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
Crayon鑫
Paddle
提交
96b4904d
P
Paddle
项目概览
Crayon鑫
/
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
提交
96b4904d
编写于
6月 12, 2018
作者:
M
mozga-intel
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
MKLDNN layout: Support for sum operator
上级
a29cb4be
变更
9
显示空白变更内容
内联
并排
Showing
9 changed file
with
373 addition
and
100 deletion
+373
-100
paddle/fluid/operators/parallel_do_op.cc
paddle/fluid/operators/parallel_do_op.cc
+1
-1
paddle/fluid/operators/recurrent_op.cc
paddle/fluid/operators/recurrent_op.cc
+2
-1
paddle/fluid/operators/sum_mkldnn_op.cc
paddle/fluid/operators/sum_mkldnn_op.cc
+242
-0
paddle/fluid/operators/sum_op.cc
paddle/fluid/operators/sum_op.cc
+26
-6
paddle/fluid/operators/while_op.cc
paddle/fluid/operators/while_op.cc
+2
-2
python/paddle/fluid/backward.py
python/paddle/fluid/backward.py
+6
-5
python/paddle/fluid/layers/nn.py
python/paddle/fluid/layers/nn.py
+73
-70
python/paddle/fluid/layers/tensor.py
python/paddle/fluid/layers/tensor.py
+17
-13
python/paddle/fluid/transpiler/distribute_transpiler.py
python/paddle/fluid/transpiler/distribute_transpiler.py
+4
-2
未找到文件。
paddle/fluid/operators/parallel_do_op.cc
浏览文件 @
96b4904d
...
@@ -295,7 +295,7 @@ class ParallelDoGradOp : public framework::OperatorBase {
...
@@ -295,7 +295,7 @@ class ParallelDoGradOp : public framework::OperatorBase {
auto
sum_op
=
framework
::
OpRegistry
::
CreateOp
(
auto
sum_op
=
framework
::
OpRegistry
::
CreateOp
(
"sum"
,
{{
"X"
,
{
s
,
tmp_name
}}},
{{
"Out"
,
{
s
}}},
"sum"
,
{{
"X"
,
{
s
,
tmp_name
}}},
{{
"Out"
,
{
s
}}},
framework
::
AttributeMap
{});
framework
::
AttributeMap
{
{
"use_mkldnn"
,
{
false
}}
});
VLOG
(
10
)
<<
sum_op
->
DebugStringEx
(
sub_scopes
[
0
]);
VLOG
(
10
)
<<
sum_op
->
DebugStringEx
(
sub_scopes
[
0
]);
sum_op
->
Run
(
*
sub_scopes
[
0
],
places
[
0
]);
sum_op
->
Run
(
*
sub_scopes
[
0
],
places
[
0
]);
WaitOnPlace
(
places
[
0
]);
WaitOnPlace
(
places
[
0
]);
...
...
paddle/fluid/operators/recurrent_op.cc
浏览文件 @
96b4904d
...
@@ -429,7 +429,8 @@ class RecurrentGradOp : public RecurrentBase {
...
@@ -429,7 +429,8 @@ class RecurrentGradOp : public RecurrentBase {
auto
sum_op
=
framework
::
OpRegistry
::
CreateOp
(
auto
sum_op
=
framework
::
OpRegistry
::
CreateOp
(
"sum"
,
{{
"X"
,
{
pg_names
[
param_id
],
new_inside_name
}}},
"sum"
,
{{
"X"
,
{
pg_names
[
param_id
],
new_inside_name
}}},
{{
"Out"
,
{
pg_names
[
param_id
]}}},
framework
::
AttributeMap
{});
{{
"Out"
,
{
pg_names
[
param_id
]}}},
framework
::
AttributeMap
{{
"use_mkldnn"
,
{
false
}}});
sum_op
->
Run
(
cur_scope
,
place
);
sum_op
->
Run
(
cur_scope
,
place
);
cur_scope
.
Rename
(
new_inside_name
,
inside_grad_name
);
cur_scope
.
Rename
(
new_inside_name
,
inside_grad_name
);
...
