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
ed478a3e
P
Paddle
项目概览
PaddlePaddle
/
Paddle
12 个月 前同步成功
通知
2292
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
ed478a3e
编写于
10月 21, 2021
作者:
Z
zhulei
提交者:
GitHub
10月 21, 2021
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
[NPU] Add p_norm_grad (#36497)
上级
7eab0fa6
变更
2
隐藏空白更改
内联
并排
Showing
2 changed file
with
136 addition
and
1 deletion
+136
-1
paddle/fluid/operators/p_norm_op_npu.cc
paddle/fluid/operators/p_norm_op_npu.cc
+120
-0
python/paddle/fluid/tests/unittests/npu/test_p_norm_op_npu.py
...on/paddle/fluid/tests/unittests/npu/test_p_norm_op_npu.py
+16
-1
未找到文件。
paddle/fluid/operators/p_norm_op_npu.cc
浏览文件 @
ed478a3e
...
...
@@ -81,6 +81,122 @@ class PnormNPUKernel : public framework::OpKernel<T> {
}
};
template
<
typename
DeviceContext
,
typename
T
>
class
PnormGradNPUKernel
:
public
framework
::
OpKernel
<
T
>
{
public:
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
using
Tensor
=
framework
::
Tensor
;
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"
));
auto
place
=
ctx
.
GetPlace
();
dx
->
mutable_data
<
T
>
(
place
);
auto
xdim
=
x
->
dims
();
float
porder
=
ctx
.
Attr
<
float
>
(
"porder"
);
bool
keepdim
=
ctx
.
Attr
<
bool
>
(
"keepdim"
);
int
axis
=
ctx
.
Attr
<
int
>
(
"axis"
);
axis
=
axis
<
0
?
xdim
.
size
()
+
axis
:
axis
;
auto
stream
=
ctx
.
template
device_context
<
paddle
::
platform
::
NPUDeviceContext
>()
.
stream
();
Tensor
y_share
(
y
->
type
());
Tensor
dy_share
(
dy
->
type
());
y_share
.
ShareDataWith
(
*
y
);
dy_share
.
ShareDataWith
(
*
dy
);
auto
ydim
=
xdim
;
if
(
!
keepdim
)
{
ydim
[
axis
]
=
1
;
}
else
{
ydim
=
y
->
dims
();
}
y_share
.
Resize
(
ydim
);
dy_share
.
Resize
(
ydim
);
if
(
porder
==
0
)
{
FillNpuTensorWithConstant
(
dx
,
static_cast
<
T
>
(
0
));
dx
->
Resize
(
xdim
);
}
else
if
(
porder
==
INFINITY
||
porder
==
-
INFINITY
)
{
Tensor
x_abs
;
x_abs
.
mutable_data
<
T
>
(
xdim
,
place
);
const
auto
&
r_abs
=
NpuOpRunner
(
"Abs"
,
{
*
x
},
{
x_abs
},
{});
r_abs
.
Run
(
stream
);
Tensor
t_cond
;
t_cond
.
mutable_data
<
bool
>
(
xdim
,
place
);
const
auto
&
r_equal
=
NpuOpRunner
(
"Equal"
,
{
x_abs
,
y_share
},
{
t_cond
},
{});
r_equal
.
Run
(
stream
);
Tensor
t_zero
;
t_zero
.
mutable_data
<
T
>
({
1
},
place
);
FillNpuTensorWithConstant
(
&
t_zero
,
static_cast
<
T
>
(
0
));
Tensor
x_sign
;
x_sign
.
mutable_data
<
T
>
(
xdim
,
place
);
const
auto
&
r_sign
=
NpuOpRunner
(
"Sign"
,
{
*
x
},
{
x_sign
},
{});
r_sign
.
Run
(
stream
);
const
auto
&
r_mul
=
NpuOpRunner
(
"Mul"
,
{
x_sign
,
dy_share
},
{
*
dx
},
{});
r_mul
.
Run
(
stream
);
const
auto
&
r_sel
=
NpuOpRunner
(
"SelectV2"
,
{
t_cond
,
*
dx
,
t_zero
},
{
*
dx
},
{});
r_sel
.
Run
(
stream
);
}
else
{
Tensor
x_abs
;
x_abs
.
mutable_data
<
T
>
(
xdim
,
place
);
const
auto
&
r_abs
=
NpuOpRunner
(
"Abs"
,
{
*
x
},
{
x_abs
},
{});
r_abs
.
Run
(
stream
);
Tensor
x_sign
;
x_sign
.
mutable_data
<
T
>
(
xdim
,
place
);
const
auto
&
r_sign
=
NpuOpRunner
(
"Sign"
,
{
*
x
},
{
x_sign
},
{});
r_sign
.
