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
d876952d
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
d876952d
编写于
6月 21, 2022
作者:
Z
zhaoying9105
提交者:
GitHub
6月 21, 2022
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
[MLU]: add argsort/argsort_grad kernel (#43574)
上级
829723f2
变更
5
隐藏空白更改
内联
并排
Showing
5 changed file
with
224 addition
and
5 deletion
+224
-5
paddle/fluid/operators/argsort_op_mlu.cc
paddle/fluid/operators/argsort_op_mlu.cc
+109
-0
paddle/fluid/operators/gather_op_mlu.cc
paddle/fluid/operators/gather_op_mlu.cc
+3
-3
paddle/fluid/operators/mlu/mlu_baseop.cc
paddle/fluid/operators/mlu/mlu_baseop.cc
+14
-1
paddle/fluid/operators/mlu/mlu_baseop.h
paddle/fluid/operators/mlu/mlu_baseop.h
+10
-1
python/paddle/fluid/tests/unittests/mlu/test_argsort_op_mlu.py
...n/paddle/fluid/tests/unittests/mlu/test_argsort_op_mlu.py
+88
-0
未找到文件。
paddle/fluid/operators/argsort_op_mlu.cc
0 → 100644
浏览文件 @
d876952d
/* 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/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/mlu/mlu_baseop.h"
namespace
paddle
{
namespace
operators
{
template
<
typename
T
>
class
ArgsortMLUKernel
:
public
framework
::
OpKernel
<
T
>
{
public:
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
auto
*
input
=
ctx
.
Input
<
framework
::
LoDTensor
>
(
"X"
);
auto
*
output
=
ctx
.
Output
<
framework
::
LoDTensor
>
(
"Out"
);
auto
*
indices
=
ctx
.
Output
<
framework
::
LoDTensor
>
(
"Indices"
);
const
auto
&
place
=
ctx
.
GetPlace
();
const
auto
&
sorted
=
true
;
const
bool
descending
=
ctx
.
Attr
<
bool
>
(
"descending"
);
// axis < 0, cacluate the real axis
int
axis
=
static_cast
<
int
>
(
ctx
.
Attr
<
int
>
(
"axis"
));
if
(
axis
<
0
)
{
const
auto
&
in_dims
=
input
->
dims
();
axis
+=
in_dims
.
size
();
}
auto
in_dims
=
input
->
dims
();
size_t
k
=
in_dims
[
axis
];
output
->
mutable_data
<
T
>
(
place
);
indices
->
mutable_data
<
int64_t
>
(
place
);
// cnnl only support int32/int16 type of indices
framework
::
Tensor
indices_int32
(
framework
::
TransToPhiDataType
(
VT
::
INT32
));
indices_int32
.
Resize
(
indices
->
dims
());
indices_int32
.
mutable_data
<
int32_t
>
(
place
);
MLUCnnlTensorDesc
input_desc
(
*
input
);
MLUCnnlTensorDesc
values_output_desc
(
*
output
);
MLUCnnlTensorDesc
indices_int32_desc
(
indices_int32
);
MLUCnnl
::
TopK
(
ctx
,
k
,
axis
,
descending
,
sorted
,
input_desc
.
get
(),
GetBasePtr
(
input
),
values_output_desc
.
get
(),
GetBasePtr
(
output
),
indices_int32_desc
.
get
(),
GetBasePtr
(
&
indices_int32
));
// cast indices type to int64
MLUCnnlTensorDesc
cast_output_desc
(
*
indices
);
cnnlCastDataType_t
cast_type
=
GetCastDataType
(
VT
::
INT32
,
VT
::
INT64
);
MLUCnnl
::
Cast
(
ctx
,
cast_type
,
indices_int32_desc
.
get
(),
GetBasePtr
(
&
indices_int32
),
cast_output_desc
.
get
(),
GetBasePtr
(
indices
));
}
};
template
<
typename
T
>
class
ArgsortGradMLUKernel
:
public
framework
::
OpKernel
<
T
>
{
public:
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
auto
*
indices
=
ctx
.
