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
sven357
Paddle
提交
10fb4cee
P
Paddle
项目概览
sven357
/
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
10fb4cee
编写于
11月 20, 2018
作者:
T
tensor-tang
提交者:
GitHub
11月 20, 2018
浏览文件
操作
浏览文件
下载
差异文件
Merge pull request #14351 from tpatejko/tpatejko/mkldnn-elementwise_mul
[MKLDNN][JIT][AVX512] Elementwise Mul
上级
05280674
def272cf
变更
10
隐藏空白更改
内联
并排
Showing
10 changed file
with
600 addition
and
14 deletion
+600
-14
AUTHORS.md
AUTHORS.md
+1
-0
paddle/fluid/framework/operator.h
paddle/fluid/framework/operator.h
+1
-0
paddle/fluid/operators/elementwise/elementwise_mul_mkldnn_op.cc
.../fluid/operators/elementwise/elementwise_mul_mkldnn_op.cc
+201
-0
paddle/fluid/operators/elementwise/elementwise_op.h
paddle/fluid/operators/elementwise/elementwise_op.h
+14
-0
paddle/fluid/operators/math/jit_code.h
paddle/fluid/operators/math/jit_code.h
+36
-0
paddle/fluid/operators/math/jit_kernel.h
paddle/fluid/operators/math/jit_kernel.h
+9
-0
paddle/fluid/operators/math/jit_kernel_blas.cc
paddle/fluid/operators/math/jit_kernel_blas.cc
+41
-0
python/paddle/fluid/tests/unittests/op_test.py
python/paddle/fluid/tests/unittests/op_test.py
+3
-1
python/paddle/fluid/tests/unittests/test_elementwise_mul_mkldnn_op.py
...e/fluid/tests/unittests/test_elementwise_mul_mkldnn_op.py
+263
-0
python/paddle/fluid/tests/unittests/test_elementwise_mul_op.py
...n/paddle/fluid/tests/unittests/test_elementwise_mul_op.py
+31
-13
未找到文件。
AUTHORS.md
浏览文件 @
10fb4cee
...
@@ -42,6 +42,7 @@
...
@@ -42,6 +42,7 @@
| QiJune | Jun Qi |
| QiJune | Jun Qi |
| qingqing01 | Qing-Qing Dang |
| qingqing01 | Qing-Qing Dang |
| reyoung | Yang Yu |
| reyoung | Yang Yu |
| Sand3r- | Michal Gallus |
| Superjom | Chun-Wei Yan |
| Superjom | Chun-Wei Yan |
| tensor-tang | Jian Tang |
| tensor-tang | Jian Tang |
| tianbingsz | Tian-Bing Xu |
| tianbingsz | Tian-Bing Xu |
...
...
paddle/fluid/framework/operator.h
浏览文件 @
10fb4cee
...
@@ -100,6 +100,7 @@ class OperatorBase {
...
@@ -100,6 +100,7 @@ class OperatorBase {
const
std
::
string
&
Type
()
const
{
return
type_
;
}
const
std
::
string
&
Type
()
const
{
return
type_
;
}
bool
HasAttr
(
const
std
::
string
&
name
)
const
{
return
attrs_
.
count
(
name
);
}
template
<
typename
T
>
template
<
typename
T
>
inline
const
T
&
Attr
(
const
std
::
string
&
name
)
const
{
inline
const
T
&
Attr
(
const
std
::
string
&
name
)
const
{
PADDLE_ENFORCE
(
attrs_
.
count
(
name
)
!=
0
,
"%s should be in AttributeMap"
,
PADDLE_ENFORCE
(
attrs_
.
count
(
name
)
!=
0
,
"%s should be in AttributeMap"
,
...
...
paddle/fluid/operators/elementwise/elementwise_mul_mkldnn_op.cc
0 → 100644
浏览文件 @
10fb4cee
/* Copyright (c) 2016 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 <mkldnn/include/mkldnn.hpp>
#include "paddle/fluid/operators/elementwise/elementwise_op.h"
#include "paddle/fluid/operators/elementwise/elementwise_op_function.h"
#include "paddle/fluid/platform/mkldnn_helper.h"
#include "paddle/fluid/operators/math/jit_kernel.h"
#include "xbyak.h"
#include "xbyak_util.h"
namespace
paddle
{
namespace
operators
{
using
framework
::
DataLayout
;
using
mkldnn
::
memory
;
static
mkldnn
::
memory
::
format
StringToMKLDNNFormat
(
std
::
string
&
format
)
{
std
::
transform
(
format
.
begin
(),
format
.
end
(),
format
.
begin
(),
::
tolower
);
if
(
!
format
.
