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
04d324b2
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
04d324b2
编写于
2月 25, 2022
作者:
F
fwenguang
提交者:
GitHub
2月 25, 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 elementwise_mul mlu kernel (#39864)
上级
0615815d
变更
2
显示空白变更内容
内联
并排
Showing
2 changed file
with
409 addition
and
0 deletion
+409
-0
paddle/fluid/operators/elementwise/elementwise_mul_op_mlu.cc
paddle/fluid/operators/elementwise/elementwise_mul_op_mlu.cc
+169
-0
python/paddle/fluid/tests/unittests/mlu/test_elementwise_mul_op_mlu.py
.../fluid/tests/unittests/mlu/test_elementwise_mul_op_mlu.py
+240
-0
未找到文件。
paddle/fluid/operators/elementwise/elementwise_mul_op_mlu.cc
0 → 100644
浏览文件 @
04d324b2
/* 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/operators/elementwise/elementwise_mul_op.h"
#include "paddle/fluid/operators/mlu/mlu_baseop.h"
namespace
paddle
{
namespace
operators
{
using
Tensor
=
framework
::
Tensor
;
using
MLUDeviceContext
=
platform
::
MLUDeviceContext
;
static
void
GetReduceAxes
(
const
int
axis
,
const
framework
::
DDim
&
src_ddims
,
const
framework
::
DDim
&
target_ddims
,
std
::
vector
<
int
>*
axes
)
{
int64_t
src_dim_size
=
src_ddims
.
size
();
int64_t
target_dim_size
=
target_ddims
.
size
();
for
(
int64_t
i
=
0
;
i
<
src_dim_size
;
++
i
)
{
if
(
i
<
axis
||
i
>=
target_dim_size
+
axis
)
{
axes
->
push_back
(
i
);
continue
;
}
if
(
src_ddims
[
i
]
>
target_ddims
[
i
-
axis
])
{
axes
->
push_back
(
i
);
}
}
}
template
<
typename
T
>
class
ElementwiseMulMLUKernel
:
public
framework
::
OpKernel
<
T
>
{
public:
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
auto
*
x
=
ctx
.
Input
<
Tensor
>
(
"X"
);
auto
*
y
=
ctx
.
Input
<
Tensor
>
(
"Y"
);
auto
*
out
=
ctx
.
Output
<
Tensor
>
(
"Out"
);
out
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
int
axis
=
ctx
.
Attr
<
int
>
(
"axis"
);
const
auto
&
x_dims
=
x
->
dims
();
const
auto
&
y_dims
=
y
->
dims
();
axis
=
(
axis
<
0
?
(
std
::
abs
(
x_dims
.
size
()
-
y_dims
.
size
())
+
axis
+
1
)
:
axis
);
int
max_dim
=
std
::
max
(
x_dims
.
size
(),
y_dims
.
size
());
std
::
vector
<
int
>
x_dims_array
(
max_dim
);
std
::
vector
<
int
>
y_dims_array
(
max_dim
);
std
::
vector
<
int
>
out_dims_array
(
max_dim
);
GetBroadcastDimsArrays
(
x_dims
,
y_dims
,
x_dims_array
.
data
(),
y_dims_array
.
data
(),
out_dims_array
.
data
(),
max_dim
,
axis
);
MLUCnnlTensorDesc
x_desc
(
max_dim
,
x_dims_array
.
data
(),
ToCnnlDataType
<
T
>
());
MLUCnnlTensorDesc
y_desc
(
max_dim
,
y_dims_array
.
data
(),
ToCnnlDataType
<
T
>
());
MLUCnnlTensorDesc
out_desc
(
*
out
);
MLUCnnlOpTensorDesc
op_tensor_desc
(
CNNL_OP_TENSOR_MUL
,
ToCnnlDataType
<
T
>
(),
CNNL_NOT_PROPAGATE_NAN
);
MLUCnnl
::
OpTensor
(
ctx
,
op_tensor_desc
.
get
(),
x_desc
.
get
(),
GetBasePtr
(
x
),
y_desc
.
get
(),
GetBasePtr
(
y
),
out_desc
.
get
(),
GetBasePtr
(
out
),
ToCnnlDataType
<
T
>
());
}
};
template
<
typename
T
>
class
ElementwiseMulGradMLUKernel
:
public
framework
::
OpKernel
<
T
>
{
public:
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
auto
*
x
=
ctx
.
Input
<
Tensor
>
(
"X"
);
auto
*
y
=
ctx
.