Input
<
Tensor
>
(
"Indices"
);
auto
*
dx
=
ctx
.
Output
<
Tensor
>
(
framework
::
GradVarName
(
"X"
));
auto
*
dout
=
ctx
.
Input
<
Tensor
>
(
framework
::
GradVarName
(
"Out"
));
int
axis
=
ctx
.
Attr
<
int
>
(
"axis"
);
dx
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
auto
in_dims
=
indices
->
dims
();
axis
=
(
axis
<
0
)
?
(
in_dims
.
size
()
+
axis
)
:
axis
;
if
(
dout
->
numel
()
==
0
)
return
;
MLUCnnlTensorDesc
dout_desc
(
*
dout
);
MLUCnnlTensorDesc
indices_desc
(
*
indices
);
MLUCnnlTensorDesc
dx_desc
(
*
dx
);
MLUCnnl
::
ScatterFunctor
(
ctx
,
dx_desc
.
get
(),
GetBasePtr
(
dx
),
dout_desc
.
get
(),
GetBasePtr
(
dout
),
indices_desc
.
get
(),
GetBasePtr
(
indices
),
axis
);
}
};
}
// namespace operators
}
// namespace paddle
namespace
ops
=
paddle
::
operators
;
REGISTER_OP_MLU_KERNEL
(
argsort
,
ops
::
ArgsortMLUKernel
<
paddle
::
platform
::
float16
>
,
ops
::
ArgsortMLUKernel
<
float
>
,
ops
::
ArgsortMLUKernel
<
int8_t
>
,
ops
::
ArgsortMLUKernel
<
uint8_t
>
,
ops
::
ArgsortMLUKernel
<
int16_t
>
,
ops
::
ArgsortMLUKernel
<
int
>
);
REGISTER_OP_MLU_KERNEL
(
argsort_grad
,
ops
::
ArgsortGradMLUKernel
<
paddle
::
platform
::
float16
>
,
ops
::
ArgsortGradMLUKernel
<
float
>
,
ops
::
ArgsortGradMLUKernel
<
int8_t
>
,
ops
::
ArgsortGradMLUKernel
<
uint8_t
>
,
ops
::
ArgsortGradMLUKernel
<
int16_t
>
,
ops
::
ArgsortGradMLUKernel
<
int
>
);
paddle/fluid/operators/gather_op_mlu.cc
浏览文件 @
d876952d
...
@@ -91,9 +91,9 @@ class GatherGradOpMLUKernel : public framework::OpKernel<T> {
...
@@ -91,9 +91,9 @@ class GatherGradOpMLUKernel : public framework::OpKernel<T> {
ToCnnlDataType
(
index
->
dtype
()));
ToCnnlDataType
(
index
->
dtype
()));
MLUCnnlTensorDesc
dout_desc
(
*
dout
);
MLUCnnlTensorDesc
dout_desc
(
*
dout
);
const
cnnlScatterRefMode_t
mode
=
CNNL_SCATTERREF_UPDATE
;
const
cnnlScatterRefMode_t
mode
=
CNNL_SCATTERREF_UPDATE
;
MLUCnnl
::
Scatter
Functor
(
ctx
,
dx_desc
.
get
(),
GetBasePtr
(
dx
),
dout_desc
.
get
(
),
MLUCnnl
::
Scatter
RefFunctor
(
ctx
,
dx_desc
.
get
(),
GetBasePtr
(
dx
),
GetBasePtr
(
dout
),
index_desc
.
get
(
),
dout_desc
.
get
(),
GetBasePtr
(
dout
),
GetBasePtr
(
index
),
mode
);
index_desc
.
get
(),
GetBasePtr
(
index
),
mode
);
}
}
};
};
...
...
paddle/fluid/operators/mlu/mlu_baseop.cc
浏览文件 @
d876952d
...
@@ -1134,7 +1134,7 @@ MLUCnnlDCNDesc::~MLUCnnlDCNDesc() {
...