compare
(
"nchw"
))
{
return
memory
::
format
::
nchw
;
}
else
if
(
!
format
.
compare
(
"nchw16c"
))
{
return
memory
::
format
::
nChw16c
;
}
else
if
(
!
format
.
compare
(
"nchw8c"
))
{
return
memory
::
format
::
nChw8c
;
}
else
if
(
!
format
.
compare
(
"nhwc"
))
{
return
memory
::
format
::
nhwc
;
}
else
{
return
memory
::
format
::
any
;
}
}
static
void
UpdateDataFormat
(
const
framework
::
ExecutionContext
&
ctx
,
framework
::
Tensor
*
tensor
,
const
char
*
attribute
)
{
if
(
ctx
.
op
().
HasAttr
(
attribute
))
{
auto
format_as_string
=
ctx
.
Attr
<
std
::
string
>
(
attribute
);
auto
format
=
StringToMKLDNNFormat
(
format_as_string
);
if
(
format
!=
memory
::
format
::
any
)
{
tensor
->
set_format
(
format
);
}
}
}
template
<
typename
T
>
static
void
ReorderInput
(
framework
::
Tensor
*
tensor
,
const
platform
::
Place
&
place
,
const
mkldnn
::
engine
&
engine
,
bool
isFourDim
)
{
using
platform
::
to_void_cast
;
auto
dims
=
paddle
::
framework
::
vectorize2int
(
tensor
->
dims
());
framework
::
Tensor
out_tensor
;
out_tensor
.
Resize
(
tensor
->
dims
());
out_tensor
.
set_format
(
isFourDim
?
memory
::
format
::
nchw
:
memory
::
format
::
nc
);
out_tensor
.
set_layout
(
tensor
->
layout
());
mkldnn
::
memory
input_memory
=
{
{{
dims
,
platform
::
MKLDNNGetDataType
<
T
>
(),
tensor
->
format
()},
engine
},
to_void_cast
<
T
>
(
tensor
->
data
<
T
>
())};
mkldnn
::
memory
output_memory
=
{
{{
dims
,
platform
::
MKLDNNGetDataType
<
T
>
(),
out_tensor
.
format
()},
engine
},
to_void_cast
<
T
>
(
out_tensor
.
mutable_data
<
T
>
(
place
))};
platform
::
Reorder
(
input_memory
,
output_memory
);
tensor
->
ShareDataWith
(
out_tensor
);
}
template
<
typename
T
>
class
ElementwiseMulMKLDNNKernel
:
public
framework
::
OpKernel
<
T
>
{
public:
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
using
Tensor
=
framework
::
Tensor
;
int
axis
=
ctx
.
Attr
<
int
>
(
"axis"
);
auto
*
x
=
ctx
.
Input
<
Tensor
>
(
"X"
);
auto
*
y
=
ctx
.
Input
<
Tensor
>
(
"Y"
);
auto
*
z
=
ctx
.
Output
<
Tensor
>
(
"Out"
);
const
T
*
x_data
=
x
->
data
<
T
>
();
const
T
*
y_data
=
y
->
data
<
T
>
();
T
*
z_data
=
z
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
auto
x_dims
=
x
->
dims
();
auto
y_dims_untrimmed
=
y
->
dims
();
auto
x_int_dims
=
paddle
::
framework
::
vectorize2int
(
x_dims
);
UpdateDataFormat
(
ctx
,
(
Tensor
*
)
x
,
"x_data_format"
);
UpdateDataFormat
(
ctx
,
(
Tensor
*
)
y
,
"y_data_format"
);
Xbyak
::
util
::
Cpu
cpu
;
const
bool
is_avx512_enabled
=
cpu
.
has
(
Xbyak
::
util
::
Cpu
::
tAVX512F
);
const
bool
are_dims_divisable
=
!
(
x_int_dims
[
1
]
%
16
);
const
bool
is_x_format_correct
=
x
->
format
()
==
memory
::
format
::
nChw16c
;
const
bool
is_y_format_correct
=
y
->
format
()
==
memory
::
format
::
nc
;
if
(
is_x_format_correct
&&
is_y_format_correct
&&
are_dims_divisable
&&
is_avx512_enabled
)
{
int
pre
,
n
,
post
;
get_mid_dims
(
x_dims
,
y_dims_untrimmed
,
axis
,
&
pre
,
&
n
,
&
post
);
if
(
post
==
1
)
{
PADDLE_THROW
(
"Not implemented when post is 1"
);
}
else
{
// Just check whether it works for RE-Resnext.
