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
25fc2a1f
P
Paddle
项目概览
PaddlePaddle
/
Paddle
大约 1 年 前同步成功
通知
2298
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
25fc2a1f
编写于
3月 19, 2021
作者:
J
Jacek Czaja
提交者:
GitHub
3月 19, 2021
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
[oneDNN] Added Elementwise Mul grad fp32/bf16 (#31647)
上级
878e117b
变更
7
隐藏空白更改
内联
并排
Showing
7 changed file
with
206 addition
and
15 deletion
+206
-15
paddle/fluid/operators/elementwise/elementwise_op.h
paddle/fluid/operators/elementwise/elementwise_op.h
+2
-3
paddle/fluid/operators/elementwise/mkldnn/elementwise_add_mkldnn_op.cc
...operators/elementwise/mkldnn/elementwise_add_mkldnn_op.cc
+11
-0
paddle/fluid/operators/elementwise/mkldnn/elementwise_mkldnn_op.h
...luid/operators/elementwise/mkldnn/elementwise_mkldnn_op.h
+0
-1
paddle/fluid/operators/elementwise/mkldnn/elementwise_mul_mkldnn_op.cc
...operators/elementwise/mkldnn/elementwise_mul_mkldnn_op.cc
+116
-0
paddle/fluid/platform/mkldnn_reuse.h
paddle/fluid/platform/mkldnn_reuse.h
+9
-1
python/paddle/fluid/tests/unittests/mkldnn/test_elementwise_mul_bf16_mkldnn_op.py
...s/unittests/mkldnn/test_elementwise_mul_bf16_mkldnn_op.py
+59
-7
python/paddle/fluid/tests/unittests/mkldnn/test_elementwise_mul_mkldnn_op.py
.../tests/unittests/mkldnn/test_elementwise_mul_mkldnn_op.py
+9
-3
未找到文件。
paddle/fluid/operators/elementwise/elementwise_op.h
浏览文件 @
25fc2a1f
...
@@ -276,7 +276,7 @@ class ElementwiseOpGrad : public framework::OperatorWithKernel {
...
@@ -276,7 +276,7 @@ class ElementwiseOpGrad : public framework::OperatorWithKernel {
#ifdef PADDLE_WITH_MKLDNN
#ifdef PADDLE_WITH_MKLDNN
// If broadcasting is needed, use native implementation
// If broadcasting is needed, use native implementation
auto
CanMKLDNNElementwise
Add
GradBeUsed
=
[
&
]()
{
auto
CanMKLDNNElementwiseGradBeUsed
=
[
&
]()
{
auto
dx_dims
=
ctx
.
Input
<
Tensor
>
(
"X"
)
->
dims
();
auto
dx_dims
=
ctx
.
Input
<
Tensor
>
(
"X"
)
->
dims
();
auto
dy_dims
=
ctx
.
Input
<
Tensor
>
(
"Y"
)
->
dims
();
auto
dy_dims
=
ctx
.
Input
<
Tensor
>
(
"Y"
)
->
dims
();
// No broadcast or broadcasting of data on inner dims is supported
// No broadcast or broadcasting of data on inner dims is supported
...
@@ -284,8 +284,7 @@ class ElementwiseOpGrad : public framework::OperatorWithKernel {
...
@@ -284,8 +284,7 @@ class ElementwiseOpGrad : public framework::OperatorWithKernel {
};
};
if
(
this
->
CanMKLDNNBeUsed
(
ctx
,
input_data_type
)
&&
if
(
this
->
CanMKLDNNBeUsed
(
ctx
,
input_data_type
)
&&
(
ctx
.
Type
()
!=
"elementwise_add_grad"
||
CanMKLDNNElementwiseGradBeUsed
())
{
CanMKLDNNElementwiseAddGradBeUsed
()))
{
return
framework
::
OpKernelType
(
input_data_type
,
ctx
.
GetPlace
(),
return
framework
::
OpKernelType
(
input_data_type
,
ctx
.
GetPlace
(),
framework
::
DataLayout
::
kMKLDNN
,
framework
::
DataLayout
::
kMKLDNN
,
framework
::
LibraryType
::
kMKLDNN
);
framework
::
LibraryType
::
kMKLDNN
);
...
...
paddle/fluid/operators/elementwise/mkldnn/elementwise_add_mkldnn_op.cc
浏览文件 @
25fc2a1f
...
