if
(
ctx
.
HasAttr
(
"use_cudnn"
)
&&
ctx
.
Attr
<
bool
>
(
"use_cudnn"
)
&&
dev_ctx
.
cudnn_handle
()
!=
nullptr
)
{
return
framework
::
OpKernelType
(
data_type
,
return
framework
::
OpKernelType
(
data_type
,
ctx
.
GetPlace
(),
ctx
.
GetPlace
(),
phi
::
DataLayout
::
kAnyLayout
,
phi
::
DataLayout
::
kAnyLayout
,
framework
::
LibraryType
::
kCUDNN
);
framework
::
LibraryType
::
kCUDNN
);
}
}
}
#endif
#endif
return
framework
::
OpKernelType
(
data_type
,
ctx
.
GetPlace
());
return
framework
::
OpKernelType
(
data_type
,
ctx
.
GetPlace
());
}
}
...
@@ -268,28 +267,16 @@ Example:
...
@@ -268,28 +267,16 @@ Example:
framework
::
OpKernelType
ConvTransposeOpGrad
::
GetExpectedKernelType
(
framework
::
OpKernelType
ConvTransposeOpGrad
::
GetExpectedKernelType
(
const
framework
::
ExecutionContext
&
ctx
)
const
{
const
framework
::
ExecutionContext
&
ctx
)
const
{
bool
use_cudnn
=
auto
data_type
=
OperatorWithKernel
::
IndicateVarDataType
(
ctx
,
"Input"
);
ctx
.
HasAttr
(
"use_cudnn"
)
?
ctx
.
Attr
<
bool
>
(
"use_cudnn"
)
:
false
;
use_cudnn
&=
platform
::
is_gpu_place
(
ctx
.
GetPlace
());
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
if
(
platform
::
is_gpu_place
(
ctx
.
GetPlace
()))
{
if
(
platform
::
CanCUDNNBeUsed
(
ctx
))
{
auto
&
dev_ctx
=
ctx
.
template
device_context
<
phi
::
GPUContext
>();
return
framework
::
OpKernelType
(
data_type
,
use_cudnn
&=
dev_ctx
.
cudnn_handle
()
!=
nullptr
;
ctx
.
GetPlace
(),
phi
::
DataLayout
::
kAnyLayout
,
framework
::
LibraryType
::
kCUDNN
);
}
}
#endif
#endif
framework
::
LibraryType
library_
;
return
framework
::
OpKernelType
(
data_type
,
ctx
.
GetPlace
());
if
(
use_cudnn
)
{
library_
=
framework
::
LibraryType
::
kCUDNN
;
}
else
{
library_
=
framework
::
LibraryType
::
kPlain
;
}
phi
::
DataLayout
layout_
=
phi
::
DataLayout
::
kAnyLayout
;
return
framework
::
OpKernelType
(
OperatorWithKernel
::
IndicateVarDataType
(
ctx
,
"Input"
),
ctx
.
GetPlace
(),
layout_
,
library_
);
}
}
template
<
typename
T
>
template
<
typename
T
>
...
@@ -355,28 +342,16 @@ class ConvTransposeDoubleGradMaker : public framework::SingleGradOpMaker<T> {
...
@@ -355,28 +342,16 @@ class ConvTransposeDoubleGradMaker : public framework::SingleGradOpMaker<T> {
framework
::
OpKernelType
ConvTransposeOpDoubleGrad
::
GetExpectedKernelType
(
framework
::
OpKernelType
ConvTransposeOpDoubleGrad
::
GetExpectedKernelType
(
const
framework
::
ExecutionContext
&
ctx
)
const
{
const
framework
::
ExecutionContext
&
ctx
)
const
{
bool
use_cudnn
=
auto
data_type
=
OperatorWithKernel
::
IndicateVarDataType
(
ctx
,
"Input"
);
ctx
.
HasAttr
(
"use_cudnn"
)
?
ctx
.
