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
csdn_franckjun
Paddle
提交
84b72c5f
P
Paddle
项目概览
csdn_franckjun
/
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
1
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
1
Issue
1
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
84b72c5f
编写于
7月 14, 2022
作者:
Y
ykkk2333
提交者:
GitHub
7月 14, 2022
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
add xpu pnorm op and fix pool op, *test=kunlun (#44214)
上级
270ba570
变更
5
隐藏空白更改
内联
并排
Showing
5 changed file
with
654 addition
and
0 deletion
+654
-0
paddle/fluid/operators/p_norm_op_xpu.cc
paddle/fluid/operators/p_norm_op_xpu.cc
+354
-0
paddle/fluid/operators/pool_op_xpu.cc
paddle/fluid/operators/pool_op_xpu.cc
+40
-0
paddle/fluid/platform/device/xpu/xpu2_op_list.h
paddle/fluid/platform/device/xpu/xpu2_op_list.h
+2
-0
python/paddle/fluid/tests/unittests/xpu/test_p_norm_op_xpu.py
...on/paddle/fluid/tests/unittests/xpu/test_p_norm_op_xpu.py
+186
-0
python/paddle/fluid/tests/unittests/xpu/test_pool2d_op_xpu.py
...on/paddle/fluid/tests/unittests/xpu/test_pool2d_op_xpu.py
+72
-0
未找到文件。
paddle/fluid/operators/p_norm_op_xpu.cc
0 → 100644
浏览文件 @
84b72c5f
/* 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. */
#ifdef PADDLE_WITH_XPU
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/operators/reduce_ops/reduce_op_xpu.h"
#include "paddle/fluid/platform/device/device_wrapper.h"
#include "paddle/fluid/platform/device/xpu/xpu_header.h"
namespace
paddle
{
namespace
operators
{
inline
void
GetDims
(
const
phi
::
DDim
&
dim
,
int
axis
,
int
*
m
,
int
*
t
,
int
*
n
,
bool
asvector
)
{
*
m
=
1
;
*
n
=
1
;
*
t
=
dim
[
axis
];
if
(
asvector
)
{
*
t
=
product
(
dim
);
}
else
{
for
(
int
i
=
0
;
i
<
axis
;
++
i
)
{
(
*
m
)
*=
dim
[
i
];
}
for
(
int
i
=
axis
+
1
;
i
<
dim
.
size
();
++
i
)
{
(
*
n
)
*=
dim
[
i
];
}
}
}
using
Tensor
=
framework
::
Tensor
;
template
<
typename
DeviceContext
,
typename
T
>
class
P_NormXPUKernel
:
public
framework
::
OpKernel
<
T
>
{
using
XPUType
=
typename
XPUTypeTrait
<
T
>::
Type
;
public:
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
auto
*
in
=
ctx
.
Input
<
framework
::
Tensor
>
(
"X"
);
auto
*
out
=
ctx
.
Output
<
framework
::
Tensor
>
(
"Out"
);
out
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
float
porder
=
ctx
.
Attr
<
float
>
(
"porder"
);
int
axis
=
ctx
.
Attr
<
int
>
(
"axis"
);
bool
asvector
=
ctx
.
Attr
<
bool
>
(
"asvector"
);
auto
&
dev_ctx
=
ctx
.
template
device_context
<
DeviceContext
>();
auto
xdim
=
in
->
dims
();
if
(
axis
<
0
)
axis
=
xdim
.
size
()
+
axis
;
std
::
vector
<
int
>
r_dim
;
std
::
vector
<
int
>
x_dim
;
std
::
vector
<
int
>
y_dim
;
int
m
=
1
;
int
n
=
1
;
int
t
=
1
;
GetDims
(
xdim
,
axis
,
&
m
,
&
t
,
&
n
,
asvector
);
x_dim
.
push_back
(
m
);
x_dim
.
push_back
(
t
);
x_dim
.
push_back
(
n
);
r_dim
.
push_back
(
1
);
y_dim
.
push_back
(
m
);
y_dim
.
push_back
(
n
);
int
r
=
0
;
xpu
::
ctx_guard
RAII_GUARD
(
dev_ctx
.
x_context
());
XPUType
*
tmp_x
=
RAII_GUARD
.
alloc_l3_or_gm
<
XPUType
>
(
m
*
t
*
n
);
PADDLE_ENFORCE_XDNN_NOT_NULL
(
tmp_x
);
r
=
xpu
::
abs
(
dev_ctx
.