@@ -61,6 +61,9 @@ class EltwiseAddMKLDNNGradKernel : public ElemwiseGradKernel<T> {
...
@@ -61,6 +61,9 @@ class EltwiseAddMKLDNNGradKernel : public ElemwiseGradKernel<T> {
platform
::
EventRole
::
kUniqueOp
);
platform
::
EventRole
::
kUniqueOp
);
reorder_p
->
execute
(
astream
,
*
reorder_src_memory_p
,
*
reorder_dst_memory_p
);
reorder_p
->
execute
(
astream
,
*
reorder_src_memory_p
,
*
reorder_dst_memory_p
);
astream
.
wait
();
astream
.
wait
();
dx
->
set_layout
(
DataLayout
::
kMKLDNN
);
dx
->
set_format
(
platform
::
GetMKLDNNFormat
(
*
reorder_dst_memory_p
));
}
}
if
(
dy
)
{
if
(
dy
)
{
...
@@ -75,6 +78,9 @@ class EltwiseAddMKLDNNGradKernel : public ElemwiseGradKernel<T> {
...
@@ -75,6 +78,9 @@ class EltwiseAddMKLDNNGradKernel : public ElemwiseGradKernel<T> {
reorder_p
->
execute
(
astream
,
*
reorder_src_memory_p
,
reorder_p
->
execute
(
astream
,
*
reorder_src_memory_p
,
*
reorder_dst_memory_p
);
*
reorder_dst_memory_p
);
astream
.
wait
();
astream
.
wait
();
dy
->
set_layout
(
DataLayout
::
kMKLDNN
);
dy
->
set_format
(
platform
::
GetMKLDNNFormat
(
*
reorder_dst_memory_p
));
}
else
{
}
else
{
// Broadcasting
// Broadcasting
platform
::
ReductionMKLDNNHandler
<
T
>
handler_sum
(
platform
::
ReductionMKLDNNHandler
<
T
>
handler_sum
(
...
@@ -86,6 +92,11 @@ class EltwiseAddMKLDNNGradKernel : public ElemwiseGradKernel<T> {
...
@@ -86,6 +92,11 @@ class EltwiseAddMKLDNNGradKernel : public ElemwiseGradKernel<T> {
reduction_p
->
execute
(
astream
,
{{
DNNL_ARG_SRC
,
*
reorder_src_memory_p
},
reduction_p
->
execute
(
astream
,
{{
DNNL_ARG_SRC
,
*
reorder_src_memory_p
},
{
DNNL_ARG_DST
,
*
dy_memory_p
}});
{
DNNL_ARG_DST
,
*
dy_memory_p
}});
astream
.
wait
();
astream
.
wait
();
dy
->
set_layout
(
DataLayout
::
kMKLDNN
);
dy
->
set_format
(
platform
::
GetMKLDNNFormat
(
dy_memory_p
->
get_desc
().
reshape
(
paddle
::
framework
::
vectorize
<
int64_t
>
(
dy
->
dims
()))));
}
}
}
}
}
}
...
...
paddle/fluid/operators/elementwise/mkldnn/elementwise_mkldnn_op.h
浏览文件 @
25fc2a1f
...
@@ -15,7 +15,6 @@
...
@@ -15,7 +15,6 @@
#pragma once
#pragma once
#include <string>
#include <string>
#include <unordered_map>
#include <unordered_map>
#include "paddle/fluid/memory/memcpy.h"
#include "paddle/fluid/operators/elementwise/elementwise_add_op.h"
#include "paddle/fluid/operators/elementwise/elementwise_add_op.h"
#include "paddle/fluid/operators/elementwise/elementwise_op_function.h"
#include "paddle/fluid/operators/elementwise/elementwise_op_function.h"
...
...
paddle/fluid/operators/elementwise/mkldnn/elementwise_mul_mkldnn_op.cc
浏览文件 @
25fc2a1f
...
@@ -14,6 +14,118 @@ limitations under the License. */
...