Input
<
Tensor
>
(
"Y"
);
auto
*
dout
=
ctx
.
Input
<
Tensor
>
(
framework
::
GradVarName
(
"Out"
));
auto
*
dx
=
ctx
.
Output
<
Tensor
>
(
framework
::
GradVarName
(
"X"
));
auto
*
dy
=
ctx
.
Output
<
Tensor
>
(
framework
::
GradVarName
(
"Y"
));
int
axis
=
ctx
.
Attr
<
int
>
(
"axis"
);
const
auto
&
x_dims
=
x
->
dims
();
const
auto
&
y_dims
=
y
->
dims
();
axis
=
(
axis
<
0
?
(
std
::
abs
(
x_dims
.
size
()
-
y_dims
.
size
())
+
axis
+
1
)
:
axis
);
int
max_dim
=
std
::
max
(
x_dims
.
size
(),
y_dims
.
size
());
std
::
vector
<
int
>
x_dims_array
(
max_dim
);
std
::
vector
<
int
>
y_dims_array
(
max_dim
);
std
::
vector
<
int
>
out_dims_array
(
max_dim
);
GetBroadcastDimsArrays
(
x_dims
,
y_dims
,
x_dims_array
.
data
(),
y_dims_array
.
data
(),
out_dims_array
.
data
(),
max_dim
,
axis
);
MLUCnnlTensorDesc
x_desc
(
max_dim
,
x_dims_array
.
data
(),
ToCnnlDataType
<
T
>
());
MLUCnnlTensorDesc
y_desc
(
max_dim
,
y_dims_array
.
data
(),
ToCnnlDataType
<
T
>
());
MLUCnnlTensorDesc
dout_desc
(
*
dout
);
MLUCnnlOpTensorDesc
mul_op_desc
(
CNNL_OP_TENSOR_MUL
,
ToCnnlDataType
<
T
>
(),
CNNL_NOT_PROPAGATE_NAN
);
if
(
dx
)
{
dx
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
if
(
dx
->
dims
()
==
dout
->
dims
())
{
MLUCnnl
::
OpTensor
(
ctx
,
mul_op_desc
.
get
(),
dout_desc
.
get
(),
GetBasePtr
(
dout
),
y_desc
.
get
(),
GetBasePtr
(
y
),
x_desc
.
get
(),
GetBasePtr
(
dx
),
ToCnnlDataType
<
T
>
());
}
else
{
Tensor
dx_temp
(
x
->
dtype
());
dx_temp
.
Resize
(
dout
->
dims
());
dx_temp
.
mutable_data
<
T
>
(
ctx
.
GetPlace
());
MLUCnnl
::
OpTensor
(
ctx
,
mul_op_desc
.
get
(),
dout_desc
.
get
(),
GetBasePtr
(
dout
),
y_desc
.
get
(),
GetBasePtr
(
y
),
dout_desc
.
get
(),
GetBasePtr
(
&
dx_temp
),
ToCnnlDataType
<
T
>
());
std
::
vector
<
int
>
reduce_axes
;
GetReduceAxes
(
axis
,
dx_temp
.
dims
(),
dx
->
dims
(),
&
reduce_axes
);
MLUCnnlReduceDesc
reduction_desc
(
reduce_axes
,
CNNL_REDUCE_ADD
,
ToCnnlDataType
<
T
>
(),
CNNL_NOT_PROPAGATE_NAN
,
CNNL_REDUCE_NO_INDICES
,
CNNL_32BIT_INDICES
);
MLUCnnlTensorDesc
dx_desc
(
*
dx
);
MLUCnnl
::
Reduce
(
ctx
,
true
/*need_workspace*/
,
reduction_desc
.
get
(),
nullptr
,
dout_desc
.
get
(),
GetBasePtr
(
&
dx_temp
),
0
,
nullptr
,
nullptr
,
dx_desc
.
get
(),
GetBasePtr
(
dx
));
}
}
if
(
dy
)
{
dy
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
if
(
dy
->
dims
()
==
dout
->
dims
())
{
MLUCnnl
::
OpTensor
(
ctx
,
mul_op_desc
.
get
(),
dout_desc
.
get
(),
GetBasePtr
(
dout
),
x_desc
.
get
(),
GetBasePtr
(
x
),
y_desc
.
get
(),
GetBasePtr
(
dy
),
ToCnnlDataType
<
T
>
());
}
else
{
Tensor
dy_temp
(
y
->
dtype
());
dy_temp
.