...
paddle/fluid/operators/sum_mkldnn_op.cc
0 → 100644
浏览文件 @
96b4904d
// 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.
/*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.hpp"
#include "paddle/fluid/framework/tensor.h"
#include "paddle/fluid/operators/math/selected_rows_functor.h"
#include "paddle/fluid/operators/sum_op.h"
#include "paddle/fluid/platform/device_context.h"
#include "paddle/fluid/platform/mkldnn_helper.h"
namespace
paddle
{
namespace
operators
{
using
paddle
::
framework
::
Tensor
;
using
paddle
::
platform
::
MKLDNNDeviceContext
;
using
paddle
::
platform
::
CPUDeviceContext
;
using
framework
::
DataLayout
;
using
mkldnn
::
memory
;
using
mkldnn
::
primitive
;
using
mkldnn
::
stream
;
using
mkldnn
::
sum
;
using
mkldnn
::
reorder
;
using
platform
::
to_void_cast
;
template
<
typename
T
>
class
SumMKLDNNOpKernel
:
public
paddle
::
framework
::
OpKernel
<
T
>
{
public:
void
Compute
(
const
paddle
::
framework
::
ExecutionContext
&
ctx
)
const
override
{
PADDLE_ENFORCE
(
paddle
::
platform
::
is_cpu_place
(
ctx
.
GetPlace
()),
"It must use CPUPlace."
);
auto
&
dev_ctx
=
ctx
.
template
device_context
<
MKLDNNDeviceContext
>();
const
auto
&
mkldnn_engine
=
dev_ctx
.
GetEngine
();
auto
in_vars
=
ctx
.
MultiInputVar
(
"X"
);
const
int
N
=
in_vars
.
size
();
auto
out_var
=
ctx
.
OutputVar
(
"Out"
);
bool
in_place
=
out_var
==
in_vars
[
0
];
if
(
out_var
->
IsType
<
framework
::
LoDTensor
>
())
{
LoDTensor
*
output
=
ctx
.
Output
<
LoDTensor
>
(
"Out"
);
T
*
output_data
=
output
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
std
::
vector
<
int
>
dst_tz
=
framework
::
vectorize2int
(
output
->
dims
());
auto
src_tz
=
dst_tz
;
memory
::
format
output_format
{
memory
::
format
::
format_undef
};
std
::
vector
<
float
>
scales
;
std
::
vector
<
memory
::
primitive_desc
>
srcs_mpd
;
std
::
vector
<
mkldnn
::
memory
>
srcs_mem
;
PADDLE_ENFORCE
(
in_vars
[
0
]
->
IsType
<
LoDTensor
>
(),
"Input[0] must be LoDTensors"
);
auto
&
input0
=
in_vars
[
0
]
->
Get
<
LoDTensor
>
();
PADDLE_ENFORCE
(
input0
.
layout
()
==
DataLayout
::
kMKLDNN
&&
input0
.
format
()
!=
memory
::
format
::
format_undef
,
"Wrong layout/format for inputs[0]"
);
memory
::
format
input_format
=
input0
.
format
();
if
(
src_tz
.
size
()
==
1
&&
(
input_format
==
memory
::
format
::
nchw
||
input_format
==
memory
::
format
::
nhwc
))
{
input_format
=
memory
::
format
::
x
;
}
if
(
src_tz
.
size
()
==
2
&&
(
input_format
==
memory
::
format
::
nchw
||
input_format
==
memory
::
format
::
nhwc
))
{
input_format
=
memory
::
format
::
nc
;
}
for
(
int
i
=
in_place
?
