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
飞不过沧海
Paddle
提交
1cb883da
P
Paddle
项目概览
飞不过沧海
/
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
体验新版 GitCode,发现更多精彩内容 >>
未验证
提交
1cb883da
编写于
8月 12, 2022
作者:
Z
zyfncg
提交者:
GitHub
8月 12, 2022
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
fix extra output of kernels for inference (#45048)
上级
615b15a3
变更
6
隐藏空白更改
内联
并排
Showing
6 changed file
with
61 addition
and
34 deletion
+61
-34
paddle/fluid/operators/dropout_impl.cu.h
paddle/fluid/operators/dropout_impl.cu.h
+1
-1
paddle/phi/infermeta/ternary.cc
paddle/phi/infermeta/ternary.cc
+6
-12
paddle/phi/kernels/cpu/dropout_kernel.cc
paddle/phi/kernels/cpu/dropout_kernel.cc
+2
-2
paddle/phi/kernels/cpu/instance_norm_kernel.cc
paddle/phi/kernels/cpu/instance_norm_kernel.cc
+18
-6
paddle/phi/kernels/gpu/dropout_kernel.cu
paddle/phi/kernels/gpu/dropout_kernel.cu
+7
-3
paddle/phi/kernels/gpu/instance_norm_kernel.cu
paddle/phi/kernels/gpu/instance_norm_kernel.cu
+27
-10
未找到文件。
paddle/fluid/operators/dropout_impl.cu.h
浏览文件 @
1cb883da
...
...
@@ -307,7 +307,7 @@ void DropoutFwGPUKernelDriver(const phi::GPUContext& dev_ctx,
auto
*
x_data
=
x
.
data
<
T
>
();
auto
*
y_data
=
y
->
data
<
T
>
();
if
(
!
is_test
)
{
if
(
!
is_test
&&
mask
)
{
auto
*
mask_data
=
mask
->
data
<
uint8_t
>
();
size_t
size
=
phi
::
product
(
mask
->
dims
());
...
...
paddle/phi/infermeta/ternary.cc
浏览文件 @
1cb883da
...
...
@@ -317,16 +317,6 @@ void InstanceNormInferMeta(const MetaTensor& x,
nullptr
,
phi
::
errors
::
InvalidArgument
(
"The y in InstanceNormInferMeta can't be nullptr."
));
PADDLE_ENFORCE_NE
(
saved_mean
,
nullptr
,
phi
::
errors
::
InvalidArgument
(
"The saved_mean in InstanceNormInferMeta can't be nullptr."
));
PADDLE_ENFORCE_NE
(
saved_variance
,
nullptr
,
phi
::
errors
::
InvalidArgument
(
"The saved_variance in InstanceNormInferMeta can't be nullptr."
));
const
auto
x_dims
=
x
.
dims
();
PADDLE_ENFORCE_NE
(
phi
::
product
(
x_dims
),
0
,
...
...
@@ -401,11 +391,15 @@ void InstanceNormInferMeta(const MetaTensor& x,
}
}
y
->
set_dims
(
x_dims
);
saved_mean
->
set_dims
({
NxC
});
saved_variance
->
set_dims
({
NxC
});
y
->
share_lod
(
x
);
y
->
set_dtype
(
x
.
dtype
());
y
->
set_layout
(
x
.
layout
());
if
(
saved_mean
)
{
saved_mean
->
set_dims
({
NxC
});
}
if
(
saved_variance
)
{
saved_variance
->
set_dims
({
NxC
});
}
}
void
GraphSendRecvInferMeta
(
const
MetaTensor
&
x
,
...
...
paddle/phi/kernels/cpu/dropout_kernel.cc
浏览文件 @
1cb883da
...
...
@@ -63,7 +63,7 @@ void DropoutRawKernel(const Context& dev_ctx,
auto
&
dropout_implementation
=
mode
;
bool
upscale_in_train
=
(
dropout_implementation
==
"upscale_in_train"
);
if
(
!
is_test
)
{
if
(
!
is_test
&&
mask
)
{
auto
*
mask_data
=
dev_ctx
.
template
Alloc
<
uint8_t
>(
mask
);
size_t
size
=
phi
::
product
(
mask
->
dims
());
...
...
@@ -124,7 +124,7 @@ void DropoutNdKernel(const Context& dev_ctx,
auto
&
dropout_implementation
=
mode
;
bool
upscale_in_train
=
(
dropout_implementation
==
"upscale_in_train"
);
if
(
!
is_test
)
{
if
(
!
is_test
&&
mask
)
{
DenseTensor
t_mask
;
t_mask
.