Resize
(
dout
->
dims
());
dy_temp
.
mutable_data
<
T
>
(
ctx
.
GetPlace
());
MLUCnnl
::
OpTensor
(
ctx
,
mul_op_desc
.
get
(),
dout_desc
.
get
(),
GetBasePtr
(
dout
),
x_desc
.
get
(),
GetBasePtr
(
x
),
dout_desc
.
get
(),
GetBasePtr
(
&
dy_temp
),
ToCnnlDataType
<
T
>
());
std
::
vector
<
int
>
reduce_axes
;
GetReduceAxes
(
axis
,
dy_temp
.
dims
(),
dy
->
dims
(),
&
reduce_axes
);
MLUCnnlReduceDesc
reduction_desc
(
reduce_axes
,
CNNL_REDUCE_ADD
,
ToCnnlDataType
<
T
>
(),
CNNL_NOT_PROPAGATE_NAN
,
CNNL_REDUCE_NO_INDICES
,
CNNL_32BIT_INDICES
);
MLUCnnlTensorDesc
dy_desc
(
*
dy
);
MLUCnnl
::
Reduce
(
ctx
,
true
/*need_workspace*/
,
reduction_desc
.
get
(),
nullptr
,
dout_desc
.
get
(),
GetBasePtr
(
&
dy_temp
),
0
,
nullptr
,
nullptr
,
dy_desc
.
get
(),
GetBasePtr
(
dy
));
}
}
}
};
}
// namespace operators
}
// namespace paddle
namespace
ops
=
paddle
::
operators
;
REGISTER_OP_MLU_KERNEL
(
elementwise_mul
,
ops
::
ElementwiseMulMLUKernel
<
float
>
,
ops
::
ElementwiseMulMLUKernel
<
paddle
::
platform
::
float16
>
,
ops
::
ElementwiseMulMLUKernel
<
int
>
);
REGISTER_OP_MLU_KERNEL
(
elementwise_mul_grad
,
ops
::
ElementwiseMulGradMLUKernel
<
float
>
,
ops
::
ElementwiseMulGradMLUKernel
<
paddle
::
platform
::
float16
>
,
ops
::
ElementwiseMulGradMLUKernel
<
int
>
);
python/paddle/fluid/tests/unittests/mlu/test_elementwise_mul_op_mlu.py
0 → 100644
浏览文件 @
04d324b2
# 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
unittest
import
numpy
as
np
import
paddle
import
paddle.fluid
as
fluid
import
paddle.fluid.core
as
core
from
paddle.fluid
import
Program
,
compiler
,
program_guard
from
paddle.fluid.op
import
Operator
import
sys
sys
.
path
.
append
(
'..'
)
from
op_test
import
OpTest
,
skip_check_grad_ci
paddle
.
enable_static
()
class
ElementwiseMulOp
(
OpTest
):
def
init_kernel_type
(
self
):
self
.
__class__
.
use_mlu
=
True
self
.
place
=
paddle
.
device
.
MLUPlace
(
0
)
def
setUp
(
self
):
self
.
op_type
=
"elementwise_mul"
self
.
dtype
=
np
.
float32
self
.
axis
=
-
1
self
.
init_dtype
()
self
.
init_input_output
()
self
.
init_kernel_type
()
self
.
init_axis
()
self
.
inputs
=
{
'X'
:
OpTest
.
np_dtype_to_fluid_dtype
(
self
.
x
),
'Y'
:
OpTest
.
np_dtype_to_fluid_dtype
(
self
.
y
)
}
self
.
outputs
=
{
'Out'
:
self
.
out
}
self
.
attrs
=
{
'axis'
:
self
.
axis
}
def
test_check_output
(
self
):
self
.
check_output_with_place
(
self
.
place
)
def
test_check_grad_normal
(
self
):
self
.
check_grad_with_place
(
self
.
place
,
[
'X'
,
'Y'
],
'Out'
)
def
test_check_grad_ingore_x
(
self
):
self
.
check_grad_with_place
(
self
.
place
,
[
'Y'
],
'Out'
,
no_grad_set
=
set
(
"X"
))
def
test_check_grad_ingore_y
(
self
):
self
.
check_grad_with_place
(
self
.
place
,
[
'X'
],
'Out'
,
no_grad_set
=
set
(
'Y'
))
def
init_input_output
(
self
):
self
.
x
=
np
.
random
.
uniform
(
0.1
,
1
,
[
13
,
17
]).
astype
(
self
.
dtype
)
self
.
y
=
np
.
random
.
uniform
(
0.1
,
1
,
[
13
,
17
]).
astype
(
self
.
dtype
)
self
.
out
=
np
.
multiply
(
self
.
x
,
self
.
y
)
def
init_dtype
(
self
):
pass
def
init_axis
(
self
):
pass
@
skip_check_grad_ci
(
reason
=
"[skip shape check] Use y_shape(1) to test broadcast."
