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
yushenjjy
Paddle
提交
5098891f
P
Paddle
项目概览
yushenjjy
/
Paddle
与 Fork 源项目一致
Fork自
PaddlePaddle / Paddle
通知
1
Star
0
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
5098891f
编写于
10月 10, 2020
作者:
Z
zhupengyang
提交者:
GitHub
10月 10, 2020
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
add softmax xpu kernel (#27700)
上级
65c06141
变更
2
隐藏空白更改
内联
并排
Showing
2 changed file
with
192 addition
and
0 deletion
+192
-0
paddle/fluid/operators/softmax_op_xpu.cc
paddle/fluid/operators/softmax_op_xpu.cc
+99
-0
python/paddle/fluid/tests/unittests/xpu/test_softmax_op_xpu.py
...n/paddle/fluid/tests/unittests/xpu/test_softmax_op_xpu.py
+93
-0
未找到文件。
paddle/fluid/operators/softmax_op_xpu.cc
0 → 100644
浏览文件 @
5098891f
/* Copyright (c) 2020 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/operators/softmax_op.h"
#include "paddle/fluid/framework/op_registry.h"
namespace
paddle
{
namespace
operators
{
using
Tensor
=
framework
::
Tensor
;
using
DDim
=
framework
::
DDim
;
template
<
typename
DeviceContext
,
typename
T
>
class
SoftmaxXPUKernel
:
public
framework
::
OpKernel
<
T
>
{
public:
void
Compute
(
const
framework
::
ExecutionContext
&
context
)
const
override
{
auto
*
x
=
context
.
Input
<
Tensor
>
(
"X"
);
auto
*
out
=
context
.
Output
<
Tensor
>
(
"Out"
);
const
int
rank
=
x
->
dims
().
size
();
const
int
axis
=
CanonicalAxis
(
context
.
Attr
<
int
>
(
"axis"
),
rank
);
PADDLE_ENFORCE_EQ
(
axis
==
-
1
||
axis
==
rank
-
1
,
true
,
platform
::
errors
::
InvalidArgument
(
"xpu softmax kernel only support last dimension of x "
"(axis==-1 or axis==x_dims-1), but received axis: "
"%d, x's shape: %s."
,
axis
,
x
->
dims
()));
// allocate memory on device.
out
->
mutable_data
<
T
>
(
context
.
GetPlace
());
const
int
n
=
SizeToAxis
(
axis
,
x
->
dims
());
const
int
d
=
SizeFromAxis
(
axis
,
x
->
dims
());
auto
&
dev_ctx
=
context
.
template
device_context
<
DeviceContext
>();
int
r
=
xpu
::
softmax2d_forward
(
dev_ctx
.
x_context
(),
x
->
data
<
float
>
(),
out
->
data
<
float
>
(),
n
,
d
,
d
<=
2048
);
PADDLE_ENFORCE_EQ
(
r
,
XPU_SUCCESS
,
platform
::
errors
::
External
(
"XPU API(softmax2d_forward) return wrong "
"value[%d], please check whether "
"Baidu Kunlun Card is properly installed."
,
r
));
}
};
template
<
typename
DeviceContext
,
typename
T
>
class
SoftmaxGradXPUKernel
:
public
framework
::
OpKernel
<
T
>
{
public:
void
Compute
(
const
framework
::
ExecutionContext
&
context
)
const
override
{
auto
*
out
=
context
.
Input
<
Tensor
>
(
"Out"
);
auto
*
dout
=
context
.
Input
<
Tensor
>
(
framework
::
GradVarName
(
"Out"
));
auto
*
dx
=
context
.
Output
<
Tensor
>
(
framework
::
GradVarName
(
"X"
));
const
int
rank
=
dx
->
dims
().
size
();
const
int
axis
=
CanonicalAxis
(
context
.
Attr
<
int
>
(
"axis"
),
rank
);
// allocate memory on device.
dx
->
mutable_data
<
T
>
(
context
.