@@ -14,6 +14,118 @@ limitations under the License. */
#include "paddle/fluid/operators/elementwise/mkldnn/elementwise_mkldnn_op.h"
#include "paddle/fluid/operators/elementwise/mkldnn/elementwise_mkldnn_op.h"
namespace
paddle
{
namespace
framework
{
class
ExecutionContext
;
}
// namespace framework
namespace
platform
{
class
CPUDeviceContext
;
struct
CPUPlace
;
}
// namespace platform
}
// namespace paddle
namespace
paddle
{
namespace
operators
{
template
<
typename
T
>
class
EltwiseMulMKLDNNGradKernel
:
public
ElemwiseGradKernel
<
T
>
{
public:
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
ElemwiseGradKernel
<
T
>::
Compute
(
ctx
);
auto
&
dev_ctx
=
ctx
.
template
device_context
<
paddle
::
platform
::
MKLDNNDeviceContext
>();
const
auto
&
mkldnn_engine
=
dev_ctx
.
GetEngine
();
auto
*
x
=
ctx
.
Input
<
framework
::
Tensor
>
(
"X"
);
auto
*
y
=
ctx
.
Input
<
framework
::
Tensor
>
(
"Y"
);
auto
*
dout
=
ctx
.
Input
<
framework
::
Tensor
>
(
framework
::
GradVarName
(
"Out"
));
auto
*
dx
=
ctx
.
Output
<
framework
::
Tensor
>
(
framework
::
GradVarName
(
"X"
));
auto
*
dy
=
ctx
.
Output
<
framework
::
Tensor
>
(
framework
::
GradVarName
(
"Y"
));
int
axis
=
ctx
.
Attr
<
int
>
(
"axis"
);
auto
&
astream
=
platform
::
MKLDNNDeviceContext
::
tls
().
get_stream
();
if
(
dx
)
{
// dx = dout*y
platform
::
BinaryMKLDNNHandler
<
T
>
handler
(
dnnl
::
algorithm
::
binary_mul
,
axis
,
dev_ctx
,
mkldnn_engine
,
ctx
.
GetPlace
(),
dout
,
y
,
dx
,
1.0
f
,
1.0
f
,
1.0
f
,
ctx
.
InputName
(
framework
::
GradVarName
(
"Out"
)));
const
auto
src_dout_memory
=
handler
.
AcquireSrcMemory
(
dout
);
const
auto
src_y_memory
=
handler
.
AcquireSecondSrcMemory
(
y
);
const
auto
dst_dx_memory
=
handler
.
AcquireDstMemory
(
dx
);
const
auto
binary_prim
=
handler
.
AcquireForwardPrimitive
();
const
std
::
unordered_map
<
int
,
dnnl
::
memory
>
args
=
{
{
DNNL_ARG_SRC_0
,
*
src_dout_memory
},
{
DNNL_ARG_SRC_1
,
*
src_y_memory
},
{
DNNL_ARG_DST
,
*
dst_dx_memory
}};
binary_prim
->
execute
(
astream
,
args
);
astream
.
wait
();
dx
->
set_layout
(
framework
::
DataLayout
::
kMKLDNN
);
dx
->
set_format
(
platform
::
GetMKLDNNFormat
(
*
dst_dx_memory
));
}
if
(
dy
)
{
// dy = dout*x
// Handler is having nullptr passed instead of output tensor as
// we want Dst buffer to be allocated by oneDNN not to use Tensor
platform
::
BinaryMKLDNNHandler
<
T
>
handler
(
dnnl
::
algorithm
::
binary_mul
,
axis
,
dev_ctx
,
mkldnn_engine
,
ctx
.
GetPlace
(),
dout
,
x
,
nullptr
,
1.0
f
,
1.0
f
,
1.0
f
,
ctx
.
InputName
(
framework
::
GradVarName
(
"Out"
)));
const
auto
src_dout_memory
=
handler
.
AcquireSrcMemory
(
dout
);
const
auto
src_x_memory
=
handler
.
AcquireSecondSrcMemory
(
x
);
// If broadcasting is in use then let's write to temporary
// buffer allocated by oneDNN
const
auto
dst_dy_memory
=
(
dout
->
dims
()
==
dy
->
dims
())
?
handler
.
AcquireDstMemory
(
dy
)
:
handler
.
AcquireDstMemory
();
const
auto
binary_prim
=
handler
.
