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
1f1a7835
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
1f1a7835
编写于
8月 26, 2022
作者:
R
Ruibiao Chen
提交者:
GitHub
8月 26, 2022
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Move conv2d_transpose XPU kernel to PHI, test=kunlun (#45419)
上级
6ab80b64
变更
2
隐藏空白更改
内联
并排
Showing
2 changed file
with
126 addition
and
103 deletion
+126
-103
paddle/fluid/operators/conv_transpose_op_xpu.cc
paddle/fluid/operators/conv_transpose_op_xpu.cc
+0
-103
paddle/phi/kernels/xpu/conv_transpose_kernel.cc
paddle/phi/kernels/xpu/conv_transpose_kernel.cc
+126
-0
未找到文件。
paddle/fluid/operators/conv_transpose_op_xpu.cc
浏览文件 @
1f1a7835
...
...
@@ -24,106 +24,6 @@ namespace operators {
using
Tensor
=
framework
::
Tensor
;
// target_len == 2 || target_len == 4
inline
std
::
vector
<
int
>
vector_extend
(
const
std
::
vector
<
int
>&
src
,
int
target_len
)
{
if
(
target_len
==
2
&&
src
.
size
()
==
1
)
{
return
{
src
[
0
],
src
[
0
]};
}
if
(
target_len
==
4
&&
src
.
size
()
==
1
)
{
return
{
src
[
0
],
src
[
0
],
src
[
0
],
src
[
0
]};
}
if
(
target_len
==
4
&&
src
.
size
()
==
2
)
{
return
{
src
[
0
],
src
[
0
],
src
[
1
],
src
[
1
]};
}
return
src
;
}
template
<
typename
DeviceContext
,
typename
T
>
class
Conv2DTransposeXPUKernel
:
public
framework
::
OpKernel
<
T
>
{
public:
void
Compute
(
const
framework
::
ExecutionContext
&
context
)
const
override
{
const
Tensor
*
input
=
context
.
Input
<
Tensor
>
(
"Input"
);
// The filter will be reshaped in the calculations,
// so here use an assignment operation,
// that avoids modifying the variable in the Scope.
Tensor
filter
=
*
context
.
Input
<
Tensor
>
(
"Filter"
);
Tensor
*
output
=
context
.
Output
<
Tensor
>
(
"Output"
);
output
->
mutable_data
<
T
>
(
context
.
GetPlace
());
int
groups
=
context
.
Attr
<
int
>
(
"groups"
);
std
::
vector
<
int
>
strides
=
context
.
Attr
<
std
::
vector
<
int
>>
(
"strides"
);
std
::
vector
<
int
>
paddings
=
context
.
Attr
<
std
::
vector
<
int
>>
(
"paddings"
);
std
::
vector
<
int
>
dilations
=
context
.
Attr
<
std
::
vector
<
int
>>
(
"dilations"
);
const
std
::
string
data_format
=
context
.
Attr
<
std
::
string
>
(
"data_format"
);
const
std
::
string
padding_algorithm
=
context
.
Attr
<
std
::
string
>
(
"padding_algorithm"
);
PADDLE_ENFORCE_EQ
(
data_format
==
"NHWC"
||
data_format
==
"NDHWC"
,
false
,
platform
::
errors
::
InvalidArgument
(
(
"XPU do support data_format is NCHW in conv_transpose op."
)));
framework
::
DDim
in_data_dims
=
phi
::
slice_ddim
(
input
->
dims
(),
2
,
input
->
dims
().
size
());
framework
::
DDim
filter_data_dims
=
phi
::
slice_ddim
(
filter
.
dims
(),
2
,
filter
.
dims
().