1
:
0
;
i
<
N
;
i
++
)
{
PADDLE_ENFORCE
(
in_vars
[
i
]
->
IsType
<
LoDTensor
>
(),
"all inputs must be all LoDTensors"
);
auto
&
input
=
in_vars
[
i
]
->
Get
<
LoDTensor
>
();
PADDLE_ENFORCE
(
input
.
layout
()
==
DataLayout
::
kMKLDNN
&&
input
.
format
()
!=
memory
::
format
::
format_undef
,
"Wrong layout/format for inputs"
);
if
(
input
.
numel
()
==
0
)
{
continue
;
}
const
T
*
input_data
=
input
.
data
<
T
>
();
auto
src_md
=
memory
::
desc
(
src_tz
,
memory
::
data_type
::
f32
,
input_format
);
auto
src_mpd
=
memory
::
primitive_desc
(
src_md
,
mkldnn_engine
);
auto
src_mem
=
memory
(
src_mpd
,
to_void_cast
(
input_data
));
srcs_mpd
.
push_back
(
src_mpd
);
srcs_mem
.
push_back
(
src_mem
);
scales
.
push_back
(
1.0
);
}
auto
dst_md
=
memory
::
desc
(
dst_tz
,
memory
::
data_type
::
f32
,
memory
::
format
::
any
);
auto
sum_pd
=
sum
::
primitive_desc
(
dst_md
,
scales
,
srcs_mpd
);
std
::
shared_ptr
<
memory
>
dst_mem
;
if
(
in_place
)
dst_mem
.
reset
(
new
memory
(
sum_pd
.
dst_primitive_desc
()));
else
dst_mem
.
reset
(
new
memory
(
sum_pd
.
dst_primitive_desc
(),
output_data
));
std
::
vector
<
mkldnn
::
primitive
::
at
>
inputs
;
for
(
size_t
i
=
0
;
i
<
srcs_mem
.
size
();
++
i
)
{
inputs
.
push_back
(
srcs_mem
[
i
]);
}
auto
sum_prim
=
mkldnn
::
sum
(
sum_pd
,
inputs
,
*
dst_mem
);
output_format
=
(
memory
::
format
)
sum_pd
.
dst_primitive_desc
().
desc
().
data
.
format
;
primitive
reorder_prim
;
std
::
shared_ptr
<
memory
>
target_mem
;
if
(
in_place
)
{
output_format
=
input_format
;
target_mem
.
reset
(
new
memory
(
{{{
src_tz
},
memory
::
data_type
::
f32
,
output_format
},
mkldnn_engine
},
output_data
));
reorder_prim
=
reorder
(
*
dst_mem
,
*
target_mem
);
}
std
::
vector
<
primitive
>
pipeline
;
pipeline
.
push_back
(
sum_prim
);
if
(
in_place
)
pipeline
.
push_back
(
reorder_prim
);
stream
(
stream
::
kind
::
eager
).
submit
(
pipeline
).
wait
();
output
->
set_layout
(
DataLayout
::
kMKLDNN
);
output
->
set_format
(
output_format
);
}
else
if
(
out_var
->
IsType
<
framework
::
SelectedRows
>
())
{
// TODO(@mozga-intel) Add MKLDNN SelectedRows support
std
::
unique_ptr
<
framework
::
SelectedRows
>
in0
;
if
(
in_place
)
{
// If is in_place, we store the input[0] to in0
auto
&
in_sel0
=
in_vars
[
0
]
->
Get
<
SelectedRows
>
();
auto
&
rows
=
in_sel0
.
rows
();
in0
.
reset
(
new
framework
::
SelectedRows
(
rows
,
in_sel0
.
height
()));
in0
->
mutable_value
()
->
ShareDataWith
(
in_sel0
.
value
());
}
auto
get_selected_row
=
[
&
](
size_t
i
)
->
const
SelectedRows
&
{
if
(
i
==
0
&&
in0
)
{
return
*
in0
.
get
();
}
else
{
return
in_vars
[
i
]
->
Get
<
SelectedRows
>
();
}
};
auto
*
out
=
ctx
.