AcquireForwardPrimitive
();
const
std
::
unordered_map
<
int
,
dnnl
::
memory
>
args
=
{
{
DNNL_ARG_SRC_0
,
*
src_dout_memory
},
{
DNNL_ARG_SRC_1
,
*
src_x_memory
},
{
DNNL_ARG_DST
,
*
dst_dy_memory
}};
binary_prim
->
execute
(
astream
,
args
);
astream
.
wait
();
dy
->
set_layout
(
framework
::
DataLayout
::
kMKLDNN
);
// Reduction is needed for broadcasting scenario
if
(
dout
->
dims
()
!=
dy
->
dims
())
{
platform
::
ReductionMKLDNNHandler
<
T
>
handler_sum
(
dnnl
::
algorithm
::
reduction_sum
,
0.0
f
,
0.0
f
,
dev_ctx
,
mkldnn_engine
,
ctx
.
GetPlace
(),
dout
,
dy
,
ctx
.
InputName
(
framework
::
GradVarName
(
"Out"
)));
auto
dy_memory_p
=
handler_sum
.
AcquireDstMemory
(
dy
);
auto
reduction_p
=
handler_sum
.
AcquireForwardPrimitive
();
// As source we use mem object with results from binary operation
reduction_p
->
execute
(
astream
,
{{
DNNL_ARG_SRC
,
*
dst_dy_memory
},
{
DNNL_ARG_DST
,
*
dy_memory_p
}});
astream
.
wait
();
dy
->
set_format
(
platform
::
GetMKLDNNFormat
(
dy_memory_p
->
get_desc
().
reshape
(
paddle
::
framework
::
vectorize
<
int64_t
>
(
dy
->
dims
()))));
}
else
{
dy
->
set_format
(
platform
::
GetMKLDNNFormat
(
*
dst_dy_memory
));
}
}
}
};
}
// namespace operators
}
// namespace paddle
namespace
ops
=
paddle
::
operators
;
namespace
ops
=
paddle
::
operators
;
REGISTER_OP_KERNEL
(
REGISTER_OP_KERNEL
(
...
@@ -23,3 +135,7 @@ REGISTER_OP_KERNEL(
...
@@ -23,3 +135,7 @@ REGISTER_OP_KERNEL(
dnnl
::
algorithm
::
binary_mul
>
,
dnnl
::
algorithm
::
binary_mul
>
,
ops
::
EltwiseMKLDNNKernel
<
int8_t
,
dnnl
::
algorithm
::
binary_mul
>
,
ops
::
EltwiseMKLDNNKernel
<
int8_t
,
dnnl
::
algorithm
::
binary_mul
>
,
ops
::
EltwiseMKLDNNKernel
<
uint8_t
,
dnnl
::
algorithm
::
binary_mul
>
)
ops
::
EltwiseMKLDNNKernel
<
uint8_t
,
dnnl
::
algorithm
::
binary_mul
>
)
REGISTER_OP_KERNEL
(
elementwise_mul_grad
,
MKLDNN
,
::
paddle
::
platform
::
CPUPlace
,
ops
::
EltwiseMulMKLDNNGradKernel
<
paddle
::
platform
::
bfloat16
>
,
ops
::
EltwiseMulMKLDNNGradKernel
<
float
>
)
paddle/fluid/platform/mkldnn_reuse.h
浏览文件 @
25fc2a1f
...
@@ -87,6 +87,11 @@ class MKLDNNHandlerT {
...
@@ -87,6 +87,11 @@ class MKLDNNHandlerT {
"@dst_mem_p"
);
"@dst_mem_p"
);
}
}
template
<
typename
T_out
=
T
>
std
::
shared_ptr
<
mkldnn
::
memory
>
AcquireDstMemory
(
void
)
{
return
this
->
AcquireMemoryFromPrimitive
(
fwd_pd_
->
dst_desc
(),
"@dstt_mem_p"
);
}
template
<
typename
T_out
=
T
>
template
<
typename
T_out
=
T
>
std
::
shared_ptr
<
mkldnn
::
memory
>
AcquireDstMemory
(
std
::
shared_ptr
<
mkldnn
::
memory
>
AcquireDstMemory
(
const
framework
::
Tensor
*
output
)
{
const
framework
::
Tensor
*
output
)
{
...
@@ -561,7 +566,10 @@ class BinaryMKLDNNHandler : public platform::MKLDNNHandlerT<T, dnnl::binary> {
...
