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
9ad05afd
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
9ad05afd
编写于
6月 10, 2022
作者:
F
fuyou765
提交者:
GitHub
6月 10, 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 mlu kernel for scatter op (#43292)
上级
ac75617a
变更
2
隐藏空白更改
内联
并排
Showing
2 changed file
with
313 addition
and
0 deletion
+313
-0
paddle/fluid/operators/scatter_op_mlu.cc
paddle/fluid/operators/scatter_op_mlu.cc
+66
-0
python/paddle/fluid/tests/unittests/mlu/test_scatter_op_mlu.py
...n/paddle/fluid/tests/unittests/mlu/test_scatter_op_mlu.py
+247
-0
未找到文件。
paddle/fluid/operators/scatter_op_mlu.cc
0 → 100644
浏览文件 @
9ad05afd
/* 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
ScatterMLUKernel
:
public
framework
::
OpKernel
<
T
>
{
public:
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
auto
*
x
=
ctx
.
Input
<
Tensor
>
(
"X"
);
auto
*
indices
=
ctx
.
Input
<
Tensor
>
(
"Ids"
);
auto
*
updates
=
ctx
.
Input
<
Tensor
>
(
"Updates"
);
bool
overwrite
=
ctx
.
Attr
<
bool
>
(
"overwrite"
);
auto
*
out
=
ctx
.
Output
<
Tensor
>
(
"Out"
);
auto
place
=
ctx
.
GetPlace
();
out
->
mutable_data
<
T
>
(
place
);
MLUCnnlTensorDesc
x_desc
(
*
x
);
MLUCnnlTensorDesc
indices_desc
(
*
indices
);
MLUCnnlTensorDesc
updates_desc
(
*
updates
);
MLUCnnlTensorDesc
out_desc
(
*
out
);
cnnlScatterRefMode_t
mode
;
if
(
overwrite
)
{
mode
=
CNNL_SCATTERREF_UPDATE
;
MLUCnnl
::
ScatterFunctor
(
ctx
,
x_desc
.
get
(),
GetBasePtr
(
x
),
updates_desc
.
get
(),
GetBasePtr
(
updates
),
indices_desc
.
get
(),
GetBasePtr
(
indices
),
mode
);
}
else
{
Tensor
tensor_zeros
(
updates
->
type
());
tensor_zeros
.
mutable_data
<
T
>
(
updates
->
dims
(),
ctx
.
GetPlace
());
MLUCnnlTensorDesc
tensor_zeros_desc
(
tensor_zeros
);
float
value
=
0.0
;
auto
value_t
=
static_cast
<
T
>
(
value
);
MLUCnnl
::
Fill
(
ctx
,
CNNL_POINTER_MODE_HOST
,
&
value_t
,
tensor_zeros_desc
.
get
(),
GetBasePtr
(
&
tensor_zeros
));
mode
=
CNNL_SCATTERREF_UPDATE
;
MLUCnnl
::
ScatterFunctor
(
ctx
,
x_desc
.
get
(),
GetBasePtr
(
x
),
tensor_zeros_desc
.
get
(),
GetBasePtr
(
&
tensor_zeros
),
indices_desc
.
get
(),
GetBasePtr
(
indices
),
mode
);
mode
=
CNNL_SCATTERREF_ADD
;
MLUCnnl
::
ScatterFunctor
(
ctx
,
x_desc
.
get
(),
GetBasePtr
(
x
),
updates_desc
.
get
(),
GetBasePtr
(
updates
),
indices_desc
.
get
(),
GetBasePtr
(
indices
),
mode
);
}
paddle
::
framework
::
TensorCopy
(
*
x
,
place
,
out
);
}
};
}
// namespace operators
}
// namespace paddle
namespace
ops
=
paddle
::
operators
;
REGISTER_OP_MLU_KERNEL
(
scatter
,
ops
::
ScatterMLUKernel
<
float
>
,
ops
::
ScatterMLUKernel
<
paddle
::
platform
::
float16
>
);
python/paddle/fluid/tests/unittests/mlu/test_scatter_op_mlu.py
0 → 100644
浏览文件 @
9ad05afd
# 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
sys
sys
.
path
.
append
(
".."