x_context
(),
reinterpret_cast
<
const
XPUType
*>
(
in
->
data
<
T
>
()),
tmp_x
,
m
*
t
*
n
);
PADDLE_ENFORCE_XDNN_SUCCESS
(
r
,
"abs"
);
if
(
porder
==
INFINITY
)
{
r
=
xpu
::
reduce_max
(
dev_ctx
.
x_context
(),
tmp_x
,
reinterpret_cast
<
XPUType
*>
(
out
->
data
<
T
>
()),
x_dim
,
r_dim
);
PADDLE_ENFORCE_XDNN_SUCCESS
(
r
,
"reduce_max"
);
}
else
if
(
porder
==
-
INFINITY
)
{
r
=
xpu
::
reduce_min
(
dev_ctx
.
x_context
(),
tmp_x
,
reinterpret_cast
<
XPUType
*>
(
out
->
data
<
T
>
()),
x_dim
,
r_dim
);
PADDLE_ENFORCE_XDNN_SUCCESS
(
r
,
"reduce_min"
);
}
else
if
(
porder
==
0
)
{
XPUType
*
zeros
=
RAII_GUARD
.
alloc_l3_or_gm
<
XPUType
>
(
1
);
PADDLE_ENFORCE_XDNN_NOT_NULL
(
zeros
);
r
=
xpu
::
constant
(
dev_ctx
.
x_context
(),
zeros
,
1
,
0.0
f
);
std
::
vector
<
int
>
zeros_dim
(
1
,
1
);
bool
*
tmp2_x
=
RAII_GUARD
.
alloc_l3_or_gm
<
bool
>
(
m
*
t
*
n
);
PADDLE_ENFORCE_XDNN_NOT_NULL
(
tmp2_x
);
r
=
xpu
::
broadcast_not_equal
(
dev_ctx
.
x_context
(),
tmp_x
,
zeros
,
tmp2_x
,
x_dim
,
zeros_dim
);
PADDLE_ENFORCE_XDNN_SUCCESS
(
r
,
"broadcast_not_equal"
);
XPUType
*
x_mid
=
tmp_x
;
r
=
xpu
::
cast
<
bool
,
XPUType
>
(
dev_ctx
.
x_context
(),
tmp2_x
,
x_mid
,
m
*
t
*
n
);
PADDLE_ENFORCE_XDNN_SUCCESS
(
r
,
"cast"
);
r
=
xpu
::
reduce_sum
(
dev_ctx
.
x_context
(),
x_mid
,
reinterpret_cast
<
XPUType
*>
(
out
->
data
<
T
>
()),
x_dim
,
r_dim
);
PADDLE_ENFORCE_XDNN_SUCCESS
(
r
,
"reduce_sum"
);
}
else
{
Tensor
porder_tensor
;
framework
::
DDim
pdim
=
phi
::
make_ddim
({
1
});
porder_tensor
.
mutable_data
<
float
>
(
pdim
,
in
->
place
());
r
=
xpu
::
constant
(
dev_ctx
.
x_context
(),
porder_tensor
.
data
<
float
>
(),
1
,
porder
);
PADDLE_ENFORCE_XDNN_SUCCESS
(
r
,
"constant"
);
std
::
vector
<
int
>
p_dim
(
1
,
1
);
XPUType
*
tmp2_x
=
RAII_GUARD
.
alloc_l3_or_gm
<
XPUType
>
(
m
*
t
*
n
);
PADDLE_ENFORCE_XDNN_NOT_NULL
(
tmp2_x
);
r
=
xpu
::
broadcast_pow
(
dev_ctx
.
x_context
(),
reinterpret_cast
<
const
XPUType
*>
(
tmp_x
),
reinterpret_cast
<
const
XPUType
*>
(
porder_tensor
.
data
<
float
>
()),
tmp2_x
,
x_dim
,
p_dim
);
PADDLE_ENFORCE_XDNN_SUCCESS
(
r
,
"broadcast_pow"
);
XPUType
*
tmp_y
=
RAII_GUARD
.
alloc_l3_or_gm
<
XPUType
>
(
m
*
n
);
PADDLE_ENFORCE_XDNN_NOT_NULL
(
tmp_y
);
r
=
xpu
::
reduce_sum
(
dev_ctx
.
x_context
(),
reinterpret_cast
<
const
XPUType
*>
(
tmp2_x
),
tmp_y
,
x_dim
,
r_dim
);
PADDLE_ENFORCE_XDNN_SUCCESS
(
r
,
"reduce_sum"
);
r
=
xpu
::
constant
(
dev_ctx
.