@@ -561,7 +566,10 @@ class BinaryMKLDNNHandler : public platform::MKLDNNHandlerT<T, dnnl::binary> {
const
auto
src_x_tz
=
framework
::
vectorize
(
x
->
dims
());
const
auto
src_x_tz
=
framework
::
vectorize
(
x
->
dims
());
const
auto
src_y_tz
=
framework
::
vectorize
(
y
->
dims
());
const
auto
src_y_tz
=
framework
::
vectorize
(
y
->
dims
());
const
auto
dst_tz
=
framework
::
vectorize
(
z
->
dims
());
// if output tensor(z) is nullptr then we are computing into oneDNN
// managed buffer
const
auto
dst_tz
=
(
z
==
nullptr
)
?
src_x_tz
:
framework
::
vectorize
(
z
->
dims
());
const
auto
src0_md
=
dnnl
::
memory
::
desc
(
const
auto
src0_md
=
dnnl
::
memory
::
desc
(
src_x_tz
,
platform
::
MKLDNNGetDataType
<
T
>
(),
x
->
format
());
src_x_tz
,
platform
::
MKLDNNGetDataType
<
T
>
(),
x
->
format
());
...
...
python/paddle/fluid/tests/unittests/mkldnn/test_elementwise_mul_bf16_mkldnn_op.py
浏览文件 @
25fc2a1f
...
@@ -30,10 +30,9 @@ class TestElementwiseMulBf16MklDNNOp(OpTest):
...
@@ -30,10 +30,9 @@ class TestElementwiseMulBf16MklDNNOp(OpTest):
self
.
axis
=
-
1
self
.
axis
=
-
1
self
.
generate_data
()
self
.
generate_data
()
self
.
inputs
=
{
self
.
x_bf16
=
convert_float_to_uint16
(
self
.
x
)
'X'
:
convert_float_to_uint16
(
self
.
x
),
self
.
y_bf16
=
convert_float_to_uint16
(
self
.
y
)
'Y'
:
convert_float_to_uint16
(
self
.
y
)
self
.
inputs
=
{
'X'
:
self
.
x_bf16
,
'Y'
:
self
.
y_bf16
}
}
self
.
attrs
=
{
'axis'
:
self
.
axis
,
'use_mkldnn'
:
self
.
use_mkldnn
}
self
.
attrs
=
{
'axis'
:
self
.
axis
,
'use_mkldnn'
:
self
.
use_mkldnn
}
self
.
outputs
=
{
'Out'
:
convert_float_to_uint16
(
self
.
out
)}
self
.
outputs
=
{
'Out'
:
convert_float_to_uint16
(
self
.
out
)}
...
@@ -46,13 +45,66 @@ class TestElementwiseMulBf16MklDNNOp(OpTest):
...
@@ -46,13 +45,66 @@ class TestElementwiseMulBf16MklDNNOp(OpTest):
self
.
check_output_with_place
(
core
.
CPUPlace
())
self
.
check_output_with_place
(
core
.
CPUPlace
())
def
test_check_grad_normal
(
self
):
def
test_check_grad_normal
(
self
):
pass
self
.
check_grad_with_place
(
core
.
CPUPlace
(),
[
"X"
,
"Y"
],
"Out"
,
check_dygraph
=
False
,
user_defined_grads
=
[
np
.
multiply
(
self
.
x
,
self
.
y
),
np
.
multiply
(
self
.
x
,
self
.
x
)
],
user_defined_grad_outputs
=
[
self
.
x_bf16
])
def
test_check_grad_ingore_x
(
self
):
def
test_check_grad_ingore_x
(
self
):
pass
self
.
check_grad_with_place
(
core
.
CPUPlace
(),
[
"Y"
],
"Out"
,
check_dygraph
=
False
,
user_defined_grads
=
[
np
.
multiply
(
self
.
y
,
self
.
x
)],
user_defined_grad_outputs
=
[
self
.
y_bf16
])
def
test_check_grad_ingore_y
(
self
):
def
test_check_grad_ingore_y
(
self
):
pass
self
.
check_grad_with_place
(
core
.