Resize
(
mask
->
dims
());
T
*
t_mask_data
=
dev_ctx
.
template
Alloc
<
T
>(
&
t_mask
);
...
...
paddle/phi/kernels/cpu/instance_norm_kernel.cc
浏览文件 @
1cb883da
...
...
@@ -21,6 +21,7 @@
#include "paddle/phi/backends/cpu/cpu_context.h"
#include "paddle/phi/common/layout.h"
#include "paddle/phi/core/kernel_registry.h"
#include "paddle/phi/kernels/full_kernel.h"
#include "paddle/phi/kernels/funcs/eigen/common.h"
#include "paddle/phi/kernels/funcs/eigen/eigen_function.h"
#include "paddle/phi/kernels/funcs/eigen/extensions.h"
...
...
@@ -63,14 +64,25 @@ void InstanceNormKernel(const Context& dev_ctx,
#endif
phi
::
funcs
::
SetConstant
<
CPUContext
,
T
>
set_constant
;
dev_ctx
.
template
Alloc
<
T
>(
saved_mean
);
dev_ctx
.
template
Alloc
<
T
>(
saved_variance
);
set_constant
(
dev_ctx
,
saved_mean
,
static_cast
<
T
>
(
0
));
set_constant
(
dev_ctx
,
saved_variance
,
static_cast
<
T
>
(
0
));
DenseTensor
saved_mean_tmp
,
saved_variance_tmp
;
if
(
saved_mean
)
{
dev_ctx
.
template
Alloc
<
T
>(
saved_mean
);
set_constant
(
dev_ctx
,
saved_mean
,
static_cast
<
T
>
(
0
));
}
else
{
saved_mean_tmp
=
phi
::
Full
<
T
>
(
dev_ctx
,
{
NxC
},
0
);
}
if
(
saved_variance
)
{
dev_ctx
.
template
Alloc
<
T
>(
saved_variance
);
set_constant
(
dev_ctx
,
saved_variance
,
static_cast
<
T
>
(
0
));
}
else
{
saved_variance_tmp
=
phi
::
Full
<
T
>
(
dev_ctx
,
{
NxC
},
0
);
}
auto
saved_mean_a
=
EigenVector
<
T
>::
Flatten
(
*
saved_mean
);
auto
saved_mean_a
=
EigenVector
<
T
>::
Flatten
(
saved_mean
?
*
saved_mean
:
saved_mean_tmp
);
auto
saved_mean_e
=
saved_mean_a
.
reshape
(
NxC_shape
);
auto
saved_variance_a
=
EigenVector
<
T
>::
Flatten
(
*
saved_variance
);
auto
saved_variance_a
=
EigenVector
<
T
>::
Flatten
(
saved_variance
?
*
saved_variance
:
saved_variance_tmp
);
auto
saved_variance_e
=
saved_variance_a
.
reshape
(
NxC_shape
);
auto
x_e
=
EigenVector
<
T
>::
Flatten
(
x
);
...
...
paddle/phi/kernels/gpu/dropout_kernel.cu
浏览文件 @
1cb883da
...
...
@@ -32,8 +32,10 @@ void DropoutRawKernel(const Context& dev_ctx,
DenseTensor
*
out
,
DenseTensor
*
mask
)
{
bool
upscale_in_train
=
(
mode
==
"upscale_in_train"
);
out
->
mutable_data
<
T
>
(
dev_ctx
.
GetPlace
());
mask
->
mutable_data
<
uint8_t
>
(
dev_ctx
.
GetPlace
());
dev_ctx
.
template
Alloc
<
T
>(
out
);
if
(
mask
)
{
dev_ctx
.
template
Alloc
<
uint8_t
>(
mask
);
}
paddle
::
operators
::
DropoutFwGPUKernelDriver
<
T
>
(
dev_ctx
,
is_test
,
p
.
to
<
float
>
(),
...
...
@@ -61,7 +63,9 @@ void DropoutNdKernel(const Context& dev_ctx,
DenseTensor
*
mask
)
{
bool
upscale_in_train
=
(
mode
==
"upscale_in_train"
);
dev_ctx
.
template
Alloc
<
T
>(
out
);
dev_ctx
.
template
Alloc
<
uint8_t
>(
mask
);
if
(
mask
)
{
dev_ctx
.
template
Alloc
<
uint8_t
>(
mask
);
}
paddle
::
operators
::
DropoutFwGPUKernelDriver
<
T
>
(
dev_ctx
,
is_test
,
p
.
to
<
float
>
(),
...