size
());
std
::
vector
<
int
>
ksize
=
phi
::
vectorize
<
int
>
(
filter_data_dims
);
phi
::
UpdatePaddingAndDilation
(
&
paddings
,
&
dilations
,
padding_algorithm
,
in_data_dims
,
strides
,
ksize
);
const
int
batch_size
=
static_cast
<
int
>
(
input
->
dims
()[
0
]);
const
int
img_yc
=
static_cast
<
int
>
(
input
->
dims
()[
1
]);
const
int
img_yh
=
static_cast
<
int
>
(
input
->
dims
()[
2
]);
const
int
img_yw
=
static_cast
<
int
>
(
input
->
dims
()[
3
]);
const
int
img_xc
=
static_cast
<
int
>
(
output
->
dims
()[
1
]);
const
int
img_xh
=
static_cast
<
int
>
(
output
->
dims
()[
2
]);
const
int
img_xw
=
static_cast
<
int
>
(
output
->
dims
()[
3
]);
{
std
::
vector
<
int
>
ksize_check
=
vector_extend
(
ksize
,
2
);
std
::
vector
<
int
>
stride_check
=
vector_extend
(
strides
,
2
);
std
::
vector
<
int
>
pad_check
=
vector_extend
(
paddings
,
4
);
std
::
vector
<
int
>
dilation_check
=
vector_extend
(
dilations
,
2
);
int
xh_check
=
(
img_yh
-
1
)
*
stride_check
[
0
]
-
pad_check
[
0
]
-
pad_check
[
1
]
+
(
dilation_check
[
0
]
*
(
ksize_check
[
0
]
-
1
)
+
1
);
int
xw_check
=
(
img_yw
-
1
)
*
stride_check
[
1
]
-
pad_check
[
2
]
-
pad_check
[
3
]
+
(
dilation_check
[
1
]
*
(
ksize_check
[
1
]
-
1
)
+
1
);
PADDLE_ENFORCE_EQ
(
xh_check
==
img_xh
&&
xw_check
==
img_xw
,
true
,
platform
::
errors
::
InvalidArgument
(
(
"XPU output size check error in conv_transpose op."
)));
}
auto
&
dev_ctx
=
context
.
template
device_context
<
DeviceContext
>();
int
r
=
xpu
::
conv2d_transpose
<
float
,
float
,
float
,
int16_t
>
(
dev_ctx
.
x_context
(),
input
->
data
<
float
>
(),
filter
.
data
<
float
>
(),
output
->
data
<
float
>
(),
batch_size
,
img_yc
,
img_yh
,
img_yw
,
img_xc
,
ksize
,
strides
,
paddings
,
dilations
,
groups
,
nullptr
,
nullptr
,
nullptr
,
true
);
PADDLE_ENFORCE_XDNN_SUCCESS
(
r
,
"conv2d_transpose"
);
}
};
template
<
typename
DeviceContext
,
typename
T
>
class
Conv2DTransposeGradXPUKernel
:
public
framework
::
OpKernel
<
T
>
{
public:
...
...
@@ -209,9 +109,6 @@ class Conv2DTransposeGradXPUKernel : public framework::OpKernel<T> {
}
// namespace operators
}
// namespace paddle
namespace
ops
=
paddle
::
operators
;
REGISTER_OP_XPU_KERNEL
(
conv2d_transpose
,
ops
::
Conv2DTransposeXPUKernel
<
paddle
::
platform
::
XPUDeviceContext
,
float
>
);
REGISTER_OP_XPU_KERNEL
(
conv2d_transpose_grad
,
ops
::
Conv2DTransposeGradXPUKernel
<
paddle
::
platform
::
XPUDeviceContext
,
...
...
paddle/phi/kernels/xpu/conv_transpose_kernel.cc
0 → 100644
浏览文件 @
1f1a7835
// Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "paddle/phi/kernels/conv_transpose_kernel.h"
#include "paddle/phi/backends/xpu/enforce_xpu.h"
#include "paddle/phi/core/kernel_registry.h"
#include "paddle/phi/kernels/cpu/conv_util.h"
namespace
phi
{
// target_len == 2 || target_len == 4
inline
std
::
vector
<
int
>
vector_extend
(
const
std
::
vector
<
int
>&
src
,
int
target_len
)
{
if
(
target_len
==
2
&&
src
.
size
()
==
1
)
{
return
{
src
[
0
],
src
[
0
]};
}
if
(
target_len
==
4
&&
src
.
size
()
==
1
)
{
return
{
src
[
0
],
src
[
0
],
src
[
0
],
src
[
0
]};
}
if
(
target_len
==
4
&&
src
.
size
()
==
2
)
{
return
{
src
[
0
],
src
[
0
],
src
[
1
],
src
[
1
]};
}
return
src
;
}
template
<
typename
T
,
typename
Context
>
void
Conv2dTransposeKernel
(
const
Context
&
ctx
,
const
DenseTensor
&
x
,
const
DenseTensor
&
filter
,
const
std
::
vector
<
int
>&
strides
,
const
std
::
vector
<
int
>&
paddings
,
const
std
::
vector
<
int
>&
output_padding
,
const
std
::
vector
<
int
>&
output_size
,
const
std
::
string
&
padding_algorithm
,
int
groups
,
const
std
::
vector
<
int
>&
dilations
,
const
std
::
string
&
data_format
,
DenseTensor
*
out
)
{
// The filter will be reshaped in the calculations,
// so here use an assignment operation,
// that avoids modifying the variable in the Scope.