x_context
(),
porder_tensor
.
data
<
float
>
(),
1
,
1.0
f
/
porder
);
PADDLE_ENFORCE_XDNN_SUCCESS
(
r
,
"constant"
);
r
=
xpu
::
broadcast_pow
(
dev_ctx
.
x_context
(),
reinterpret_cast
<
const
XPUType
*>
(
tmp_y
),
reinterpret_cast
<
const
XPUType
*>
(
porder_tensor
.
data
<
float
>
()),
reinterpret_cast
<
XPUType
*>
(
out
->
data
<
T
>
()),
y_dim
,
p_dim
);
PADDLE_ENFORCE_XDNN_SUCCESS
(
r
,
"broadcast_pow"
);
dev_ctx
.
Wait
();
}
}
};
template
<
typename
DeviceContext
,
typename
T
>
class
P_NormGradXPUKernel
:
public
framework
::
OpKernel
<
T
>
{
using
XPUType
=
typename
XPUTypeTrait
<
T
>::
Type
;
public:
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
auto
*
x
=
ctx
.
Input
<
Tensor
>
(
"X"
);
auto
*
y
=
ctx
.
Input
<
Tensor
>
(
"Out"
);
auto
*
dy
=
ctx
.
Input
<
Tensor
>
(
framework
::
GradVarName
(
"Out"
));
auto
*
dx
=
ctx
.
Output
<
Tensor
>
(
framework
::
GradVarName
(
"X"
));
dx
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
auto
xdim
=
x
->
dims
();
float
porder
=
ctx
.
Attr
<
float
>
(
"porder"
);
bool
asvector
=
ctx
.
Attr
<
bool
>
(
"asvector"
);
int
axis
=
ctx
.
Attr
<
int
>
(
"axis"
);
axis
=
axis
<
0
?
xdim
.
size
()
+
axis
:
axis
;
auto
&
dev_ctx
=
ctx
.
template
device_context
<
DeviceContext
>();
int
m
,
t
,
n
;
GetDims
(
xdim
,
axis
,
&
m
,
&
t
,
&
n
,
asvector
);
std
::
vector
<
int
>
r_dim
;
std
::
vector
<
int
>
x_dim
;
std
::
vector
<
int
>
y_dim
;
x_dim
.
push_back
(
m
);
x_dim
.
push_back
(
t
);
x_dim
.
push_back
(
n
);
y_dim
.
push_back
(
m
);
y_dim
.
push_back
(
1
);
y_dim
.
push_back
(
n
);
int
r
=
0
;
if
(
porder
==
0
)
{
r
=
xpu
::
constant
(
dev_ctx
.
x_context
(),
reinterpret_cast
<
XPUType
*>
(
dx
->
data
<
T
>
()),
m
*
t
*
n
,
static_cast
<
T
>
(
0
));
PADDLE_ENFORCE_XDNN_SUCCESS
(
r
,
"constant"
);
}
else
if
(
porder
==
INFINITY
||
porder
==
-
INFINITY
)
{
xpu
::
ctx_guard
RAII_GUARD
(
dev_ctx
.
x_context
());
XPUType
*
x_abs
=
RAII_GUARD
.
alloc_l3_or_gm
<
XPUType
>
(
m
*
t
*
n
);
PADDLE_ENFORCE_XDNN_NOT_NULL
(
x_abs
);
r
=
xpu
::
abs
(
dev_ctx
.
x_context
(),
reinterpret_cast
<
const
XPUType
*>
(
x
->
data
<
T
>
()),
x_abs
,
m
*
t
*
n
);
PADDLE_ENFORCE_XDNN_SUCCESS
(
r
,
"abs"
);
bool
*
dx_t
=
RAII_GUARD
.
alloc_l3_or_gm
<
bool
>
(
m
*
t
*
n
);
PADDLE_ENFORCE_XDNN_NOT_NULL
(
dx_t
);
XPUType
*
dx_mid
=
RAII_GUARD
.
alloc_l3_or_gm
<
XPUType
>
(
m
*
t
*
n
);
PADDLE_ENFORCE_XDNN_NOT_NULL
(
dx_mid
);
r
=
xpu
::
broadcast_equal
<
XPUType
>
(
dev_ctx
.