@@ -1134,7 +1134,7 @@ MLUCnnlDCNDesc::~MLUCnnlDCNDesc() {
indices_desc
,
indices
,
output_desc
,
output
));
indices_desc
,
indices
,
output_desc
,
output
));
}
}
/* static */
void
MLUCnnl
::
ScatterFunctor
(
/* static */
void
MLUCnnl
::
Scatter
Ref
Functor
(
const
ExecutionContext
&
ctx
,
const
cnnlTensorDescriptor_t
params_desc
,
const
ExecutionContext
&
ctx
,
const
cnnlTensorDescriptor_t
params_desc
,
const
void
*
params
,
const
cnnlTensorDescriptor_t
updates_desc
,
const
void
*
params
,
const
cnnlTensorDescriptor_t
updates_desc
,
const
void
*
updates
,
const
cnnlTensorDescriptor_t
indices_desc
,
const
void
*
updates
,
const
cnnlTensorDescriptor_t
indices_desc
,
...
@@ -1146,6 +1146,19 @@ MLUCnnlDCNDesc::~MLUCnnlDCNDesc() {
...
@@ -1146,6 +1146,19 @@ MLUCnnlDCNDesc::~MLUCnnlDCNDesc() {
updates
,
0
,
mode
));
updates
,
0
,
mode
));
}
}
/* static */
void
MLUCnnl
::
ScatterFunctor
(
const
ExecutionContext
&
ctx
,
const
cnnlTensorDescriptor_t
params_desc
,
void
*
params
,
const
cnnlTensorDescriptor_t
updates_desc
,
const
void
*
updates
,
const
cnnlTensorDescriptor_t
indices_desc
,
const
void
*
indices
,
const
int
dim
,
const
cnnlScatterMode_t
mode
)
{
cnnlHandle_t
handle
=
GetHandleFromCTX
(
ctx
);
PADDLE_ENFORCE_MLU_SUCCESS
(
cnnlScatter
(
handle
,
dim
,
params_desc
,
params
,
indices_desc
,
indices
,
updates_desc
,
updates
,
params_desc
,
params
,
/* output_desc, output, same with params*/
mode
));
}
/* static */
void
MLUCnnl
::
StridedSliceGrad
(
/* static */
void
MLUCnnl
::
StridedSliceGrad
(
const
ExecutionContext
&
ctx
,
const
int
begin
[],
const
int
end
[],
const
ExecutionContext
&
ctx
,
const
int
begin
[],
const
int
end
[],
const
int
strides
[],
const
cnnlTensorDescriptor_t
input_desc
,
const
int
strides
[],
const
cnnlTensorDescriptor_t
input_desc
,
...
...
paddle/fluid/operators/mlu/mlu_baseop.h
浏览文件 @
d876952d
...
@@ -626,12 +626,21 @@ class MLUCnnl {
...
@@ -626,12 +626,21 @@ class MLUCnnl {
const
cnnlTensorDescriptor_t
indices_desc
,
const
void
*
indices
,
const
cnnlTensorDescriptor_t
indices_desc
,
const
void
*
indices
,
const
cnnlTensorDescriptor_t
output_desc
,
void
*
output
);
const
cnnlTensorDescriptor_t
output_desc
,
void
*
output
);
static
void
ScatterFunctor
(
static
void
Scatter
Ref
Functor
(
const
ExecutionContext
&
ctx
,
const
cnnlTensorDescriptor_t
params_desc
,
const
ExecutionContext
&
ctx
,
const
cnnlTensorDescriptor_t
params_desc
,
const
void
*
params
,
const
cnnlTensorDescriptor_t
updates_desc
,
const
void
*
params
,
const
cnnlTensorDescriptor_t
updates_desc
,
const
void
*
updates
,
const
cnnlTensorDescriptor_t
indices_desc
,
const
void
*
updates
,
const
cnnlTensorDescriptor_t
indices_desc
,
const
void
*
indices
,
const
cnnlScatterRefMode_t
mode
);
const
void
*
indices
,
const
cnnlScatterRefMode_t
mode
);
static
void
ScatterFunctor
(
const
ExecutionContext
&
ctx
,
const
cnnlTensorDescriptor_t
params_desc
,
const
void
*
params
,
const
cnnlTensorDescriptor_t
updates_desc
,
const
void
*
updates
,
const
cnnlTensorDescriptor_t
indices_desc
,
const
void
*
indices
,
const
int
dim
,
const
cnnlScatterMode_t
mode
=
CNNL_SCATTER
);
static
void
Range
(
const
ExecutionContext
&
ctx
,
const
void
*
start
,
static
void
Range
(
const
ExecutionContext
&
ctx
,
const
void
*
start
,
const
void
*
end
,
const
void
*
step
,
const
void
*
end
,
const
void
*
step
,
const
cnnlDataType_t
output_dtype
,
void
*
output
);
const
cnnlDataType_t
output_dtype
,
void
*
output
);
...