Attr
<
bool
>
(
"use_cudnn"
)
:
false
;
use_cudnn
&=
platform
::
is_gpu_place
(
ctx
.
GetPlace
());
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
if
(
platform
::
is_gpu_place
(
ctx
.
GetPlace
()))
{
if
(
platform
::
CanCUDNNBeUsed
(
ctx
))
{
auto
&
dev_ctx
=
ctx
.
template
device_context
<
phi
::
GPUContext
>();
return
framework
::
OpKernelType
(
data_type
,
use_cudnn
&=
dev_ctx
.
cudnn_handle
()
!=
nullptr
;
ctx
.
GetPlace
(),
phi
::
DataLayout
::
kAnyLayout
,
framework
::
LibraryType
::
kCUDNN
);
}
}
#endif
#endif
framework
::
LibraryType
library_
;
return
framework
::
OpKernelType
(
data_type
,
ctx
.
GetPlace
());
if
(
use_cudnn
)
{
library_
=
framework
::
LibraryType
::
kCUDNN
;
}
else
{
library_
=
framework
::
LibraryType
::
kPlain
;
}
phi
::
DataLayout
layout_
=
phi
::
DataLayout
::
kAnyLayout
;
return
framework
::
OpKernelType
(
OperatorWithKernel
::
IndicateVarDataType
(
ctx
,
"Input"
),
ctx
.
GetPlace
(),
layout_
,
library_
);
}
}
}
// namespace operators
}
// namespace operators
...
...
paddle/fluid/platform/device/gpu/cuda/cudnn_helper.h
浏览文件 @
afd5a96b
...
@@ -617,7 +617,7 @@ class ScopedActivationDescriptor {
...
@@ -617,7 +617,7 @@ class ScopedActivationDescriptor {
};
};
inline
bool
CanCUDNNBeUsed
(
const
framework
::
ExecutionContext
&
ctx
)
{
inline
bool
CanCUDNNBeUsed
(
const
framework
::
ExecutionContext
&
ctx
)
{
bool
use_cudnn
=
ctx
.
Attr
<
bool
>
(
"use_cudnn"
);
bool
use_cudnn
=
ctx
.
HasAttr
(
"use_cudnn"
)
&&
ctx
.
Attr
<
bool
>
(
"use_cudnn"
);
use_cudnn
&=
paddle
::
platform
::
is_gpu_place
(
ctx
.
GetPlace
());
use_cudnn
&=
paddle
::
platform
::
is_gpu_place
(
ctx
.
GetPlace
());
#ifdef PADDLE_WITH_CUDA
#ifdef PADDLE_WITH_CUDA
if
(
use_cudnn
)
{
if
(
use_cudnn
)
{
...
...
paddle/fluid/platform/device/gpu/rocm/miopen_helper.h
浏览文件 @
afd5a96b
...
@@ -554,7 +554,7 @@ class ScopedActivationDescriptor {
...
@@ -554,7 +554,7 @@ class ScopedActivationDescriptor {
};
};
inline
bool
CanCUDNNBeUsed
(
const
framework
::
ExecutionContext
&
ctx
)
{
inline
bool
CanCUDNNBeUsed
(
const
framework
::
ExecutionContext
&
ctx
)
{
bool
use_cudnn
=
ctx
.
Attr
<
bool
>
(
"use_cudnn"
);
bool
use_cudnn
=
ctx
.
HasAttr
(
"use_cudnn"
)
&&
ctx
.
Attr
<
bool
>
(
"use_cudnn"
);
use_cudnn
&=
paddle
::
platform
::
is_gpu_place
(
ctx
.
GetPlace
());
use_cudnn
&=
paddle
::
platform
::
is_gpu_place
(
ctx
.
GetPlace
());
#ifdef PADDLE_WITH_HIP
#ifdef PADDLE_WITH_HIP
if
(
use_cudnn
)
{
if
(
use_cudnn
)
{
...
...
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