x_context
(),
reinterpret_cast
<
const
XPUType
*>
(
x_abs
),
reinterpret_cast
<
const
XPUType
*>
(
y
->
data
<
T
>
()),
dx_t
,
x_dim
,
y_dim
);
PADDLE_ENFORCE_XDNN_SUCCESS
(
r
,
"broadcast_equal"
);
r
=
xpu
::
cast
<
bool
,
XPUType
>
(
dev_ctx
.
x_context
(),
dx_t
,
dx_mid
,
m
*
t
*
n
);
PADDLE_ENFORCE_XDNN_SUCCESS
(
r
,
"cast"
);
XPUType
*
x_sign
=
RAII_GUARD
.
alloc_l3_or_gm
<
XPUType
>
(
m
*
t
*
n
);
PADDLE_ENFORCE_XDNN_NOT_NULL
(
x_sign
);
r
=
xpu
::
sign
(
dev_ctx
.
x_context
(),
reinterpret_cast
<
const
XPUType
*>
(
x
->
data
<
T
>
()),
x_sign
,
m
*
t
*
n
);
PADDLE_ENFORCE_XDNN_SUCCESS
(
r
,
"sign"
);
XPUType
*
dx_pre_dy
=
x_abs
;
r
=
xpu
::
mul
(
dev_ctx
.
x_context
(),
reinterpret_cast
<
const
XPUType
*>
(
dx_mid
),
reinterpret_cast
<
const
XPUType
*>
(
x_sign
),
dx_pre_dy
,
m
*
t
*
n
);
PADDLE_ENFORCE_XDNN_SUCCESS
(
r
,
"mul"
);
r
=
xpu
::
broadcast_mul
(
dev_ctx
.
x_context
(),
dx_pre_dy
,
reinterpret_cast
<
const
XPUType
*>
(
dy
->
data
<
T
>
()),
reinterpret_cast
<
XPUType
*>
(
dx
->
data
<
T
>
()),
x_dim
,
y_dim
);
PADDLE_ENFORCE_XDNN_SUCCESS
(
r
,
"broadcast_mul"
);
}
else
{
xpu
::
ctx_guard
RAII_GUARD
(
dev_ctx
.
x_context
());
XPUType
*
x_abs
=
RAII_GUARD
.
alloc_l3_or_gm
<
XPUType
>
(
m
*
t
*
n
);
PADDLE_ENFORCE_XDNN_NOT_NULL
(
x_abs
);
r
=
xpu
::
abs
(
dev_ctx
.
x_context
(),
reinterpret_cast
<
const
XPUType
*>
(
x
->
data
<
T
>
()),
x_abs
,
m
*
t
*
n
);
PADDLE_ENFORCE_XDNN_SUCCESS
(
r
,
"abs"
);
Tensor
porder_tensor
;
framework
::
DDim
pdim
=
phi
::
make_ddim
({
1
});
porder_tensor
.
mutable_data
<
float
>
(
pdim
,
x
->
place
());
r
=
xpu
::
constant
(
dev_ctx
.
x_context
(),
porder_tensor
.
data
<
float
>
(),
1
,
porder
-
1.0
f
);
PADDLE_ENFORCE_XDNN_SUCCESS
(
r
,
"constant"
);
std
::
vector
<
int
>
p_dim
(
1
,
1
);
XPUType
*
x_pow
=
RAII_GUARD
.
alloc_l3_or_gm
<
XPUType
>
(
m
*
t
*
n
);
PADDLE_ENFORCE_XDNN_NOT_NULL
(
x_pow
);
r
=
xpu
::
broadcast_pow
(
dev_ctx
.
x_context
(),
reinterpret_cast
<
const
XPUType
*>
(
x_abs
),
reinterpret_cast
<
const
XPUType
*>
(
porder_tensor
.
data
<
float
>
()),
x_pow
,
x_dim
,
p_dim
);
PADDLE_ENFORCE_XDNN_SUCCESS
(
r
,
"broadcast_pow"
);
XPUType
*
y_pow
=
RAII_GUARD
.
alloc_l3_or_gm
<
XPUType
>
(
m
*
n
);
PADDLE_ENFORCE_XDNN_NOT_NULL
(
y_pow
);
r
=
xpu
::
broadcast_pow
(
dev_ctx
.
x_context
(),
reinterpret_cast
<
const
XPUType
*>
(
y
->
data
<
T
>
()),
reinterpret_cast
<
const
XPUType
*>
(
porder_tensor
.
data
<
float
>
()),
y_pow
,
y_dim
,
p_dim
);
PADDLE_ENFORCE_XDNN_SUCCESS
(
r
,
"broadcast_pow"
);
dev_ctx
.