...
python/paddle/fluid/tests/unittests/mlu/test_argsort_op_mlu.py
0 → 100644
浏览文件 @
d876952d
# 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.
from
__future__
import
print_function
import
paddle
import
numpy
as
np
import
unittest
import
sys
sys
.
path
.
append
(
".."
)
from
op_test
import
OpTest
paddle
.
enable_static
()
SEED
=
2022
def
gen_test_class
(
dtype
,
axis
,
descending
):
class
TestArgsortOp
(
OpTest
):
def
setUp
(
self
):
np
.
random
.
seed
(
SEED
)
self
.
set_mlu
()
self
.
op_type
=
"argsort"
self
.
place
=
paddle
.
MLUPlace
(
0
)
self
.
init_inputshape
()
if
'int'
in
dtype
:
self
.
x
=
np
.
random
.
choice
(
255
,
self
.
size
,
replace
=
False
)
self
.
x
=
self
.
x
.
reshape
(
self
.
input_shape
).
astype
(
dtype
)
else
:
self
.
x
=
np
.
random
.
random
(
self
.
input_shape
).
astype
(
dtype
)
self
.
inputs
=
{
"X"
:
self
.
x
}
self
.
attrs
=
{
"axis"
:
axis
,
"descending"
:
descending
}
self
.
get_output
()
self
.
outputs
=
{
"Out"
:
self
.
sorted_x
,
"Indices"
:
self
.
indices
}
def
get_output
(
self
):
if
descending
:
self
.
indices
=
np
.
flip
(
np
.
argsort
(
self
.
x
,
kind
=
'heapsort'
,
axis
=
axis
),
axis
)
self
.
sorted_x
=
np
.
flip
(
np
.
sort
(
self
.
x
,
kind
=
'heapsort'
,
axis
=
axis
),
axis
)
else
:
self
.
indices
=
np
.
argsort
(
self
.
x
,
kind
=
'heapsort'
,
axis
=
axis
)
self
.
sorted_x
=
np
.
sort
(
self
.
x
,
kind
=
'heapsort'
,
axis
=
axis
)
def
test_check_grad
(
self
):
if
dtype
in
[
'float16'
,
'int8'
,
'uint8'
,
'int32'
]:
self
.
__class__
.
no_need_check_grad
=
True
else
:
self
.
check_grad_with_place
(
self
.
place
,
[
"X"
],
"Out"
)
def
set_mlu
(
self
):
self
.
__class__
.
use_mlu
=
True
def
init_inputshape
(
self
):
self
.
input_shape
=
(
5
,
2
,
2
,
3
,
3
)
self
.
size
=
np
.
prod
(
self
.
input_shape
)
def
test_check_output
(
self
):
self
.
check_output_with_place
(
self
.
place
)
def
init_direction
(
self
):
self
.
descending
=
False
cls_name
=
"{}_{}_{}_TestArgsortOp"
.
format
(
dtype
,
axis
,
descending
)
TestArgsortOp
.
__name__
=
cls_name
globals
()[
cls_name
]
=
TestArgsortOp
for
dtype
in
[
'float32'
,
'float16'
,
'int8'
,
'uint8'
,
'int32'
]:
for
axis
in
[
1
,
2
,
3
,
-
1
]:
for
descending
in
[
False
]:
gen_test_class
(
dtype
,
axis
,
descending
)
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