Output
<
SelectedRows
>
(
"Out"
);
out
->
mutable_rows
()
->
clear
();
auto
*
out_value
=
out
->
mutable_value
();
// Runtime InferShape
size_t
first_dim
=
0
;
for
(
int
i
=
0
;
i
<
N
;
i
++
)
{
auto
&
sel_row
=
get_selected_row
(
i
);
first_dim
+=
sel_row
.
rows
().
size
();
}
auto
in_dim
=
framework
::
vectorize
(
get_selected_row
(
N
-
1
).
value
().
dims
());
in_dim
[
0
]
=
static_cast
<
int64_t
>
(
first_dim
);
out_value
->
Resize
(
framework
::
make_ddim
(
in_dim
));
// if all the input sparse vars are empty, no need to
// merge these vars.
if
(
first_dim
==
0UL
)
{
return
;
}
out_value
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
math
::
SelectedRowsAddTo
<
CPUDeviceContext
,
T
>
functor
;
int64_t
offset
=
0
;
for
(
int
i
=
0
;
i
<
N
;
i
++
)
{
auto
&
sel_row
=
get_selected_row
(
i
);
if
(
sel_row
.
rows
().
size
()
==
0
)
{
continue
;
}
PADDLE_ENFORCE_EQ
(
out
->
height
(),
sel_row
.
height
());
functor
(
ctx
.
template
device_context
<
CPUDeviceContext
>(),
sel_row
,
offset
,
out
);
offset
+=
sel_row
.
value
().
numel
();
}
}
else
if
(
out_var
->
IsType
<
framework
::
LoDTensorArray
>
())
{
// TODO(@mozga-intel) Add MKLDNN LoDTensorArray support
auto
&
out_array
=
*
out_var
->
GetMutable
<
framework
::
LoDTensorArray
>
();
for
(
size_t
i
=
in_place
?
1
:
0
;
i
<
in_vars
.
size
();
++
i
)
{
PADDLE_ENFORCE
(
in_vars
[
i
]
->
IsType
<
framework
::
LoDTensorArray
>
(),
"Only support all inputs are TensorArray"
);
auto
&
in_array
=
in_vars
[
i
]
->
Get
<
framework
::
LoDTensorArray
>
();
for
(
size_t
i
=
0
;
i
<
in_array
.
size
();
++
i
)
{
if
(
in_array
[
i
].
numel
()
!=
0
)
{
if
(
i
>=
out_array
.
size
())
{
out_array
.
resize
(
i
+
1
);
}
if
(
out_array
[
i
].
numel
()
==
0
)
{
framework
::
TensorCopy
(
in_array
[
i
],
in_array
[
i
].
place
(),
ctx
.
device_context
(),
&
out_array
[
i
]);
out_array
[
i
].
set_lod
(
in_array
[
i
].
lod
());
}
else
{
PADDLE_ENFORCE
(
out_array
[
i
].
lod
()
==
in_array
[
i
].
lod
());
auto
in
=
EigenVector
<
T
>::
Flatten
(
in_array
[
i
]);
auto
result
=
EigenVector
<
T
>::
Flatten
(
out_array
[
i
]);
result
.
device
(
*
ctx
.
template
device_context
<
MKLDNNDeviceContext
>()
.
eigen_device
())
=
result
+
in
;
}
}
}
}
}
else
{
PADDLE_THROW
(
"Unexpected branch, output variable type is %s"
,
out_var
->
Type
().
name
());
}
}
};
}
// namespace operators
}
// namespace paddle
REGISTER_OP_KERNEL
(
sum
,
MKLDNN
,
::
paddle
::
platform
::
CPUPlace
,
paddle
::
operators
::
SumMKLDNNOpKernel
<
float
>
);
paddle/fluid/operators/sum_op.cc
浏览文件 @
96b4904d
...
@@ -18,6 +18,10 @@ limitations under the License. */
...
@@ -18,6 +18,10 @@ limitations under the License. */
#include "paddle/fluid/framework/var_type_inference.h"
#include "paddle/fluid/framework/var_type_inference.h"
#include "paddle/fluid/operators/detail/safe_ref.h"
#include "paddle/fluid/operators/detail/safe_ref.h"
#ifdef PADDLE_WITH_MKLDNN
#include "paddle/fluid/platform/mkldnn_helper.h"
#endif
namespace
paddle
{
namespace
paddle
{
namespace
operators
{
namespace
operators
{
using
framework
::
Tensor
;
using
framework
::
Tensor
;
...