Wait
();
XPUType
*
dx_t
=
x_abs
;
r
=
xpu
::
broadcast_div
(
dev_ctx
.
x_context
(),
x_pow
,
y_pow
,
dx_t
,
x_dim
,
y_dim
);
PADDLE_ENFORCE_XDNN_SUCCESS
(
r
,
"broadcast_div"
);
XPUType
*
x_sign
=
x_pow
;
r
=
xpu
::
sign
(
dev_ctx
.
x_context
(),
reinterpret_cast
<
const
XPUType
*>
(
x
->
data
<
T
>
()),
x_sign
,
m
*
t
*
n
);
PADDLE_ENFORCE_XDNN_SUCCESS
(
r
,
"sign"
);
XPUType
*
dx_mid
=
RAII_GUARD
.
alloc_l3_or_gm
<
XPUType
>
(
m
*
t
*
n
);
PADDLE_ENFORCE_XDNN_NOT_NULL
(
dx_mid
);
r
=
xpu
::
broadcast_mul
(
dev_ctx
.
x_context
(),
reinterpret_cast
<
const
XPUType
*>
(
x_sign
),
reinterpret_cast
<
const
XPUType
*>
(
dy
->
data
<
T
>
()),
dx_mid
,
x_dim
,
y_dim
);
PADDLE_ENFORCE_XDNN_SUCCESS
(
r
,
"broadcast_mul"
);
r
=
xpu
::
broadcast_mul
(
dev_ctx
.
x_context
(),
reinterpret_cast
<
const
XPUType
*>
(
dx_t
),
reinterpret_cast
<
const
XPUType
*>
(
dx_mid
),
reinterpret_cast
<
XPUType
*>
(
dx
->
data
<
T
>
()),
x_dim
,
x_dim
);
PADDLE_ENFORCE_XDNN_SUCCESS
(
r
,
"broadcast_mul"
);
}
}
};
}
// namespace operators
}
// namespace paddle
namespace
ops
=
paddle
::
operators
;
REGISTER_OP_XPU_KERNEL
(
p_norm
,
ops
::
P_NormXPUKernel
<
paddle
::
platform
::
XPUDeviceContext
,
float
>
);
REGISTER_OP_XPU_KERNEL
(
p_norm_grad
,
ops
::
P_NormGradXPUKernel
<
paddle
::
platform
::
XPUDeviceContext
,
float
>
);
#endif
paddle/fluid/operators/pool_op_xpu.cc
浏览文件 @
84b72c5f
...
...
@@ -13,6 +13,7 @@ limitations under the License. */
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/tensor.h"
#include "paddle/phi/kernels/funcs/pooling.h"
#ifdef PADDLE_WITH_XPU
namespace
paddle
{
...
...
@@ -51,6 +52,9 @@ class PoolXPUKernel : public framework::OpKernel<T> {
std
::
vector
<
int
>
paddings
=
context
.
Attr
<
std
::
vector
<
int
>>
(
"paddings"
);
bool
exclusive
=
context
.
Attr
<
bool
>
(
"exclusive"
);
bool
adaptive
=
context
.
Attr
<
bool
>
(
"adaptive"
);
bool
ceil_mode
=
context
.
Attr
<
bool
>
(
"ceil_mode"
);
std
::
string
padding_algorithm
=
context
.
Attr
<
std
::
string
>
(
"padding_algorithm"
);
PADDLE_ENFORCE_EQ
(
ksize
.
size
(),
2
,
...
...
@@ -70,10 +74,27 @@ class PoolXPUKernel : public framework::OpKernel<T> {
ksize
[
i
]
=
static_cast
<
int
>
(
in_x
->
dims
()[
i
+
2
]);
}
}
const
int
n
=
in_x
->
dims
()[
0
];
const
int
c
=
in_x
->
dims
()[
1
];
const
int
in_h
=
in_x
->
dims
()[
2
];
const
int
in_w
=
in_x
->
dims
()[
3
];
framework
::
DDim
data_dims
;
data_dims
=
phi
::
slice_ddim
(
in_x
->
dims
(),
2
,
in_x
->
dims
().
size
());
phi
::
funcs
::
UpdatePadding
(
&
paddings
,
global_pooling
,
adaptive
,
padding_algorithm
,
data_dims
,
strides
,
ksize
);
if
(
ceil_mode
)
{
paddings
[
1
]
+=
(
strides
[
0
]
-
1
);
paddings
[
3
]
+=
(
strides
[
1
]
-
1
);
}
auto
input
=
reinterpret_cast
<
const
XPUType
*>
(
in_x
->
data
<
T
>
());
out
->
mutable_data
<
T
>
(
context
.