)
import
unittest
import
numpy
as
np
import
os
import
paddle
import
paddle.fluid
as
fluid
from
op_test
import
OpTest
import
paddle.fluid.core
as
core
from
paddle.fluid.dygraph.base
import
switch_to_static_graph
class
TestScatterOp
(
OpTest
):
def
setUp
(
self
):
self
.
op_type
=
"scatter"
self
.
place
=
paddle
.
device
.
MLUPlace
(
0
)
self
.
__class__
.
use_mlu
=
True
self
.
python_api
=
paddle
.
scatter
ref_np
=
np
.
ones
((
3
,
50
)).
astype
(
"float32"
)
index_np
=
np
.
array
([
1
,
2
]).
astype
(
"int32"
)
updates_np
=
np
.
random
.
random
((
2
,
50
)).
astype
(
"float32"
)
output_np
=
np
.
copy
(
ref_np
)
output_np
[
index_np
]
=
updates_np
self
.
inputs
=
{
'X'
:
ref_np
,
'Ids'
:
index_np
,
'Updates'
:
updates_np
}
self
.
outputs
=
{
'Out'
:
output_np
}
def
test_check_output
(
self
):
self
.
check_output_with_place
(
self
.
place
,
check_eager
=
False
)
def
test_check_grad
(
self
):
self
.
check_grad
([
"X"
,
"Updates"
],
"Out"
,
check_eager
=
False
)
class
TestScatterOp0
(
OpTest
):
def
setUp
(
self
):
self
.
op_type
=
"scatter"
self
.
place
=
paddle
.
device
.
MLUPlace
(
0
)
self
.
__class__
.
use_mlu
=
True
self
.
python_api
=
paddle
.
scatter
ref_np
=
np
.
ones
((
3
,
3
)).
astype
(
"float32"
)
index_np
=
np
.
array
([
1
,
2
]).
astype
(
"int32"
)
updates_np
=
np
.
random
.
random
((
2
,
3
)).
astype
(
"float32"
)
output_np
=
np
.
copy
(
ref_np
)
output_np
[
index_np
]
=
updates_np
self
.
inputs
=
{
'X'
:
ref_np
,
'Ids'
:
index_np
,
'Updates'
:
updates_np
}
self
.
attrs
=
{
'overwrite'
:
True
}
self
.
outputs
=
{
'Out'
:
output_np
}
def
test_check_output
(
self
):
self
.
check_output_with_place
(
self
.
place
,
check_eager
=
False
)
def
test_check_grad
(
self
):
self
.
check_grad
([
"X"
,
"Updates"
],
"Out"
,
check_eager
=
False
)
class
TestScatterOp1
(
OpTest
):
def
setUp
(
self
):
self
.
op_type
=
"scatter"
self
.
place
=
paddle
.
device
.
MLUPlace
(
0
)
self
.
__class__
.
use_mlu
=
True
self
.
python_api
=
paddle
.
scatter
ref_np
=
np
.
ones
((
3
,
3
)).
astype
(
"float32"
)
zeros_np
=
np
.
zeros
([
2
,
3
]).
astype
(
'float32'
)
index_np
=
np
.
array
([
1
,
1
]).
astype
(
"int32"
)
updates_np
=
np
.
random
.
random
((
2
,
3
)).
astype
(
"float32"
)
output_np
=
np
.
copy
(
ref_np
)
output_np
[
index_np
]
=
zeros_np
for
i
in
range
(
0
,
len
(
index_np
)):
output_np
[
index_np
[
i
]]
+=
updates_np
[
i
]
self
.
attrs
=
{
'overwrite'
:
False
}
self
.
inputs
=
{
'X'
:
ref_np
,
'Ids'
:
index_np
,
'Updates'
:
updates_np
}
self
.
outputs
=
{
'Out'
:
output_np
}
def
test_check_output
(
self
):
self
.
check_output_with_place
(
self
.
place
,
check_eager
=
False
)
def
test_check_grad
(
self
):
self
.
check_grad
([
"X"
,
"Updates"
],
"Out"
,
check_eager
=
False
)
class
TestScatterOp2
(
OpTest
):
def
setUp
(
self
):
self
.
op_type
=
"scatter"
self
.
place
=
paddle
.
device
.