@@ -63,6 +67,18 @@ class SumOp : public framework::OperatorWithKernel {
...
@@ -63,6 +67,18 @@ class SumOp : public framework::OperatorWithKernel {
framework
::
OpKernelType
GetExpectedKernelType
(
framework
::
OpKernelType
GetExpectedKernelType
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
auto
x_vars
=
ctx
.
MultiInputVar
(
"X"
);
auto
x_vars
=
ctx
.
MultiInputVar
(
"X"
);
framework
::
LibraryType
library
{
framework
::
LibraryType
::
kPlain
};
framework
::
DataLayout
layout
{
framework
::
DataLayout
::
kAnyLayout
};
#ifdef PADDLE_WITH_MKLDNN
if
(
library
==
framework
::
LibraryType
::
kPlain
&&
platform
::
CanMKLDNNBeUsed
(
ctx
))
{
library
=
framework
::
LibraryType
::
kMKLDNN
;
layout
=
framework
::
DataLayout
::
kMKLDNN
;
}
#endif
if
(
x_vars
[
0
]
->
IsType
<
framework
::
LoDTensor
>
())
{
if
(
x_vars
[
0
]
->
IsType
<
framework
::
LoDTensor
>
())
{
int
dtype
=
-
1
;
int
dtype
=
-
1
;
for
(
auto
&
x_var
:
x_vars
)
{
for
(
auto
&
x_var
:
x_vars
)
{
...
@@ -80,26 +96,27 @@ class SumOp : public framework::OperatorWithKernel {
...
@@ -80,26 +96,27 @@ class SumOp : public framework::OperatorWithKernel {
"Sum operator should have at least one tensor"
);
"Sum operator should have at least one tensor"
);
return
framework
::
OpKernelType
(
return
framework
::
OpKernelType
(
static_cast
<
framework
::
proto
::
VarType
::
Type
>
(
dtype
),
static_cast
<
framework
::
proto
::
VarType
::
Type
>
(
dtype
),
ctx
.
GetPlace
(),
ctx
.
device_context
()
);
layout
,
library
);
}
else
if
(
x_vars
[
0
]
->
IsType
<
framework
::
SelectedRows
>
())
{
}
else
if
(
x_vars
[
0
]
->
IsType
<
framework
::
SelectedRows
>
())
{
for
(
auto
&
var
:
x_vars
)
{
for
(
auto
&
var
:
x_vars
)
{
auto
&
value
=
var
->
Get
<
framework
::
SelectedRows
>
().
value
();
auto
&
value
=
var
->
Get
<
framework
::
SelectedRows
>
().
value
();
if
(
value
.
IsInitialized
())
{
if
(
value
.
IsInitialized
())
{
return
framework
::
OpKernelType
(
framework
::
ToDataType
(
value
.
type
()),
return
framework
::
OpKernelType
(
framework
::
ToDataType
(
value
.
type
()),
ctx
.
device_context
());
ctx
.
device_context
()
,
layout
,
library
);
}
}
}
}
// if input sparse vars are not initialized, use an default kernel type.
// if input sparse vars are not initialized, use an default kernel type.
return
framework
::
OpKernelType
(
framework
::
proto
::
VarType
::
FP32
,
return
framework
::
OpKernelType
(
framework
::
proto
::
VarType
::
FP32
,
ctx
.
device_context
());
ctx
.
device_context
()
,
layout
,
library
);
}
else
if
(
x_vars
[
0
]
->
IsType
<
framework
::
LoDTensorArray
>
())
{
}
else
if
(
x_vars
[
0
]
->
IsType
<
framework
::
LoDTensorArray
>
())
{
for
(
auto
&
x_var
:
x_vars
)
{
for
(
auto
&
x_var
:
x_vars
)
{
auto
&
array
=
x_var
->
Get
<
framework
::
LoDTensorArray
>
();
auto
&
array
=
x_var
->
Get
<
framework
::
LoDTensorArray
>
();
for
(
auto
&
each
:
array
)
{
for
(
auto
&
each
:
array
)
{
if
(
each
.
numel
()
!=
0
)
{
if
(
each
.
numel
()
!=
0
)
{
return
framework
::
OpKernelType
(
framework
::
ToDataType
(
each
.
type
()),
return
framework
::
OpKernelType
(
framework
::
ToDataType
(
each
.
type
()),
ctx
.
device_context
());
ctx
.
device_context
(),
layout
,
library
);
}
}
}
}
}
}
...