GetPlace
());
auto
output
=
reinterpret_cast
<
XPUType
*>
(
out
->
data
<
T
>
());
...
...
@@ -135,6 +156,9 @@ class PoolGradXPUKernel : public framework::OpKernel<T> {
std
::
vector
<
int
>
paddings
=
context
.
Attr
<
std
::
vector
<
int
>>
(
"paddings"
);
bool
exclusive
=
context
.
Attr
<
bool
>
(
"exclusive"
);
bool
adaptive
=
context
.
Attr
<
bool
>
(
"adaptive"
);
bool
ceil_mode
=
context
.
Attr
<
bool
>
(
"ceil_mode"
);
std
::
string
padding_algorithm
=
context
.
Attr
<
std
::
string
>
(
"padding_algorithm"
);
const
int
*
index_data
=
nullptr
;
PADDLE_ENFORCE_EQ
(
ksize
.
size
(),
...
...
@@ -163,6 +187,22 @@ class PoolGradXPUKernel : public framework::OpKernel<T> {
const
int
c
=
in_x
->
dims
()[
1
];
const
int
in_h
=
in_x
->
dims
()[
2
];
const
int
in_w
=
in_x
->
dims
()[
3
];
framework
::
DDim
data_dims
;
data_dims
=
phi
::
slice_ddim
(
in_x
->
dims
(),
2
,
in_x
->
dims
().
size
());
phi
::
funcs
::
UpdatePadding
(
&
paddings
,
global_pooling
,
adaptive
,
padding_algorithm
,
data_dims
,
strides
,
ksize
);
if
(
ceil_mode
)
{
paddings
[
1
]
+=
(
strides
[
0
]
-
1
);
paddings
[
3
]
+=
(
strides
[
1
]
-
1
);
}
auto
input
=
reinterpret_cast
<
const
XPUType
*>
(
in_x
->
data
<
T
>
());
auto
output
=
reinterpret_cast
<
const
XPUType
*>
(
out
->
data
<
T
>
());
auto
output_grad
=
reinterpret_cast
<
const
XPUType
*>
(
out_grad
->
data
<
T
>
());
...
...
paddle/fluid/platform/device/xpu/xpu2_op_list.h
浏览文件 @
84b72c5f
...
...
@@ -323,6 +323,8 @@ XPUOpMap& get_kl2_ops() {
{
"one_hot_v2"
,
XPUKernelSet
({
pOpKernelType
(
vartype
::
INT32
,
XPUPlace
()),
pOpKernelType
(
vartype
::
INT64
,
XPUPlace
())})},
{
"p_norm"
,
XPUKernelSet
({
pOpKernelType
(
vartype
::
FP32
,
XPUPlace
())})},
{
"p_norm_grad"
,
XPUKernelSet
({
pOpKernelType
(
vartype
::
FP32
,
XPUPlace
())})},
{
"pool2d_grad"
,
XPUKernelSet
({
pOpKernelType
(
vartype
::
FP32
,
XPUPlace
()),
pOpKernelType
(
vartype
::
FP16
,
XPUPlace
())})},
...
...
python/paddle/fluid/tests/unittests/xpu/test_p_norm_op_xpu.py
0 → 100644
浏览文件 @
84b72c5f
# 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.
import
paddle
import
numpy
as
np
import
sys
import
unittest
from
functools
import
reduce
sys
.
path
.
append
(
".."
)
from
op_test
import
OpTest
from
op_test_xpu
import
XPUOpTest
from
operator
import
mul
from
xpu.get_test_cover_info
import
create_test_class
,
get_xpu_op_support_types
,
XPUOpTestWrapper
paddle
.
enable_static
()
def
ref_p_norm
(
x
,
axis
,
porder
,
keepdims
=
False
,
reduce_all
=
False
):
r
=
[]
if
axis
is
None
or
reduce_all
:
x
=
x
.
flatten
()
if
porder
==
np
.
inf
:
r
=
np
.
amax
(
np
.
abs
(
x
),
keepdims
=
keepdims
)
elif
porder
==
-
np
.
inf
:
r
=
np
.
amin
(
np
.
abs
(
x
),
keepdims
=
keepdims
)
else
:
r
=
np
.
linalg
.