PADDLE_ENFORCE_EQ
(
x_dims
.
size
(),
4
,
"X should have 4 dimensions"
);
int
n
=
x_dims
[
0
];
int
c
=
x_dims
[
1
];
int
h
=
x_dims
[
2
];
int
w
=
x_dims
[
3
];
PADDLE_ENFORCE
(
y_dims_untrimmed
[
0
]
==
n
&&
y_dims_untrimmed
[
1
]
==
c
,
"Y should be in nc format"
);
constexpr
int
simd_width
=
16
;
int
C
=
c
/
simd_width
;
const
auto
&
multiply
=
math
::
jitkernel
::
KernelPool
::
Instance
()
.
template
Get
<
math
::
jitkernel
::
EltwiseMulnChw16cNCKernel
<
T
>
>
(
n
);
#pragma omp parallel for collapse(2)
for
(
int
ni
=
0
;
ni
<
n
;
ni
++
)
{
for
(
int
ci
=
0
;
ci
<
C
;
ci
++
)
{
auto
ptr_x
=
x_data
+
ni
*
C
*
h
*
w
*
simd_width
+
ci
*
h
*
w
*
simd_width
;
auto
ptr_y
=
y_data
+
ni
*
C
*
simd_width
+
ci
*
simd_width
;
auto
ptr_z
=
z_data
+
ni
*
C
*
h
*
w
*
simd_width
+
ci
*
h
*
w
*
simd_width
;
multiply
->
Compute
(
ptr_x
,
ptr_y
,
ptr_z
,
h
,
w
);
}
}
}
z
->
set_layout
(
DataLayout
::
kMKLDNN
);
z
->
set_format
(
x
->
format
());
}
else
{
// Fallback to naive version:
const
bool
are_inputs_in_same_format
=
x
->
format
()
==
y
->
format
();
const
bool
is_x_nchw
=
x
->
format
()
==
memory
::
format
::
nchw
;
const
bool
is_x_nc
=
x
->
format
()
==
memory
::
format
::
nc
;
const
bool
is_y_nchw
=
y
->
format
()
==
memory
::
format
::
nchw
;
const
bool
is_y_nc
=
y
->
format
()
==
memory
::
format
::
nc
;
if
(
!
are_inputs_in_same_format
)
{
using
platform
::
MKLDNNDeviceContext
;
auto
&
dev_ctx
=
ctx
.
template
device_context
<
MKLDNNDeviceContext
>();
const
auto
&
mkldnn_engine
=
dev_ctx
.
GetEngine
();
if
(
!
(
is_x_nchw
||
is_x_nc
))
ReorderInput
<
T
>
((
Tensor
*
)
x
,
ctx
.
GetPlace
(),
mkldnn_engine
,
x
->
dims
().
size
()
==
4
);
if
(
!
(
is_y_nchw
||
is_y_nc
))
ReorderInput
<
T
>
((
Tensor
*
)
y
,
ctx
.
GetPlace
(),
mkldnn_engine
,
y
->
dims
().
size
()
==
4
);
}
auto
mul_func
=
[](
T
a
,
T
b
)
->
T
{
return
a
*
b
;
};
TransformFunctor
<
decltype
(
mul_func
),
T
,
paddle
::
platform
::
CPUDeviceContext
,
T
>
functor
(
x
,
y
,
z
,
ctx
.
template
device_context
<
paddle
::
platform
::
CPUDeviceContext
>(),
mul_func
);
axis
=
(
axis
==
-
1
?
x_dims
.
size
()
-
y_dims_untrimmed
.
size
()
:
axis
);
PADDLE_ENFORCE
(
axis
>=
0
&&
axis
<
x_dims
.
size
(),
"Axis should be in range [0, x_dims)"
);
auto
y_dims
=
trim_trailing_singular_dims
(
y_dims_untrimmed
);
axis
=
(
y_dims
.
size
()
==
0
)
?
x_dims
.
size
()
:
axis
;
int
pre
,
n
,
post
;
get_mid_dims
(
x_dims
,
y_dims
,
axis
,
&
pre
,
&
n
,
&
post
);
if
(
post
==
1
)
{
functor
.
RunRowWise
(
n
,
pre
);
}
else
{
functor
.
RunMidWise
(
n
,
pre
,
post
);
}
z
->
set_layout
(
DataLayout
::
kMKLDNN
);
z
->
set_format
(
x
->
format
());
}
}
};
}
// namespace operators
}
// namespace paddle
namespace
ops
=
paddle
::
operators
;
REGISTER_OP_KERNEL
(
elementwise_mul
,
MKLDNN
,
::
paddle
::
platform
::
CPUPlace
,
ops
::
ElementwiseMulMKLDNNKernel
<
float
>
)
paddle/fluid/operators/elementwise/elementwise_op.h
浏览文件 @
10fb4cee
...