MLUPlace
(
0
)
self
.
__class__
.
use_mlu
=
True
self
.
python_api
=
paddle
.
scatter
ref_np
=
np
.
ones
((
3
,
3
)).
astype
(
"float32"
)
index_np
=
np
.
array
([
1
,
2
]).
astype
(
"int64"
)
updates_np
=
np
.
random
.
random
((
2
,
3
)).
astype
(
"float32"
)
output_np
=
np
.
copy
(
ref_np
)
output_np
[
index_np
]
=
updates_np
self
.
inputs
=
{
'X'
:
ref_np
,
'Ids'
:
index_np
,
'Updates'
:
updates_np
}
self
.
outputs
=
{
'Out'
:
output_np
}
def
test_check_output
(
self
):
self
.
check_output_with_place
(
self
.
place
,
check_eager
=
False
)
def
test_check_grad
(
self
):
self
.
check_grad
([
'X'
,
'Updates'
],
'Out'
,
check_eager
=
False
)
class
TestScatterAPI
(
unittest
.
TestCase
):
def
setUp
(
self
):
self
.
places
=
[
paddle
.
device
.
MLUPlace
(
0
)]
self
.
__class__
.
use_mlu
=
True
self
.
executed_api
()
def
executed_api
(
self
):
self
.
scatter
=
paddle
.
scatter
def
check_static_result
(
self
,
place
):
with
fluid
.
program_guard
(
fluid
.
Program
(),
fluid
.
Program
()):
input
=
fluid
.
data
(
name
=
"input"
,
shape
=
[
3
,
2
],
dtype
=
"float32"
)
index
=
fluid
.
data
(
name
=
"index"
,
shape
=
[
4
],
dtype
=
"int64"
)
updates
=
fluid
.
data
(
name
=
"updates"
,
shape
=
[
4
,
2
],
dtype
=
"float32"
)
result
=
self
.
scatter
(
input
,
index
,
updates
,
False
)
input_data
=
np
.
array
([[
1
,
1
],
[
2
,
2
],
[
3
,
3
]]).
astype
(
np
.
float32
)
index_data
=
np
.
array
([
2
,
1
,
0
,
1
]).
astype
(
np
.
int64
)
updates_data
=
np
.
array
([[
1
,
1
],
[
2
,
2
],
[
3
,
3
],
[
4
,
4
]]).
astype
(
np
.
float32
)
exe
=
fluid
.
Executor
(
place
)
fetches
=
exe
.
run
(
fluid
.
default_main_program
(),
feed
=
{
"input"
:
input_data
,
"index"
:
index_data
,
"updates"
:
updates_data
},
fetch_list
=
[
result
])
self
.
assertEqual
((
fetches
[
0
]
==
\
np
.
array
([[
3.
,
3.
],[
6.
,
6.
],[
1.
,
1.
]])).
all
(),
True
)
def
test_static
(
self
):
for
place
in
self
.
places
:
self
.
check_static_result
(
place
=
place
)
def
test_dygraph
(
self
):
for
place
in
self
.
places
:
with
fluid
.
dygraph
.
guard
(
place
):
x_data
=
np
.
array
([[
1
,
1
],
[
2
,
2
],
[
3
,
3
]]).
astype
(
np
.
float32
)
index_data
=
np
.
array
([
2
,
1
,
0
,
1
]).
astype
(
np
.
int64
)
updates_data
=
np
.
array
([[
1
,
1
],
[
2
,
2
],
[
3
,
3
],
[
4
,
4
]]).
astype
(
np
.
float32
)
x
=
fluid
.
dygraph
.
to_variable
(
x_data
)
index
=
fluid
.
dygraph
.
to_variable
(
index_data
)
updates
=
fluid
.
dygraph
.
to_variable
(
updates_data
)
output1
=
self
.
scatter
(
x
,
index
,
updates
,
overwrite
=
False
)
self
.
assertEqual
((
output1
.
numpy
()
==
\
np
.
array
([[
3.
