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
机器未来
Paddle
提交
8493f20e
P
Paddle
项目概览
机器未来
/
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 >>
提交
8493f20e
编写于
11月 27, 2019
作者:
G
GaoWei8
提交者:
Yiqun Liu
11月 27, 2019
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Polish the codes of fc when needs padding (#21378)
test=develop
上级
5d7d5482
变更
2
隐藏空白更改
内联
并排
Showing
2 changed file
with
14 addition
and
14 deletion
+14
-14
paddle/fluid/framework/ir/fc_fuse_pass.cc
paddle/fluid/framework/ir/fc_fuse_pass.cc
+3
-3
paddle/fluid/operators/math/fc.cc
paddle/fluid/operators/math/fc.cc
+11
-11
未找到文件。
paddle/fluid/framework/ir/fc_fuse_pass.cc
浏览文件 @
8493f20e
...
...
@@ -94,20 +94,20 @@ int FCFusePass::ApplyFCPattern(Graph* graph, bool with_relu) const {
auto
*
weight
=
scope
->
FindVar
(
w
->
Name
())
->
GetMutable
<
LoDTensor
>
();
auto
place
=
weight
->
place
();
bool
use_gpu
=
Get
<
bool
>
(
"use_gpu"
);
auto
weight_data
=
weight
->
data
<
float
>
();
auto
*
weight_data
=
weight
->
data
<
float
>
();
auto
weight_dims
=
weight
->
dims
();
int
weight_num
=
product
(
weight_dims
);
int
w_h
=
weight_dims
[
0
];
int
w_w
=
weight_dims
[
1
];
if
(
!
use_gpu
)
{
if
(
w_h
%
128
==
0
&&
w_w
%
128
==
0
)
{
float
*
weight_data_tmp
=
new
float
[
weight_num
];
auto
*
weight_data_tmp
=
new
float
[
weight_num
];
for
(
int
i
=
0
;
i
<
w_h
;
i
++
)
{
memcpy
(
weight_data_tmp
+
i
*
w_w
,
weight_data
+
i
*
w_w
,
w_w
*
sizeof
(
float
));
}
weight
->
Resize
(
DDim
{
weight_dims
[
0
]
+
4
,
weight_dims
[
1
]
+
4
});
auto
weight_data_new
=
auto
*
weight_data_new
=
weight
->
mutable_data
<
float
>
(
platform
::
CPUPlace
());
for
(
int
i
=
0
;
i
<
w_h
;
i
++
)
{
memcpy
(
weight_data_new
+
i
*
(
w_w
+
4
),
weight_data_tmp
+
i
*
w_w
,
...
...
paddle/fluid/operators/math/fc.cc
浏览文件 @
8493f20e
...
...
@@ -30,28 +30,28 @@ class FCFunctor<platform::CPUDeviceContext, T> {
auto
blas
=
math
::
GetBlas
<
platform
::
CPUDeviceContext
,
T
>
(
context
);
framework
::
Tensor
Y1
;
T
*
Y1_data
=
nullptr
;
if
(
N
%
128
==
0
&&
K
%
128
==
0
)
{
auto
padding
=
N
%
128
==
0
&&
K
%
128
==
0
;
if
(
padding
)
{
const
int
NN
=
N
+
4
;
const
int
KK
=
K
+
4
;
framework
::
Tensor
X1
;
T
*
X1_data
=
X1
.
Resize
({
M
*
KK
}).
mutable_data
<
T
>
(
platform
::
CPUPlace
());
Y1_data
=
Y1
.
Resize
({
M
*
(
N
+
4
)}).
mutable_data
<
T
>
(
platform
::
CPUPlace
());
T
*
X1_data
=
X1
.
mutable_data
<
T
>
({
M
*
KK
},
platform
::
CPUPlace
());
Y1_data
=
Y1
.
mutable_data
<
T
>
({
M
*
(
N
+
4
)},
platform
::
CPUPlace
());
#ifdef PADDLE_WITH_MKLML
#pragma omp parallel for
#endif
for
(
int
i
=
0
;
i
<
M
;
i
++
)
{
memcpy
(
X1_data
+
i
*
KK
,
X
+
i
*
K
,
K
*
sizeof
(
X
[
0
]
));
memcpy
(
X1_data
+
i
*
KK
,
X
+
i
*
K
,
K
*
sizeof
(
T
));
}
framework
::
Tensor
W1
;
T
*
W1_data
=
nullptr
;
if
(
!
padding_weights
)
{
W1_data
=
W1
.
Resize
({(
K
+
4
)
*
(
N
+
4
)})
.
mutable_data
<
T
>
(
platform
::
CPUPlace
());
W1_data
=
W1
.
mutable_data
<
T
>
({(
K
+
4
)
*
(
N
+
4
)},
platform
::
CPUPlace
());
#ifdef PADDLE_WITH_MKLML
#pragma omp parallel for
#endif
for
(
int
i
=
0
;
i
<
K
;
i
++
)
{
memcpy
(
W1_data
+
i
*
NN
,
W
+
i
*
N
,
N
*
sizeof
(
W
[
0
]
));
memcpy
(
W1_data
+
i
*
NN
,
W
+
i
*
N
,
N
*
sizeof
(
T
));
}
}
blas
.
GEMM
(
false
,
false
,
M
,
N
,
K
,
static_cast
<
T
>
(
1.0
),
X1_data
,
KK
,
...
...
@@ -61,12 +61,12 @@ class FCFunctor<platform::CPUDeviceContext, T> {
blas
.
MatMul
(
M
,
N
,
K
,
X
,
W
,
Y
);
}
if
(
B
==
NULL
)
{
if
(
N
%
128
==
0
&&
K
%
128
==
0
)
{
if
(
padding
)
{
#ifdef PADDLE_WITH_MKLML
#pragma omp parallel for
#endif
for
(
int
i
=
0
;
i
<
M
;
i
++
)
{
memcpy
(
Y
+
i
*
N
,
Y1_data
+
i
*
(
N
+
4
),
N
*
sizeof
(
Y
[
0
]
));
memcpy
(
Y
+
i
*
N
,
Y1_data
+
i
*
(
N
+
4
),
N
*
sizeof
(
T
));
}
}
PADDLE_ENFORCE_EQ
(
relu
,
false
,
...
...
@@ -80,7 +80,7 @@ class FCFunctor<platform::CPUDeviceContext, T> {
.
At
(
N
);
for
(
int
i
=
0
;
i
<
M
;
i
++
)
{
T
*
dst
=
Y
+
i
*
N
;
T
*
src
=
(
N
%
128
==
0
&&
K
%
128
==
0
)
?
Y1_data
+
i
*
(
N
+
4
)
:
dst
;
T
*
src
=
(
padding
)
?
Y1_data
+
i
*
(
N
+
4
)
:
dst
;
compute
(
B
,
src
,
dst
,
N
);
}
}
else
{
...
...
@@ -92,7 +92,7 @@ class FCFunctor<platform::CPUDeviceContext, T> {
#endif
for
(
int
i
=
0
;
i
<
M
;
i
++
)
{
T
*
dst
=
Y
+
i
*
N
;
T
*
src
=
(
N
%
128
==
0
&&
K
%
128
==
0
)
?
Y1_data
+
i
*
(
N
+
4
)
:
dst
;
T
*
src
=
(
padding
)
?
Y1_data
+
i
*
(
N
+
4
)
:
dst
;
compute
(
B
,
src
,
dst
,
N
);
}
}
...
...
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