GetPlace
());
const
int
n
=
SizeToAxis
(
axis
,
dx
->
dims
());
const
int
d
=
SizeFromAxis
(
axis
,
dx
->
dims
());
auto
&
dev_ctx
=
context
.
template
device_context
<
DeviceContext
>();
int
r
=
xpu
::
softmax2d_backward
(
dev_ctx
.
x_context
(),
out
->
data
<
float
>
(),
dout
->
data
<
float
>
(),
dx
->
data
<
float
>
(),
n
,
d
);
PADDLE_ENFORCE_EQ
(
r
,
XPU_SUCCESS
,
platform
::
errors
::
External
(
"XPU API(softmax2d_backward) return wrong "
"value[%d], please check whether "
"Baidu Kunlun Card is properly installed."
,
r
));
}
};
}
// namespace operators
}
// namespace paddle
namespace
ops
=
paddle
::
operators
;
REGISTER_OP_XPU_KERNEL
(
softmax
,
ops
::
SoftmaxXPUKernel
<
paddle
::
platform
::
XPUDeviceContext
,
float
>
);
REGISTER_OP_XPU_KERNEL
(
softmax_grad
,
ops
::
SoftmaxGradXPUKernel
<
paddle
::
platform
::
XPUDeviceContext
,
float
>
);
#endif // PADDLE_WITH_XPU
python/paddle/fluid/tests/unittests/xpu/test_softmax_op_xpu.py
0 → 100644
浏览文件 @
5098891f
# Copyright (c) 2020 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
sys
.
path
.
append
(
".."
)
from
op_test
import
OpTest
paddle
.
enable_static
()
np
.
random
.
seed
(
10
)
def
stable_softmax
(
x
):
"""Compute the softmax of vector x in a numerically stable way."""
# clip to shiftx, otherwise, when calc loss with
# log(exp(shiftx)), may get log(0)=INF
shiftx
=
(
x
-
np
.
max
(
x
)).
clip
(
-
64.
)
exps
=
np
.
exp
(
shiftx
)
return
exps
/
np
.
sum
(
exps
)
def
ref_softmax
(
x
,
axis
=
None
,
dtype
=
None
):
x_t
=
x
.
copy
()
if
dtype
is
not
None
:
x_t
=
x_t
.
astype
(
dtype
)
if
axis
is
None
:
axis
=
-
1
return
np
.
apply_along_axis
(
stable_softmax
,
axis
,
x_t
)
@
unittest
.
skipIf
(
not
paddle
.
is_compiled_with_xpu
(),
"core is not compiled with XPU"
)
class
TestXPUSoftmaxOp
(
OpTest
):
def
setUp
(
self
):
self
.
op_type
=
"softmax"
self
.
dtype
=
np
.
float32
self
.
shape
=
[
2
,
3
,
4
,
5
]
self
.
axis
=
-
1
self
.
set_attrs
()
x
=
np
.
random
.
uniform
(
-
1
,
1
,
self
.
shape
).
astype
(
self
.
dtype
)
out
=
np
.
apply_along_axis
(
stable_softmax
,
self
.
axis
,
x
)
self
.
inputs
=
{
'X'
:
x
}
self
.
outputs
=
{
'Out'
:
out
}
self
.
attrs
=
{
'axis'
:
self
.
axis
,
'use_xpu'
:
True
}
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'
)
@
unittest
.
skipIf
(
not
paddle
.
is_compiled_with_xpu
(),
"core is not compiled with XPU"
)
class
TestXPUSoftmaxAxis3
(
TestXPUSoftmaxOp
):
def
set_attrs
(
self
):
self
.
axis
=
3
@
unittest
.
skipIf
(
not
paddle
.
is_compiled_with_xpu
(),
"core is not compiled with XPU"
)
class
TestXPUSoftmax2D
(
TestXPUSoftmaxOp
):
def
set_attrs
(
self
):
self
.
shape
=
[
10
,
12
]
@
unittest
.
skipIf
(
not
paddle
.
is_compiled_with_xpu
(),
"core is not compiled with XPU"
)
class
TestXPUSoftmax3D
(
TestXPUSoftmaxOp
):
def
set_attrs
(
self
):
self
.
shape
=
[
4
,
5
,
6
]
if
__name__
==
"__main__"
:
unittest
.
main
()
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