...
paddle/phi/kernels/gpu/instance_norm_kernel.cu
浏览文件 @
1cb883da
...
...
@@ -17,6 +17,7 @@
#include "paddle/phi/backends/gpu/gpu_context.h"
#include "paddle/phi/common/layout.h"
#include "paddle/phi/core/kernel_registry.h"
#include "paddle/phi/kernels/full_kernel.h"
#include "paddle/phi/kernels/funcs/math_function.h"
#include "paddle/phi/kernels/funcs/norm_utils.h"
#include "paddle/phi/kernels/gpu/instance_norm_utils.h"
...
...
@@ -143,11 +144,29 @@ void InstanceNormKernel(const Context &dev_ctx,
auto
handle
=
dev_ctx
.
cudnn_handle
();
DenseTensor
saved_mean_tmp
,
saved_variance_tmp
;
phi
::
funcs
::
SetConstant
<
GPUContext
,
BatchNormParamType
<
T
>>
functor
;
dev_ctx
.
template
Alloc
<
BatchNormParamType
<
T
>
>
(
saved_mean
);
dev_ctx
.
template
Alloc
<
BatchNormParamType
<
T
>
>
(
saved_variance
);
functor
(
dev_ctx
,
saved_mean
,
static_cast
<
BatchNormParamType
<
T
>>
(
0
));
functor
(
dev_ctx
,
saved_variance
,
static_cast
<
BatchNormParamType
<
T
>>
(
0
));
if
(
saved_mean
)
{
dev_ctx
.
template
Alloc
<
BatchNormParamType
<
T
>
>
(
saved_mean
);
functor
(
dev_ctx
,
saved_mean
,
static_cast
<
BatchNormParamType
<
T
>>
(
0
));
}
else
{
saved_mean_tmp
=
phi
::
Full
<
BatchNormParamType
<
T
>>
(
dev_ctx
,
{
NxC
},
static_cast
<
BatchNormParamType
<
T
>>
(
0
));
}
if
(
saved_variance
)
{
dev_ctx
.
template
Alloc
<
BatchNormParamType
<
T
>
>
(
saved_variance
);
functor
(
dev_ctx
,
saved_variance
,
static_cast
<
BatchNormParamType
<
T
>>
(
0
));
}
else
{
saved_variance_tmp
=
phi
::
Full
<
BatchNormParamType
<
T
>>
(
dev_ctx
,
{
NxC
},
static_cast
<
BatchNormParamType
<
T
>>
(
0
));
}
auto
*
saved_mean_data
=
saved_mean
?
saved_mean
->
data
<
BatchNormParamType
<
T
>>
()
:
saved_mean_tmp
.
data
<
BatchNormParamType
<
T
>>
();
auto
*
saved_variance_data
=
saved_variance
?
saved_variance
->
data
<
BatchNormParamType
<
T
>>
()
:
saved_variance_tmp
.
data
<
BatchNormParamType
<
T
>>
();
#ifdef PADDLE_WITH_HIP
PADDLE_ENFORCE_GPU_SUCCESS
(
...
...
@@ -171,10 +190,8 @@ void InstanceNormKernel(const Context &dev_ctx,
nullptr
,
nullptr
,
epsilon
,
static_cast
<
void
*>
(
saved_mean
->
template
data
<
BatchNormParamType
<
T
>
>
()),
static_cast
<
void
*>
(
saved_variance
->
template
data
<
BatchNormParamType
<
T
>
>
())));
static_cast
<
void
*>
(
saved_mean_data
),
static_cast
<
void
*>
(
saved_variance_data
)));
PADDLE_ENFORCE_GPU_SUCCESS
(
paddle
::
platform
::
dynload
::
miopenDestroyTensorDescriptor
(
data_desc_
));
...
...
@@ -198,8 +215,8 @@ void InstanceNormKernel(const Context &dev_ctx,
nullptr
,
nullptr
,
epsilon
,
saved_mean
->
template
data
<
BatchNormParamType
<
T
>
>
()
,
saved_variance
->
template
data
<
BatchNormParamType
<
T
>
>
()
));
saved_mean
_data
,
saved_variance
_data
));
PADDLE_ENFORCE_GPU_SUCCESS
(
paddle
::
platform
::
dynload
::
cudnnDestroyTensorDescriptor
(
data_desc_
));
...
...
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