)
class
TestElementwiseMulOp_scalar
(
ElementwiseMulOp
):
def
setUp
(
self
):
self
.
op_type
=
"elementwise_mul"
self
.
inputs
=
{
'X'
:
np
.
random
.
rand
(
10
,
3
,
4
).
astype
(
np
.
float32
),
'Y'
:
np
.
random
.
rand
(
1
).
astype
(
np
.
float32
)
}
self
.
outputs
=
{
'Out'
:
self
.
inputs
[
'X'
]
*
self
.
inputs
[
'Y'
]}
self
.
init_kernel_type
()
class
TestElementwiseMulOp_Vector
(
ElementwiseMulOp
):
def
setUp
(
self
):
self
.
op_type
=
"elementwise_mul"
self
.
inputs
=
{
'X'
:
np
.
random
.
random
((
100
,
)).
astype
(
"float32"
),
'Y'
:
np
.
random
.
random
((
100
,
)).
astype
(
"float32"
)
}
self
.
outputs
=
{
'Out'
:
np
.
multiply
(
self
.
inputs
[
'X'
],
self
.
inputs
[
'Y'
])}
self
.
init_kernel_type
()
class
TestElementwiseMulOp_broadcast_0
(
ElementwiseMulOp
):
def
init_input_output
(
self
):
self
.
x
=
np
.
random
.
rand
(
100
,
2
,
3
).
astype
(
self
.
dtype
)
self
.
y
=
np
.
random
.
rand
(
100
).
astype
(
self
.
dtype
)
self
.
out
=
self
.
x
*
self
.
y
.
reshape
(
100
,
1
,
1
)
def
init_axis
(
self
):
self
.
axis
=
0
class
TestElementwiseMulOp_broadcast_1
(
ElementwiseMulOp
):
def
setUp
(
self
):
self
.
op_type
=
"elementwise_mul"
self
.
inputs
=
{
'X'
:
np
.
random
.
rand
(
2
,
100
,
3
).
astype
(
np
.
float32
),
'Y'
:
np
.
random
.
rand
(
100
).
astype
(
np
.
float32
)
}
self
.
attrs
=
{
'axis'
:
1
}
self
.
outputs
=
{
'Out'
:
self
.
inputs
[
'X'
]
*
self
.
inputs
[
'Y'
].
reshape
(
1
,
100
,
1
)
}
self
.
init_kernel_type
()
class
TestElementwiseMulOp_broadcast_2
(
ElementwiseMulOp
):
def
setUp
(
self
):
self
.
op_type
=
"elementwise_mul"
self
.
inputs
=
{
'X'
:
np
.
random
.
rand
(
2
,
3
,
100
).
astype
(
np
.
float32
),
'Y'
:
np
.
random
.
rand
(
100
).
astype
(
np
.
float32
)
}
self
.
outputs
=
{
'Out'
:
self
.
inputs
[
'X'
]
*
self
.
inputs
[
'Y'
].
reshape
(
1
,
1
,
100
)
}
self
.
init_kernel_type
()
class
TestElementwiseMulOp_broadcast_3
(
ElementwiseMulOp
):
def
setUp
(
self
):
self
.
op_type
=
"elementwise_mul"
self
.
inputs
=
{
'X'
:
np
.
random
.
rand
(
2
,
10
,
12
,
3
).
astype
(
np
.
float32
),
'Y'
:
np
.
random
.
rand
(
10
,
12
).
astype
(
np
.
float32
)
}
self
.
attrs
=
{
'axis'
:
1
}
self
.
outputs
=
{
'Out'
:
self
.
inputs
[
'X'
]
*
self
.
inputs
[
'Y'
].
reshape
(
1
,
10
,
12
,
1
)
}
self
.
init_kernel_type
()
class
TestElementwiseMulOp_broadcast_4
(
ElementwiseMulOp
):
def
setUp
(
self
):
self
.
op_type
=
"elementwise_mul"
self
.
inputs
=
{
'X'
:
np
.