@@ -97,6 +97,20 @@ class ElementwiseOpMaker : public framework::OpProtoAndCheckerMaker {
...
@@ -97,6 +97,20 @@ class ElementwiseOpMaker : public framework::OpProtoAndCheckerMaker {
.
EqualGreaterThan
(
-
1
);
.
EqualGreaterThan
(
-
1
);
AddAttr
<
bool
>
(
"use_mkldnn"
,
"(bool, default false). Used by MKLDNN."
)
AddAttr
<
bool
>
(
"use_mkldnn"
,
"(bool, default false). Used by MKLDNN."
)
.
SetDefault
(
false
);
.
SetDefault
(
false
);
AddAttr
<
std
::
string
>
(
"x_data_format"
,
"(string, default NCHW) Only used in mkldnn"
"An optional string from:
\"
NHWC
\"
,
\"
NCHW
\"
,
\"
NCHW16C
\"
,
\"
NCHW8C
\"
. "
"Defaults to
\"\"
. Specify the data format of the output data, "
"the input will be transformed automatically. "
)
.
SetDefault
(
""
);
AddAttr
<
std
::
string
>
(
"y_data_format"
,
"(string, default
\"\"
) Only used in mkldnn"
"An optional string from:
\"
NHWC
\"
,
\"
NCHW
\"
,
\"
NCHW16C
\"
,
\"
NCHW8C
\"
. "
"Defaults to
\"\"
. Specify the data format of the output data, "
"the input will be transformed automatically. "
)
.
SetDefault
(
""
);
AddComment
(
string
::
Sprintf
(
R"DOC(
AddComment
(
string
::
Sprintf
(
R"DOC(
Elementwise %s Operator
Elementwise %s Operator
...
...
paddle/fluid/operators/math/jit_code.h
浏览文件 @
10fb4cee
...
@@ -322,6 +322,42 @@ class VActJitCode : public JitCode {
...
@@ -322,6 +322,42 @@ class VActJitCode : public JitCode {
ymm_t
ymm_dst
=
ymm_t
(
1
);
ymm_t
ymm_dst
=
ymm_t
(
1
);
};
};
#ifdef PADDLE_WITH_MKLDNN
struct
EltwiseMulnChw16cNC
:
public
Xbyak
::
CodeGenerator
{
explicit
EltwiseMulnChw16cNC
(
size_t
code_size
=
256
*
1024
)
:
Xbyak
::
CodeGenerator
(
code_size
)
{
// RDI is ptr x_input
// RSI is ptr y_input
// RDX is ptr output
// RCX is height
// r8 is width
push
(
rbx
);
xor_
(
rax
,
rax
);
xor_
(
r10
,
r10
);
vmovups
(
zmm3
,
ptr
[
rsi
]);
L
(
"h_loop"
);
xor_
(
rbx
,
rbx
);
L
(
"w_loop"
);
vmovups
(
zmm2
,
ptr
[
rdi
+
rax
]);
vmulps
(
zmm1
,
zmm2
,
zmm3
);
vmovups
(
ptr
[
rdx
+
rax
],
zmm1
);
add
(
rax
,
64
);
inc
(
rbx
);
cmp
(
r8
,
rbx
);
jnz
(
"w_loop"
);
inc
(
r10
);
cmp
(
r10
,
rcx
);
jnz
(
"h_loop"
);
pop
(
rbx
);
ret
();
}
};
#endif
}
// namespace gen
}
// namespace gen
}
// namespace jitkernel
}
// namespace jitkernel
}
// namespace math
}
// namespace math
...
...
paddle/fluid/operators/math/jit_kernel.h
浏览文件 @
10fb4cee
...
@@ -95,6 +95,15 @@ class VAddBiasKernel : public Kernel {
...
@@ -95,6 +95,15 @@ class VAddBiasKernel : public Kernel {
void
(
*
Compute
)(
const
T
*
,
const
T
*
,
T
*
,
int
);
void
(
*
Compute
)(
const
T
*
,
const
T
*
,
T
*
,
int
);
};
};
#ifdef PADDLE_WITH_MKLDNN
template
<
typename
T
>
class
EltwiseMulnChw16cNCKernel
:
public
Kernel
{
public:
// nChw16c = nChw16c .* NC
void
(
*
Compute
)(
const
float
*
,
const
float
*
,
float
*
,
int
,
int
);
};
#endif
template
<
typename
T
>
template
<
typename
T
>
class
VActKernel
:
public
Kernel
{
class
VActKernel
:
public
Kernel
{
public:
public:
...
...
paddle/fluid/operators/math/jit_kernel_blas.cc
浏览文件 @
10fb4cee
...