norm
(
x
,
ord
=
porder
,
keepdims
=
keepdims
)
elif
isinstance
(
axis
,
list
or
tuple
)
and
len
(
axis
)
==
2
:
if
porder
==
np
.
inf
:
axis
=
tuple
(
axis
)
r
=
np
.
amax
(
np
.
abs
(
x
),
axis
=
axis
,
keepdims
=
keepdims
)
elif
porder
==
-
np
.
inf
:
axis
=
tuple
(
axis
)
r
=
np
.
amin
(
np
.
abs
(
x
),
axis
=
axis
,
keepdims
=
keepdims
)
elif
porder
==
0
:
axis
=
tuple
(
axis
)
r
=
x
.
astype
(
bool
)
r
=
np
.
sum
(
r
,
axis
,
keepdims
=
keepdims
)
elif
porder
==
1
:
axis
=
tuple
(
axis
)
r
=
np
.
sum
(
np
.
abs
(
x
),
axis
,
keepdims
=
keepdims
)
else
:
axis
=
tuple
(
axis
)
xp
=
np
.
power
(
np
.
abs
(
x
),
porder
)
s
=
np
.
sum
(
xp
,
axis
=
axis
,
keepdims
=
keepdims
)
r
=
np
.
power
(
s
,
1.0
/
porder
)
else
:
if
isinstance
(
axis
,
list
):
axis
=
tuple
(
axis
)
r
=
np
.
linalg
.
norm
(
x
,
ord
=
porder
,
axis
=
axis
,
keepdims
=
keepdims
)
r
=
r
.
astype
(
x
.
dtype
)
return
r
class
XPUTestPNormOp
(
XPUOpTestWrapper
):
def
__init__
(
self
):
self
.
op_name
=
'p_norm'
self
.
use_dynamic_create_class
=
False
class
TestXPUPNormOp
(
XPUOpTest
):
def
setUp
(
self
):
self
.
op_type
=
"p_norm"
self
.
dtype
=
self
.
in_type
self
.
shape
=
[
2
,
3
,
4
,
5
]
self
.
epsilon
=
1e-12
self
.
axis
=
1
self
.
porder
=
2.0
self
.
asvector
=
False
self
.
keepdims
=
False
self
.
set_attrs
()
np
.
random
.
seed
(
12345
)
x_np
=
np
.
random
.
uniform
(
-
10
,
10
,
self
.
shape
).
astype
(
self
.
dtype
)
ref_y_np
=
ref_p_norm
(
x_np
,
self
.
axis
,
self
.
porder
,
self
.
keepdims
,
self
.
asvector
)
self
.
inputs
=
{
'X'
:
x_np
}
self
.
outputs
=
{
'Out'
:
ref_y_np
}
self
.
attrs
=
{
'epsilon'
:
self
.
epsilon
,
'axis'
:
self
.
axis
,
'porder'
:
float
(
self
.
porder
),
'asvector'
:
self
.
asvector
}
def
set_attrs
(
self
):
pass
def
test_check_output
(
self
):
self
.
check_output_with_place
(
paddle
.
XPUPlace
(
0
),
atol
=
1e-4
)
def
test_check_grad
(
self
):
self
.
check_grad_with_place
(
paddle
.
XPUPlace
(
0
),
[
'X'
],
'Out'
)
class
TestPnormOp2
(
TestXPUPNormOp
):
def
set_attrs
(
self
):
self
.
shape
=
[
3
,
20
,
3
]
self
.
axis
=
2
self
.
porder
=
2.0
class
TestPnormOp3
(
TestXPUPNormOp
):
def
set_attrs
(
self
):
self
.
shape
=
[
3
,
20
,
3
]
self
.
axis
=
2
self
.
porder
=
np
.
inf
class
TestPnormOp4
(
TestXPUPNormOp
):
def
set_attrs
(
self
):
self
.
shape
=
[
3
,
20
,
3
]
self
.
axis
=
2
self
.
porder
=
-
np
.
inf
class
TestPnormOp5
(
TestXPUPNormOp
):
def
set_attrs
(
self
):
self
.
shape
=
[
3
,
20
,
3
]
self
.
axis
=
2
self
.
porder
=
0
class
TestPnormOp6
(
TestXPUPNormOp
):
def
set_attrs
(
self
):
self
.
shape
=
[
3
,
20
,
3
]
self
.
axis
=
-
1
self
.
porder
=
2
class
TestPnormOp7
(
TestXPUPNormOp
):
def
set_attrs
(
self
):
self
.