@@ -116,6 +133,9 @@ class SumOpMaker : public framework::OpProtoAndCheckerMaker {
...
@@ -116,6 +133,9 @@ class SumOpMaker : public framework::OpProtoAndCheckerMaker {
AddInput
(
"X"
,
"(vector<Tensor>) The input tensors of sum operator."
)
AddInput
(
"X"
,
"(vector<Tensor>) The input tensors of sum operator."
)
.
AsDuplicable
();
.
AsDuplicable
();
AddOutput
(
"Out"
,
"(Tensor) The output tensor of sum operator."
).
Reuse
(
"X"
);
AddOutput
(
"Out"
,
"(Tensor) The output tensor of sum operator."
).
Reuse
(
"X"
);
AddAttr
<
bool
>
(
"use_mkldnn"
,
"(bool, default false) Only used in mkldnn kernel"
)
.
SetDefault
(
false
);
AddComment
(
R"DOC(
AddComment
(
R"DOC(
Sum operator.
Sum operator.
...
@@ -132,7 +152,6 @@ class SumOpVarTypeInference : public framework::VarTypeInference {
...
@@ -132,7 +152,6 @@ class SumOpVarTypeInference : public framework::VarTypeInference {
framework
::
BlockDesc
*
block
)
const
override
{
framework
::
BlockDesc
*
block
)
const
override
{
auto
&
inputs
=
op_desc
.
Input
(
"X"
);
auto
&
inputs
=
op_desc
.
Input
(
"X"
);
auto
var_type
=
framework
::
proto
::
VarType
::
SELECTED_ROWS
;
auto
var_type
=
framework
::
proto
::
VarType
::
SELECTED_ROWS
;
for
(
auto
&
name
:
op_desc
.
Input
(
"X"
))
{
for
(
auto
&
name
:
op_desc
.
Input
(
"X"
))
{
VLOG
(
10
)
<<
name
<<
" "
VLOG
(
10
)
<<
name
<<
" "
<<
block
->
FindRecursiveOrCreateVar
(
name
).
GetType
();
<<
block
->
FindRecursiveOrCreateVar
(
name
).
GetType
();
...
@@ -206,6 +225,7 @@ namespace ops = paddle::operators;
...
@@ -206,6 +225,7 @@ namespace ops = paddle::operators;
REGISTER_OPERATOR
(
sum
,
ops
::
SumOp
,
ops
::
SumOpMaker
,
ops
::
SumGradMaker
,
REGISTER_OPERATOR
(
sum
,
ops
::
SumOp
,
ops
::
SumOpMaker
,
ops
::
SumGradMaker
,
ops
::
SumOpVarTypeInference
);
ops
::
SumOpVarTypeInference
);
REGISTER_OP_CPU_KERNEL
(
REGISTER_OP_CPU_KERNEL
(
sum
,
ops
::
SumKernel
<
paddle
::
platform
::
CPUDeviceContext
,
float
>
,
sum
,
ops
::
SumKernel
<
paddle
::
platform
::
CPUDeviceContext
,
float
>
,
ops
::
SumKernel
<
paddle
::
platform
::
CPUDeviceContext
,
double
>
,
ops
::
SumKernel
<
paddle
::
platform
::
CPUDeviceContext
,
double
>
,
...
...
paddle/fluid/operators/while_op.cc
浏览文件 @
96b4904d
...
@@ -203,11 +203,11 @@ class WhileGradOp : public framework::OperatorBase {
...
@@ -203,11 +203,11 @@ class WhileGradOp : public framework::OperatorBase {
->
set_lod
(
inside_tensor
.
lod
());
->
set_lod
(
inside_tensor
.
lod
());
}
}
}
}
auto
new_inside_name
=
cur_scope
.