@@ -226,6 +226,44 @@ bool VAddKernelImpl<double>::useMKL(int d) {
...
@@ -226,6 +226,44 @@ bool VAddKernelImpl<double>::useMKL(int d) {
}
}
#endif
#endif
#ifdef PADDLE_WITH_MKLDNN
/* EltwiseMul for nChw16c & NC inputs JitKernel */
template
<
typename
T
>
class
EltwiseMulnChw16cNCKernelImpl
:
public
math
::
jitkernel
::
EltwiseMulnChw16cNCKernel
<
T
>
{
public:
JITKERNEL_DECLARE_STATIC_FUNC
;
explicit
EltwiseMulnChw16cNCKernelImpl
(
int
d
)
:
EltwiseMulnChw16cNCKernel
<
T
>
()
{
using
mul_func_t
=
void
(
*
)(
const
float
*
,
const
float
*
,
float
*
,
int
,
int
);
#ifdef PADDLE_WITH_XBYAK
if
(
useJIT
(
d
))
{
// roughly estimate the size of code
size_t
sz
=
96
+
d
/
YMM_FLOAT_BLOCK
*
4
*
8
;
sz
=
sz
>
4096
?
sz
:
4096
;
jitcode_
.
reset
(
new
gen
::
EltwiseMulnChw16cNC
(
sz
));
this
->
Compute
=
(
mul_func_t
)
jitcode_
->
getCode
();
return
;
}
#endif
PADDLE_THROW
(
"This kernel shouldn't be used in Non-Xbyak, Non-MKL-DNN "
"environemnt"
);
}
#ifdef PADDLE_WITH_XBYAK
private:
std
::
unique_ptr
<
gen
::
EltwiseMulnChw16cNC
>
jitcode_
{
nullptr
};
};
template
<
>
bool
EltwiseMulnChw16cNCKernelImpl
<
float
>::
useJIT
(
int
d
)
{
return
true
;
}
#endif
#endif
/* VAddRelu JitKernel */
/* VAddRelu JitKernel */
template
<
typename
T
>
template
<
typename
T
>
class
VAddReluKernelImpl
:
public
VAddReluKernel
<
T
>
{
class
VAddReluKernelImpl
:
public
VAddReluKernel
<
T
>
{
...
@@ -394,6 +432,9 @@ REGISTER_JITKERNEL(vscal, VScalKernel);
...
@@ -394,6 +432,9 @@ REGISTER_JITKERNEL(vscal, VScalKernel);
REGISTER_JITKERNEL
(
vaddbias
,
VAddBiasKernel
);
REGISTER_JITKERNEL
(
vaddbias
,
VAddBiasKernel
);
REGISTER_JITKERNEL
(
vrelu
,
VReluKernel
);
REGISTER_JITKERNEL
(
vrelu
,
VReluKernel
);
REGISTER_JITKERNEL
(
videntity
,
VIdentityKernel
);
REGISTER_JITKERNEL
(
videntity
,
VIdentityKernel
);
#ifdef PADDLE_WITH_MKLDNN
REGISTER_JITKERNEL
(
eltwise_mul_nchw16c
,
EltwiseMulnChw16cNCKernel
);
#endif
}
// namespace jitkernel
}
// namespace jitkernel
}
// namespace math
}
// namespace math
...
...
python/paddle/fluid/tests/unittests/op_test.py
浏览文件 @
10fb4cee
...
@@ -362,7 +362,9 @@ class OpTest(unittest.TestCase):
...
@@ -362,7 +362,9 @@ class OpTest(unittest.TestCase):
else
:
else
:
return
[]
return
[]
places
=
[
fluid
.
CPUPlace
()]
places
=
[
fluid
.
CPUPlace
()]
if
core
.
is_compiled_with_cuda
()
and
core
.
op_support_gpu
(
self
.
op_type
):
cpu_only
=
self
.
_cpu_only
if
hasattr
(
self
,
'_cpu_only'
)
else
False
if
core
.
is_compiled_with_cuda
()
and
core
.
op_support_gpu
(
self
.
op_type
)
\
and
not
cpu_only
:
places
.
append
(
core
.
CUDAPlace
(
0
))
places
.
append
(
core
.
CUDAPlace
(
0
))
return
places
return
places
...