CPUPlace
(),
[
"X"
],
"Out"
,
check_dygraph
=
False
,
user_defined_grads
=
[
np
.
multiply
(
self
.
x
,
self
.
y
)],
user_defined_grad_outputs
=
[
self
.
x_bf16
])
class
TestElementwiseMulBroadcastingBf16MklDNNOp
(
TestElementwiseMulBf16MklDNNOp
):
def
generate_data
(
self
):
self
.
x
=
np
.
random
.
uniform
(
1
,
2
,
[
1
,
2
,
3
,
100
]).
astype
(
np
.
float32
)
self
.
y
=
np
.
random
.
uniform
(
1
,
2
,
[
100
]).
astype
(
np
.
float32
)
self
.
out
=
np
.
multiply
(
self
.
x
,
self
.
y
)
# Compute partial sums along all axes but last one
def
compute_reduced_gradients
(
self
,
out_grads
):
part_sum
=
np
.
add
.
reduceat
(
out_grads
,
[
0
],
axis
=
0
)
part_sum
=
np
.
add
.
reduceat
(
part_sum
,
[
0
],
axis
=
1
)
part_sum
=
np
.
add
.
reduceat
(
part_sum
,
[
0
],
axis
=
2
)
return
part_sum
.
flatten
()
def
test_check_grad_normal
(
self
):
self
.
check_grad_with_place
(
core
.
CPUPlace
(),
[
"X"
,
"Y"
],
"Out"
,
check_dygraph
=
False
,
user_defined_grads
=
[
np
.
multiply
(
self
.
x
,
self
.
y
),
self
.
compute_reduced_gradients
(
np
.
multiply
(
self
.
x
,
self
.
x
))
],
user_defined_grad_outputs
=
[
self
.
x_bf16
])
def
test_check_grad_ingore_x
(
self
):
self
.
check_grad_with_place
(
core
.
CPUPlace
(),
[
"Y"
],
"Out"
,
check_dygraph
=
False
,
user_defined_grads
=
[
self
.
compute_reduced_gradients
(
np
.
multiply
(
self
.
x
,
self
.
x
))
],
user_defined_grad_outputs
=
[
self
.
x_bf16
])
if
__name__
==
'__main__'
:
if
__name__
==
'__main__'
:
...
...
python/paddle/fluid/tests/unittests/mkldnn/test_elementwise_mul_mkldnn_op.py
浏览文件 @
25fc2a1f
...
@@ -17,6 +17,7 @@ import unittest
...
@@ -17,6 +17,7 @@ import unittest
import
numpy
as
np
import
numpy
as
np
from
paddle.fluid.tests.unittests.op_test
import
skip_check_grad_ci
from
paddle.fluid.tests.unittests.op_test
import
skip_check_grad_ci
from
paddle.fluid.tests.unittests.test_elementwise_mul_op
import
ElementwiseMulOp
from
paddle.fluid.tests.unittests.test_elementwise_mul_op
import
ElementwiseMulOp
from
paddle
import
enable_static
class
TestMKLDNNElementwiseMulOp
(
ElementwiseMulOp
):
class
TestMKLDNNElementwiseMulOp
(
ElementwiseMulOp
):
...
@@ -51,13 +52,17 @@ class TestMKLDNNElementwiseMulOp4(TestMKLDNNElementwiseMulOp):
...
@@ -51,13 +52,17 @@ class TestMKLDNNElementwiseMulOp4(TestMKLDNNElementwiseMulOp):
def
test_check_grad_normal
(
self
):
def
test_check_grad_normal
(
self
):
pass
pass
def
test_check_grad_ingore_x
(
self
):
pass
def
test_check_grad_ingore_y
(
self
):
def
test_check_grad_ingore_y
(
self
):
pass
pass
class
TestMKLDNNElementwiseMulOp5
(
TestMKLDNNElementwiseMulOp
):
def
init_input_output
(
self
):
self
.
x
=
np
.
random
.
uniform
(
1
,
2
,
[
2
,
3
,
4
,
100
]).
astype
(
self
.
dtype
)
self
.
y
=
np
.
random
.
uniform
(
1
,
2
,
[
100
]).
astype
(
self
.
dtype
)
self
.
out
=
np
.
multiply
(
self
.
x
,
self
.
y
)
''' INT8 Tests '''
''' INT8 Tests '''
...
@@ -140,4 +145,5 @@ class TestUint8Scales(TestInt8Scales):
...
@@ -140,4 +145,5 @@ class TestUint8Scales(TestInt8Scales):
if
__name__
==
'__main__'
:
if
__name__
==
'__main__'
:
enable_static
()
unittest
.
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