random
.
rand
(
10
,
2
,
11
).
astype
(
np
.
float32
),
'Y'
:
np
.
random
.
rand
(
10
,
1
,
11
).
astype
(
np
.
float32
)
}
self
.
outputs
=
{
'Out'
:
self
.
inputs
[
'X'
]
*
self
.
inputs
[
'Y'
]}
self
.
init_kernel_type
()
class
TestElementwiseMulOp_broadcast_5
(
ElementwiseMulOp
):
def
setUp
(
self
):
self
.
op_type
=
"elementwise_mul"
self
.
inputs
=
{
'X'
:
np
.
random
.
rand
(
10
,
4
,
2
,
3
).
astype
(
np
.
float32
),
'Y'
:
np
.
random
.
rand
(
10
,
4
,
1
,
3
).
astype
(
np
.
float32
)
}
self
.
outputs
=
{
'Out'
:
self
.
inputs
[
'X'
]
*
self
.
inputs
[
'Y'
]}
self
.
init_kernel_type
()
class
TestElementwiseMulOpFp16
(
ElementwiseMulOp
):
def
init_dtype
(
self
):
self
.
dtype
=
np
.
float16
class
TestElementwiseMulOp_commonuse_1
(
ElementwiseMulOp
):
def
setUp
(
self
):
self
.
op_type
=
"elementwise_mul"
self
.
inputs
=
{
'X'
:
np
.
random
.
rand
(
2
,
3
,
100
).
astype
(
np
.
float32
),
'Y'
:
np
.
random
.
rand
(
1
,
1
,
100
).
astype
(
np
.
float32
)
}
self
.
outputs
=
{
'Out'
:
self
.
inputs
[
'X'
]
*
self
.
inputs
[
'Y'
]}
self
.
init_kernel_type
()
class
TestElementwiseMulOp_commonuse_2
(
ElementwiseMulOp
):
def
setUp
(
self
):
self
.
op_type
=
"elementwise_mul"
self
.
inputs
=
{
'X'
:
np
.
random
.
rand
(
30
,
3
,
1
,
5
).
astype
(
np
.
float32
),
'Y'
:
np
.
random
.
rand
(
30
,
1
,
4
,
1
).
astype
(
np
.
float32
)
}
self
.
outputs
=
{
'Out'
:
self
.
inputs
[
'X'
]
*
self
.
inputs
[
'Y'
]}
self
.
init_kernel_type
()
class
TestElementwiseMulOp_xsize_lessthan_ysize
(
ElementwiseMulOp
):
def
setUp
(
self
):
self
.
op_type
=
"elementwise_mul"
self
.
inputs
=
{
'X'
:
np
.
random
.
rand
(
10
,
10
).
astype
(
np
.
float32
),
'Y'
:
np
.
random
.
rand
(
2
,
2
,
10
,
10
).
astype
(
np
.
float32
)
}
self
.
attrs
=
{
'axis'
:
2
}
self
.
outputs
=
{
'Out'
:
self
.
inputs
[
'X'
].
reshape
(
1
,
1
,
10
,
10
)
*
self
.
inputs
[
'Y'
]
}
self
.
init_kernel_type
()
class
TestElementwiseMulOpError
(
unittest
.
TestCase
):
def
test_errors
(
self
):
with
program_guard
(
Program
(),
Program
()):
# the input of elementwise_mul must be Variable.
x1
=
fluid
.
create_lod_tensor
(
np
.
array
([
-
1
,
3
,
5
,
5
]),
[[
1
,
1
,
1
,
1
]],
fluid
.
CPUPlace
())
y1
=
fluid
.
create_lod_tensor
(
np
.
array
([
-
1
,
3
,
5
,
5
]),
[[
1
,
1
,
1
,
1
]],
fluid
.
CPUPlace
())
self
.
assertRaises
(
TypeError
,
fluid
.
layers
.
elementwise_mul
,
x1
,
y1
)
# the input dtype of elementwise_mul must be float16 or float32 or int32
x2
=
fluid
.
layers
.
data
(
name
=
'x2'
,
shape
=
[
3
,
4
,
5
,
6
],
dtype
=
"uint8"
)
y2
=
fluid
.
layers
.
data
(
name
=
'y2'
,
shape
=
[
3
,
4
,
5
,
6
],
dtype
=
"uint8"
)
self
.
assertRaises
(
TypeError
,
fluid
.
layers
.
elementwise_mul
,
x2
,
y2
)
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