...
python/paddle/fluid/tests/unittests/test_elementwise_mul_mkldnn_op.py
0 → 100644
浏览文件 @
10fb4cee
# Copyright (c) 2018 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
from
op_test
import
OpTest
import
paddle.fluid.core
as
core
from
paddle.fluid.op
import
Operator
from
test_elementwise_mul_op
import
*
class
TestElementwiseMulMKLDNNOp_BroadcastNCHW16c
(
ElementwiseMulOp
):
def
init_input_output
(
self
):
x
=
np
.
random
.
rand
(
1
,
16
,
2
,
2
).
astype
(
self
.
dtype
)
self
.
x
=
x
.
transpose
(
0
,
2
,
3
,
1
).
reshape
(
1
,
16
,
2
,
2
)
self
.
y
=
np
.
random
.
rand
(
1
,
16
).
astype
(
self
.
dtype
)
self
.
out
=
x
*
self
.
y
.
reshape
(
1
,
16
,
1
,
1
)
self
.
out
=
self
.
out
.
transpose
(
0
,
2
,
3
,
1
).
reshape
(
1
,
16
,
2
,
2
)
def
setUp
(
self
):
super
(
TestElementwiseMulMKLDNNOp_BroadcastNCHW16c
,
self
).
setUp
()
self
.
attrs
[
"x_data_format"
]
=
"nchw16c"
self
.
attrs
[
"y_data_format"
]
=
"nc"
self
.
_cpu_only
=
True
def
init_kernel_type
(
self
):
self
.
use_mkldnn
=
True
def
init_axis
(
self
):
self
.
axis
=
0
def
test_check_grad_normal
(
self
):
pass
def
test_check_grad_ingore_x
(
self
):
pass
def
test_check_grad_ingore_y
(
self
):
pass
@
unittest
.
skip
(
"Not implemented yet."
)
# TODO(mgallus): enable when implemented.
class
TestElementwiseMulMKLDNNOp_BroadcastNCHW8c
(
ElementwiseMulOp
):
def
init_input_output
(
self
):
x
=
np
.
random
.
rand
(
1
,
8
,
2
,
2
).
astype
(
self
.
dtype
)
self
.
x
=
x
.
transpose
(
0
,
2
,
3
,
1
).
reshape
(
1
,
8
,
2
,
2
)
self
.
y
=
np
.
random
.
rand
(
1
,
8
).
astype
(
self
.
dtype
)
self
.
out
=
x
*
self
.
y
.
reshape
(
1
,
8
,
1
,
1
)
self
.
out
=
self
.
out
.
transpose
(
0
,
2
,
3
,
1
).
reshape
(
1
,
8
,
2
,
2
)
def
setUp
(
self
):
super
(
TestElementwiseMulMKLDNNOp_BroadcastNCHW8c
,
self
).
setUp
()
self
.
attrs
[
"x_data_format"
]
=
"nchw8c"
self
.
attrs
[
"y_data_format"
]
=
"nc"
self
.
_cpu_only
=
True
def
init_kernel_type
(
self
):
self
.
use_mkldnn
=
True
def
init_axis
(
self
):
self
.
axis
=
0
def
test_check_grad_normal
(
self
):
pass
def
test_check_grad_ingore_x
(
self
):
pass
def
test_check_grad_ingore_y
(
self
):
pass
class
TestElementwiseMulMKLDNNOp_FallbackNCHW
(
ElementwiseMulOp
):
def
init_input_output
(
self
):
self
.
x
=
np
.
random
.
rand
(
1
,
16
,
2
,
2
).
astype
(
self
.
dtype
)
self
.
y
=
np
.
random
.
rand
(
1
,
16
).
astype
(
self
.
dtype
)
self
.
out
=
self
.
x
*
self
.
y
.
reshape
(
1
,
16
,
1
,
1
)
def
init_kernel_type
(
self
):
self
.
use_mkldnn
=
True
def
init_axis
(
self
):
self
.
axis
=
0
def
test_check_grad_normal
(
self
):
pass
def
test_check_grad_ingore_x
(
self
):
pass
def
test_check_grad_ingore_y
(
self
):
pass
class
TestElementwiseMulMKLDNNOp_FallbackNCHW16C
(
ElementwiseMulOp
):
def
init_input_output
(
self
):
x
=
np
.
random
.
rand
(
1
,
16
,
2
,
2
).
astype
(
self
.
dtype
)
self
.
x
=
x
.
transpose
(
0
,
2
,
3
,
1
).
reshape
(
1
,
16
,
2
,
2
)
y
=
np
.
random
.
rand
(
1
,
16
,
2
,
2
).
astype
(
self
.
dtype
)
self
.
y
=
y
.
transpose
(
0
,
2
,
3
,
1
).
reshape
(
1
,
16
,
2
,
2
)
self
.
out
=
self
.
x
*
self
.
y
def
setUp
(
self
):
super
(
TestElementwiseMulMKLDNNOp_FallbackNCHW16C
,
self
).
setUp
()
self
.
attrs
[
"x_data_format"
]
=
"nchw16c"
self
.
attrs
[
"y_data_format"
]
=
"nchw16c"
self
.