Rename
(
inside_grad_name
);
auto
new_inside_name
=
cur_scope
.
Rename
(
inside_grad_name
);
auto
sum_op
=
framework
::
OpRegistry
::
CreateOp
(
auto
sum_op
=
framework
::
OpRegistry
::
CreateOp
(
"sum"
,
{{
"X"
,
{
pg_names
[
param_id
],
new_inside_name
}}},
"sum"
,
{{
"X"
,
{
pg_names
[
param_id
],
new_inside_name
}}},
{{
"Out"
,
{
pg_names
[
param_id
]}}},
framework
::
AttributeMap
{});
{{
"Out"
,
{
pg_names
[
param_id
]}}},
framework
::
AttributeMap
{{
"use_mkldnn"
,
{
false
}}});
sum_op
->
Run
(
cur_scope
,
dev_place
);
sum_op
->
Run
(
cur_scope
,
dev_place
);
cur_scope
.
Rename
(
new_inside_name
,
inside_grad_name
);
cur_scope
.
Rename
(
new_inside_name
,
inside_grad_name
);
}
}
...
...
python/paddle/fluid/backward.py
浏览文件 @
96b4904d
...
@@ -132,9 +132,9 @@ def _addup_repetitive_outputs_(op_descs):
...
@@ -132,9 +132,9 @@ def _addup_repetitive_outputs_(op_descs):
for
idx
,
op_desc
in
enumerate
(
op_descs
):
for
idx
,
op_desc
in
enumerate
(
op_descs
):
for
var_name
in
op_desc
.
input_arg_names
():
for
var_name
in
op_desc
.
input_arg_names
():
if
len
(
renamed_vars
[
var_name
])
>
1
:
if
len
(
renamed_vars
[
var_name
])
>
1
:
pending_sum_ops
.
append
(
pending_sum_ops
.
append
(
(
_create_op_desc_
(
(
_create_op_desc_
(
"sum"
,
{
"X"
:
renamed_vars
[
var_name
]},
"sum"
,
{
"X"
:
renamed_vars
[
var_name
]},
{
"Out"
:
[
var_name
]},
{
"Out"
:
[
var_name
]},
{
}),
idx
))
{
"use_mkldnn"
:
False
}),
idx
))
renamed_vars
[
var_name
]
=
[
var_name
]
renamed_vars
[
var_name
]
=
[
var_name
]
for
var_name
in
op_desc
.
output_arg_names
():
for
var_name
in
op_desc
.
output_arg_names
():
if
var_name
==
core
.
empty_var_name
(
if
var_name
==
core
.
empty_var_name
(
...
@@ -161,8 +161,9 @@ def _addup_repetitive_outputs_(op_descs):
...
@@ -161,8 +161,9 @@ def _addup_repetitive_outputs_(op_descs):
renamed_vars
[
var_name
].
append
(
new_name
)
renamed_vars
[
var_name
].
append
(
new_name
)
for
var_name
,
inputs
in
renamed_vars
.
iteritems
():
for
var_name
,
inputs
in
renamed_vars
.
iteritems
():
if
len
(
inputs
)
>
1
:
if
len
(
inputs
)
>
1
:
pending_sum_ops
.
append
((
_create_op_desc_
(
pending_sum_ops
.
append
(
"sum"
,
{
"X"
:
inputs
},
{
"Out"
:
[
var_name
]},
{}),
len
(
op_descs
)))
(
_create_op_desc_
(
"sum"
,
{
"X"
:
inputs
},
{
"Out"
:
[
var_name
]},
{
"use_mkldnn"
:
False
}),
len
(
op_descs
)))
# sum_op descs are sorted according to their insert position
# sum_op descs are sorted according to their insert position
for
p
in
reversed
(
pending_sum_ops
):
for
p
in
reversed
(
pending_sum_ops
):
op_descs
.
insert
(
p
[
1
],
p
[
0
])
op_descs
.
insert
(
p
[
1
],
p
[
0
])
...