shape
=
[
3
,
20
,
3
,
10
]
self
.
axis
=
2
self
.
porder
=
2.0
class
TestPnormOp8
(
TestXPUPNormOp
):
def
set_attrs
(
self
):
self
.
shape
=
[
3
,
20
,
3
]
self
.
axis
=
2
self
.
porder
=
np
.
inf
class
TestPnormOp9
(
TestXPUPNormOp
):
def
set_attrs
(
self
):
self
.
shape
=
[
3
,
20
,
3
,
10
]
self
.
axis
=
1
self
.
porder
=
-
np
.
inf
class
TestPnormOp10
(
TestXPUPNormOp
):
def
set_attrs
(
self
):
self
.
shape
=
[
3
,
20
,
3
,
10
]
self
.
axis
=
2
self
.
porder
=
0
class
TestPnormOp11
(
TestXPUPNormOp
):
def
set_attrs
(
self
):
self
.
shape
=
[
3
,
20
,
3
,
10
]
self
.
axis
=
-
1
self
.
porder
=
2
support_types
=
get_xpu_op_support_types
(
'p_norm'
)
for
stype
in
support_types
:
create_test_class
(
globals
(),
XPUTestPNormOp
,
stype
)
if
__name__
==
"__main__"
:
unittest
.
main
()
python/paddle/fluid/tests/unittests/xpu/test_pool2d_op_xpu.py
浏览文件 @
84b72c5f
...
...
@@ -297,6 +297,7 @@ class XPUTestPool2D_Op(XPUOpTestWrapper):
'exclusive'
:
self
.
exclusive
,
'adaptive'
:
self
.
adaptive
,
"padding_algorithm"
:
self
.
padding_algorithm
,
'ceil_mode'
:
self
.
ceil_mode
}
self
.
outputs
=
{
'Out'
:
output
}
...
...
@@ -469,6 +470,77 @@ class XPUTestPool2D_Op(XPUOpTestWrapper):
def
init_shape
(
self
):
self
.
shape
=
[
2
,
3
,
7
,
7
]
class
TestCaseCeil1
(
TestPool2D_Op
):
def
init_test_case
(
self
):
self
.
ksize
=
[
3
,
3
]
self
.
strides
=
[
1
,
1
]
def
init_paddings
(
self
):
self
.
paddings
=
[
0
,
0
]
def
init_pool_type
(
self
):
self
.
pool_type
=
"avg"
self
.
pool2D_forward_naive
=
avg_pool2D_forward_naive
def
init_global_pool
(
self
):
self
.
global_pool
=
False
def
init_shape
(
self
):
self
.
shape
=
[
2
,
3
,
7
,
7
]
def
init_ceil_mode
(
self
):
self
.
ceil_mode
=
True
class
TestCaseCeil2
(
TestPool2D_Op
):
def
init_test_case
(
self
):
self
.
ksize
=
[
3
,
3
]
self
.
strides
=
[
1
,
1
]
def
init_paddings
(
self
):
self
.
paddings
=
[
1
,
1
]
def
init_pool_type
(
self
):
self
.
pool_type
=
"avg"
self
.
pool2D_forward_naive
=
avg_pool2D_forward_naive
def
init_global_pool
(
self
):
self
.
global_pool
=
False
def
init_shape
(
self
):
self
.
shape
=
[
2
,
3
,
7
,
7
]
def
init_ceil_mode
(
self
):
self
.
ceil_mode
=
True
class
TestCaseCeil3
(
TestPool2D_Op
):
def
init_pool_type
(
self
):
self
.
pool_type
=
"max"
self
.
pool2D_forward_naive
=
max_pool2D_forward_naive
def
init_ceil_mode
(
self
):
self
.
ceil_mode
=
True
class
TestCaseCeil4
(
TestCaseCeil1
):
def
init_pool_type
(
self
):
self
.
pool_type
=
"max"
self
.
pool2D_forward_naive
=
max_pool2D_forward_naive
def
init_ceil_mode
(
self
):
self
.
ceil_mode
=
True
class
TestCaseCeil5
(
TestCaseCeil2
):
def
init_pool_type
(
self
):
self
.
pool_type
=
"max"
self
.
pool2D_forward_naive
=
max_pool2D_forward_naive
def
init_ceil_mode
(
self
):
self
.
ceil_mode
=
True
support_types
=
get_xpu_op_support_types
(
'pool2d'
)
for
stype
in
support_types
:
...
...
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