_cpu_only
=
True
def
init_kernel_type
(
self
):
self
.
use_mkldnn
=
True
def
init_axis
(
self
):
self
.
axis
=
0
def
test_check_grad_normal
(
self
):
pass
def
test_check_grad_ingore_x
(
self
):
pass
def
test_check_grad_ingore_y
(
self
):
pass
class
TestElementwiseMulMKLDNNOp_FallbackNoReorders
(
ElementwiseMulOp
):
def
init_input_output
(
self
):
x
=
np
.
random
.
rand
(
1
,
16
,
2
,
2
).
astype
(
self
.
dtype
)
self
.
x
=
x
.
transpose
(
0
,
2
,
3
,
1
).
reshape
(
1
,
16
,
2
,
2
)
y
=
np
.
random
.
rand
(
1
,
16
,
2
,
2
).
astype
(
self
.
dtype
)
self
.
y
=
y
.
transpose
(
0
,
2
,
3
,
1
).
reshape
(
1
,
16
,
2
,
2
)
self
.
out
=
self
.
x
*
self
.
y
def
setUp
(
self
):
super
(
TestElementwiseMulMKLDNNOp_FallbackNoReorders
,
self
).
setUp
()
self
.
attrs
[
"x_data_format"
]
=
"nchw16c"
self
.
attrs
[
"y_data_format"
]
=
"nchw16c"
self
.
_cpu_only
=
True
def
init_kernel_type
(
self
):
self
.
use_mkldnn
=
True
def
init_axis
(
self
):
self
.
axis
=
0
def
test_check_grad_normal
(
self
):
pass
def
test_check_grad_ingore_x
(
self
):
pass
def
test_check_grad_ingore_y
(
self
):
pass
class
TestElementwiseMulMKLDNNOp_FallbackWithReorder1
(
ElementwiseMulOp
):
def
init_input_output
(
self
):
self
.
x
=
np
.
random
.
rand
(
1
,
16
,
2
,
2
).
astype
(
self
.
dtype
)
y
=
np
.
random
.
rand
(
1
,
16
,
2
,
2
).
astype
(
self
.
dtype
)
self
.
y
=
y
.
transpose
(
0
,
2
,
3
,
1
).
reshape
(
1
,
16
,
2
,
2
)
self
.
out
=
self
.
x
*
y
def
setUp
(
self
):
super
(
TestElementwiseMulMKLDNNOp_FallbackWithReorder1
,
self
).
setUp
()
self
.
attrs
[
"x_data_format"
]
=
"nchw"
self
.
attrs
[
"y_data_format"
]
=
"nchw16c"
self
.
_cpu_only
=
True
def
init_kernel_type
(
self
):
self
.
use_mkldnn
=
True
def
init_axis
(
self
):
self
.
axis
=
0
def
test_check_grad_normal
(
self
):
pass
def
test_check_grad_ingore_x
(
self
):
pass
def
test_check_grad_ingore_y
(
self
):
pass
class
TestElementwiseMulMKLDNNOp_FallbackWithReorder2
(
ElementwiseMulOp
):
def
init_input_output
(
self
):
self
.
y
=
np
.
random
.
rand
(
1
,
16
,
2
,
2
).
astype
(
self
.
dtype
)
x
=
np
.
random
.
rand
(
1
,
16
,
2
,
2
).
astype
(
self
.
dtype
)
self
.
x
=
x
.
transpose
(
0
,
2
,
3
,
1
).
reshape
(
1
,
16
,
2
,
2
)
self
.
out
=
x
*
self
.
y
def
setUp
(
self
):
super
(
TestElementwiseMulMKLDNNOp_FallbackWithReorder2
,
self
).
setUp
()
self
.
attrs
[
"x_data_format"
]
=
"nchw16c"
self
.
attrs
[
"y_data_format"
]
=
"nchw"
self
.
_cpu_only
=
True
def
init_kernel_type
(
self
):
self
.
use_mkldnn
=
True
def
init_axis
(
self
):
self
.
axis
=
0
def
test_check_grad_normal
(
self
):
pass
def
test_check_grad_ingore_x
(
self
):
pass
def
test_check_grad_ingore_y
(
self
):
pass
class
TestElementwiseMulMKLDNNOp_FallbackNoReorders2
(
ElementwiseMulOp
):
def
init_input_output
(
self
):
self
.
x
=
np
.
random
.
rand
(
1
,
16
).
astype
(
self
.
dtype
)
self
.
y
=
np
.
random
.
rand
(
1
,
16
).
astype
(
self
.
dtype
)
self
.
out
=
self
.
x
*
self
.
y
def
setUp
(
self
):
super
(
TestElementwiseMulMKLDNNOp_FallbackNoReorders2
,
self
).
setUp
()
self
.
attrs
[
"x_data_format"
]
=
"nc"
self
.
attrs
[
"y_data_format"
]
=
"nc"
self
.