...
python/paddle/fluid/layers/nn.py
浏览文件 @
96b4904d
...
@@ -197,7 +197,10 @@ def fc(input,
...
@@ -197,7 +197,10 @@ def fc(input,
else
:
else
:
pre_bias
=
helper
.
create_tmp_variable
(
dtype
)
pre_bias
=
helper
.
create_tmp_variable
(
dtype
)
helper
.
append_op
(
helper
.
append_op
(
type
=
"sum"
,
inputs
=
{
"X"
:
mul_results
},
outputs
=
{
"Out"
:
pre_bias
})
type
=
"sum"
,
inputs
=
{
"X"
:
mul_results
},
outputs
=
{
"Out"
:
pre_bias
},
attrs
=
{
"use_mkldnn"
:
use_mkldnn
})
# add bias
# add bias
pre_activation
=
helper
.
append_bias_op
(
pre_bias
,
dim_start
=
num_flatten_dims
)
pre_activation
=
helper
.
append_bias_op
(
pre_bias
,
dim_start
=
num_flatten_dims
)
# add activation
# add activation
...
...
python/paddle/fluid/layers/tensor.py
浏览文件 @
96b4904d
...
@@ -230,7 +230,11 @@ def sums(input, out=None):
...
@@ -230,7 +230,11 @@ def sums(input, out=None):
helper
=
LayerHelper
(
'sum'
,
**
locals
())
helper
=
LayerHelper
(
'sum'
,
**
locals
())
if
out
is
None
:
if
out
is
None
:
out
=
helper
.
create_tmp_variable
(
dtype
=
helper
.
input_dtype
())
out
=
helper
.
create_tmp_variable
(
dtype
=
helper
.
input_dtype
())
helper
.
append_op
(
type
=
'sum'
,
inputs
=
{
'X'
:
input
},
outputs
=
{
'Out'
:
out
})
helper
.
append_op
(
type
=
'sum'
,
inputs
=
{
'X'
:
input
},
outputs
=
{
'Out'
:
out
},
attrs
=
{
'use_mkldnn'
:
False
})
return
out
return
out
...
...
python/paddle/fluid/transpiler/distribute_transpiler.py
浏览文件 @
96b4904d
...
@@ -824,7 +824,8 @@ class DistributeTranspiler:
...
@@ -824,7 +824,8 @@ class DistributeTranspiler:
table_opt_block
.
append_op
(
table_opt_block
.
append_op
(
type
=
"sum"
,
type
=
"sum"
,
inputs
=
{
"X"
:
pserver_side_table_grad_list
},
inputs
=
{
"X"
:
pserver_side_table_grad_list
},
outputs
=
{
"Out"
:
[
grad_var
]})
outputs
=
{
"Out"
:
[
grad_var
]},
attrs
=
{
"use_mkldnn"
:
False
})
else
:
else
:
# in async_mode, for table gradient, it also need to be splited to each parameter server
# in async_mode, for table gradient, it also need to be splited to each parameter server
origin_grad_name
=
grad_var
.
name
origin_grad_name
=
grad_var
.
name
...
@@ -1056,7 +1057,8 @@ class DistributeTranspiler:
...
@@ -1056,7 +1057,8 @@ class DistributeTranspiler:
optimize_block
.
append_op
(
optimize_block
.
append_op
(
type
=
"sum"
,
type
=
"sum"
,
inputs
=
{
"X"
:
vars2merge
},
inputs
=
{
"X"
:
vars2merge
},
outputs
=
{
"Out"
:
merged_var
})
outputs
=
{
"Out"
:
merged_var
},
attrs
=
{
"use_mkldnn"
:
False
})
# TODO(panyx0718): What if it's SELECTED_ROWS.
# TODO(panyx0718): What if it's SELECTED_ROWS.
if
not
merged_var
.
type
==
core
.
VarDesc
.
VarType
.
SELECTED_ROWS
:
if
not
merged_var
.
type
==
core
.
VarDesc
.
VarType
.
SELECTED_ROWS
:
optimize_block
.
append_op
(
optimize_block
.
append_op
(
...
...
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