,
3.
],[
6.
,
6.
],[
1.
,
1.
]])).
all
(),
True
)
def
test_large_data
(
self
):
if
os
.
name
==
"nt"
:
return
x
=
np
.
random
.
rand
(
183826
,
256
).
astype
(
"float32"
)
index
=
np
.
ones
(
8388608
,
dtype
=
"int64"
)
updates
=
np
.
ones
(
shape
=
[
8388608
,
256
],
dtype
=
"float32"
)
def
test_dygraph
():
with
fluid
.
dygraph
.
guard
():
mlu_out
=
paddle
.
scatter
(
paddle
.
to_tensor
(
x
),
paddle
.
to_tensor
(
index
),
paddle
.
to_tensor
(
updates
))
return
mlu_out
.
numpy
()
@
switch_to_static_graph
def
test_static_graph
():
with
paddle
.
static
.
program_guard
(
paddle
.
static
.
Program
(),
paddle
.
static
.
Program
()):
x_t
=
paddle
.
static
.
data
(
name
=
"x"
,
dtype
=
x
.
dtype
,
shape
=
x
.
shape
)
index_t
=
paddle
.
static
.
data
(
name
=
"index"
,
dtype
=
index
.
dtype
,
shape
=
index
.
shape
)
updates_t
=
paddle
.
static
.
data
(
name
=
"updates"
,
dtype
=
updates
.
dtype
,
shape
=
updates
.
shape
)
out_t
=
paddle
.
scatter
(
x_t
,
index_t
,
updates_t
)
feed
=
{
x_t
.
name
:
x
,
index_t
.
name
:
index
,
updates_t
.
name
:
updates
}
fetch
=
[
out_t
]
mlu_exe
=
paddle
.
static
.
Executor
(
paddle
.
device
.
MLUPlace
(
0
))
mlu_value
=
mlu_exe
.
run
(
feed
=
feed
,
fetch_list
=
fetch
)[
0
]
return
mlu_value
self
.
assertTrue
(
np
.
array_equal
(
test_dygraph
(),
test_static_graph
()))
class
TestScatterOpFp16
(
OpTest
):
def
setUp
(
self
):
self
.
op_type
=
"scatter"
self
.
place
=
paddle
.
device
.
MLUPlace
(
0
)
self
.
__class__
.
use_mlu
=
True
self
.
python_api
=
paddle
.
scatter
ref_np
=
np
.
ones
((
3
,
3
)).
astype
(
"float16"
)
index_np
=
np
.
array
([
1
,
2
]).
astype
(
"int32"
)
updates_np
=
np
.
random
.
random
((
2
,
3
)).
astype
(
"float16"
)
output_np
=
np
.
copy
(
ref_np
)
output_np
[
index_np
]
=
updates_np
self
.
inputs
=
{
'X'
:
ref_np
,
'Ids'
:
index_np
,
'Updates'
:
updates_np
}
self
.
attrs
=
{
'overwrite'
:
True
}
self
.
outputs
=
{
'Out'
:
output_np
}
def
test_check_output
(
self
):
self
.
check_output_with_place
(
self
.
place
,
check_eager
=
False
)
def
test_check_grad
(
self
):
self
.
check_grad
([
"X"
,
"Updates"
],
"Out"
,
check_eager
=
False
)
class
TestScatterInplaceAPI
(
TestScatterAPI
):
def
executed_api
(
self
):
self
.
scatter
=
paddle
.
scatter_
if
__name__
==
"__main__"
:
paddle
.
enable_static
()
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