DenseTensor
filter_
=
filter
;
ctx
.
template
Alloc
<
T
>(
out
);
PADDLE_ENFORCE_EQ
(
data_format
==
"NHWC"
||
data_format
==
"NDHWC"
,
false
,
errors
::
InvalidArgument
(
(
"XPU do support data_format is NCHW in conv_transpose op."
)));
DDim
in_data_dims
=
slice_ddim
(
x
.
dims
(),
2
,
x
.
dims
().
size
());
DDim
filter_data_dims
=
slice_ddim
(
filter_
.
dims
(),
2
,
filter_
.
dims
().
size
());
std
::
vector
<
int
>
ksize
=
vectorize
<
int
>
(
filter_data_dims
);
std
::
vector
<
int
>
paddings_
=
paddings
;
std
::
vector
<
int
>
dilations_
=
dilations
;
UpdatePaddingAndDilation
(
&
paddings_
,
&
dilations_
,
padding_algorithm
,
in_data_dims
,
strides
,
ksize
);
const
int
batch_size
=
static_cast
<
int
>
(
x
.
dims
()[
0
]);
const
int
img_yc
=
static_cast
<
int
>
(
x
.
dims
()[
1
]);
const
int
img_yh
=
static_cast
<
int
>
(
x
.
dims
()[
2
]);
const
int
img_yw
=
static_cast
<
int
>
(
x
.
dims
()[
3
]);
const
int
img_xc
=
static_cast
<
int
>
(
out
->
dims
()[
1
]);
const
int
img_xh
=
static_cast
<
int
>
(
out
->
dims
()[
2
]);
const
int
img_xw
=
static_cast
<
int
>
(
out
->
dims
()[
3
]);
{
std
::
vector
<
int
>
ksize_check
=
vector_extend
(
ksize
,
2
);
std
::
vector
<
int
>
stride_check
=
vector_extend
(
strides
,
2
);
std
::
vector
<
int
>
pad_check
=
vector_extend
(
paddings_
,
4
);
std
::
vector
<
int
>
dilation_check
=
vector_extend
(
dilations_
,
2
);
int
xh_check
=
(
img_yh
-
1
)
*
stride_check
[
0
]
-
pad_check
[
0
]
-
pad_check
[
1
]
+
(
dilation_check
[
0
]
*
(
ksize_check
[
0
]
-
1
)
+
1
);
int
xw_check
=
(
img_yw
-
1
)
*
stride_check
[
1
]
-
pad_check
[
2
]
-
pad_check
[
3
]
+
(
dilation_check
[
1
]
*
(
ksize_check
[
1
]
-
1
)
+
1
);
PADDLE_ENFORCE_EQ
(
xh_check
==
img_xh
&&
xw_check
==
img_xw
,
true
,
errors
::
InvalidArgument
(
(
"XPU output size check error in conv_transpose op."
)));
}
int
r
=
xpu
::
conv2d_transpose
<
float
,
float
,
float
,
int16_t
>
(
ctx
.
x_context
(),
x
.
data
<
float
>
(),
filter_
.
data
<
float
>
(),
out
->
data
<
float
>
(),
batch_size
,
img_yc
,
img_yh
,
img_yw
,
img_xc
,
ksize
,
strides
,
paddings_
,
dilations_
,
groups
,
nullptr
,
nullptr
,
nullptr
,
true
);
PADDLE_ENFORCE_XDNN_SUCCESS
(
r
,
"conv2d_transpose"
);
}
}
// namespace phi
PD_REGISTER_KERNEL
(
conv2d_transpose
,
XPU
,
ALL_LAYOUT
,
phi
::
Conv2dTransposeKernel
,
float
)
{}
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