_cpu_only
=
True
def
init_kernel_type
(
self
):
self
.
use_mkldnn
=
True
def
init_axis
(
self
):
self
.
axis
=
0
def
test_check_grad_normal
(
self
):
pass
def
test_check_grad_ingore_x
(
self
):
pass
def
test_check_grad_ingore_y
(
self
):
pass
if
__name__
==
'__main__'
:
unittest
.
main
()
python/paddle/fluid/tests/unittests/test_elementwise_mul_op.py
浏览文件 @
10fb4cee
...
@@ -21,13 +21,24 @@ from paddle.fluid.op import Operator
...
@@ -21,13 +21,24 @@ from paddle.fluid.op import Operator
class
ElementwiseMulOp
(
OpTest
):
class
ElementwiseMulOp
(
OpTest
):
def
init_kernel_type
(
self
):
self
.
use_mkldnn
=
False
def
setUp
(
self
):
def
setUp
(
self
):
self
.
op_type
=
"elementwise_mul"
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
=
{
self
.
inputs
=
{
'X'
:
np
.
random
.
uniform
(
0.1
,
1
,
[
13
,
17
]).
astype
(
"float64"
),
'X'
:
OpTest
.
np_dtype_to_fluid_dtype
(
self
.
x
),
'Y'
:
np
.
random
.
uniform
(
0.1
,
1
,
[
13
,
17
]).
astype
(
"float64"
)
'Y'
:
OpTest
.
np_dtype_to_fluid_dtype
(
self
.
y
)
}
}
self
.
outputs
=
{
'Out'
:
np
.
multiply
(
self
.
inputs
[
'X'
],
self
.
inputs
[
'Y'
])}
self
.
outputs
=
{
'Out'
:
self
.
out
}
self
.
attrs
=
{
'axis'
:
self
.
axis
,
'use_mkldnn'
:
self
.
use_mkldnn
}
def
test_check_output
(
self
):
def
test_check_output
(
self
):
self
.
check_output
()
self
.
check_output
()
...
@@ -41,6 +52,17 @@ class ElementwiseMulOp(OpTest):
...
@@ -41,6 +52,17 @@ class ElementwiseMulOp(OpTest):
def
test_check_grad_ingore_y
(
self
):
def
test_check_grad_ingore_y
(
self
):
self
.
check_grad
([
'X'
],
'Out'
,
no_grad_set
=
set
(
'Y'
))
self
.
check_grad
([
'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
class
TestElementwiseMulOp_scalar
(
ElementwiseMulOp
):
class
TestElementwiseMulOp_scalar
(
ElementwiseMulOp
):
def
setUp
(
self
):
def
setUp
(
self
):
...
@@ -63,17 +85,13 @@ class TestElementwiseMulOp_Vector(ElementwiseMulOp):
...
@@ -63,17 +85,13 @@ class TestElementwiseMulOp_Vector(ElementwiseMulOp):
class
TestElementwiseMulOp_broadcast_0
(
ElementwiseMulOp
):
class
TestElementwiseMulOp_broadcast_0
(
ElementwiseMulOp
):
def
setUp
(
self
):
def
init_input_output
(
self
):
self
.
op_type
=
"elementwise_mul"
self
.
x
=
np
.
random
.
rand
(
2
,
3
,
4
).
astype
(
self
.
dtype
)
self
.
inputs
=
{
self
.
y
=
np
.
random
.
rand
(
2
).
astype
(
self
.
dtype
)
'X'
:
np
.
random
.
rand
(
2
,
3
,
4
).
astype
(
np
.
float64
),
self
.
out
=
self
.
x
*
self
.
y
.
reshape
(
2
,
1
,
1
)
'Y'
:
np
.
random
.
rand
(
2
).
astype
(
np
.
float64
)
}
self
.
attrs
=
{
'axis'
:
0
}
def
init_axis
(
self
):
self
.
outputs
=
{
self
.
axis
=
0
'Out'
:
self
.
inputs
[
'X'
]
*
self
.
inputs
[
'Y'
].
reshape
(
2
,
1
,
1
)
}
class
TestElementwiseMulOp_broadcast_1
(
ElementwiseMulOp
):
class
TestElementwiseMulOp_broadcast_1
(
ElementwiseMulOp
):
...
...
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