Run
(
stream
);
Tensor
y_pow
;
y_pow
.
mutable_data
<
T
>
(
ydim
,
place
);
if
(
porder
>=
1
)
{
const
auto
&
r_pow1
=
NpuOpRunner
(
"Power"
,
{
x_abs
},
{
x_abs
},
{{
"power"
,
(
porder
-
1
)},
{
"scale"
,
1.0
f
},
{
"shift"
,
0.0
f
}});
r_pow1
.
Run
(
stream
);
const
auto
&
r_pow2
=
NpuOpRunner
(
"Power"
,
{
y_share
},
{
y_pow
},
{{
"power"
,
(
porder
-
1
)},
{
"scale"
,
1.0
f
},
{
"shift"
,
0.0
f
}});
r_pow2
.
Run
(
stream
);
const
auto
&
r_div
=
NpuOpRunner
(
"DivNoNan"
,
{
x_abs
,
y_pow
},
{
*
dx
},
{});
r_div
.
Run
(
stream
);
}
else
{
const
auto
&
r_pow1
=
NpuOpRunner
(
"Power"
,
{
x_abs
},
{
x_abs
},
{{
"power"
,
(
1
-
porder
)},
{
"scale"
,
1.0
f
},
{
"shift"
,
0.0
f
}});
r_pow1
.
Run
(
stream
);
const
auto
&
r_pow2
=
NpuOpRunner
(
"Power"
,
{
y_share
},
{
y_pow
},
{{
"power"
,
(
1
-
porder
)},
{
"scale"
,
1.0
f
},
{
"shift"
,
0.0
f
}});
r_pow2
.
Run
(
stream
);
const
auto
&
r_div
=
NpuOpRunner
(
"DivNoNan"
,
{
y_pow
,
x_abs
},
{
*
dx
},
{});
r_div
.
Run
(
stream
);
}
const
auto
&
r_mul1
=
NpuOpRunner
(
"Mul"
,
{
*
dx
,
x_sign
},
{
*
dx
},
{});
r_mul1
.
Run
(
stream
);
const
auto
&
r_mul2
=
NpuOpRunner
(
"Mul"
,
{
*
dx
,
dy_share
},
{
*
dx
},
{});
r_mul2
.
Run
(
stream
);
}
}
};
}
// namespace operators
}
// namespace paddle
...
...
@@ -90,3 +206,7 @@ namespace plat = paddle::platform;
REGISTER_OP_NPU_KERNEL
(
p_norm
,
ops
::
PnormNPUKernel
<
plat
::
NPUDeviceContext
,
float
>
,
ops
::
PnormNPUKernel
<
plat
::
NPUDeviceContext
,
plat
::
float16
>
);
REGISTER_OP_NPU_KERNEL
(
p_norm_grad
,
ops
::
PnormGradNPUKernel
<
plat
::
NPUDeviceContext
,
float
>
,
ops
::
PnormGradNPUKernel
<
plat
::
NPUDeviceContext
,
plat
::
float16
>
);
python/paddle/fluid/tests/unittests/npu/test_p_norm_op_npu.py
浏览文件 @
ed478a3e
...
...
@@ -27,7 +27,6 @@ paddle.enable_static()
class
TestPnormOp
(
OpTest
):
def
set_npu
(
self
):
self
.
__class__
.
use_npu
=
True
self
.
__class__
.
no_need_check_grad
=
True
def
setUp
(
self
):
self
.
set_npu
()
...
...
@@ -51,6 +50,12 @@ class TestPnormOp(OpTest):
else
:
self
.
check_output_with_place
(
paddle
.
NPUPlace
(
0
))
def
test_check_grad
(
self
):
if
self
.
dtype
==
"float16"
:
return
self
.
check_grad_with_place
(
paddle
.
NPUPlace
(
0
),
[
'X'
],
'Out'
,
user_defined_grads
=
self
.
gradient
)
def
init_test_case
(
self
):
self
.
shape
=
[
2
,
3
,
4
,
5
]
self
.
axis
=
1
...
...
@@ -131,6 +136,16 @@ class TestPnormOp5(TestPnormOp3):
self
.
init_dtype
()
class
TestPnormOp6
(
TestPnormOp3
):
def
init_test_case
(
self
):
self
.
shape
=
[
2
,
3
,
4
,
5
]
self
.
axis
=
1
self
.
epsilon
=
1e-12
self
.
porder
=
0.5
self
.
keepdim
=
False
self
.
init_dtype
()
class
TestPnormOpfp16
(
TestPnormOp
):
def
init_dtype
(
self
):
self
.
dtype
=
"float16"
...
...
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