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-Lite
提交
412ee03a
P
Paddle-Lite
项目概览
PaddlePaddle
/
Paddle-Lite
通知
331
Star
4
Fork
1
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
271
列表
看板
标记
里程碑
合并请求
78
Wiki
0
Wiki
分析
仓库
DevOps
项目成员
Pages
P
Paddle-Lite
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
271
Issue
271
列表
看板
标记
里程碑
合并请求
78
合并请求
78
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
未验证
提交
412ee03a
编写于
9月 06, 2018
作者:
S
smilejames
提交者:
GitHub
9月 06, 2018
浏览文件
操作
浏览文件
下载
差异文件
Merge pull request #919 from smilejames/develop
add border check in gemm
上级
4ea1c935
de258be5
变更
2
隐藏空白更改
内联
并排
Showing
2 changed file
with
93 addition
and
69 deletion
+93
-69
src/operators/math/gemm.cpp
src/operators/math/gemm.cpp
+84
-39
src/operators/math/math_function.cpp
src/operators/math/math_function.cpp
+9
-30
未找到文件。
src/operators/math/gemm.cpp
浏览文件 @
412ee03a
...
...
@@ -3002,14 +3002,21 @@ void Sgemm(int m, int n, int k, float alpha, const float *A, int lda,
NC
=
L2
/
(
KC
*
sizeof
(
float
));
// make sure MC is multiple of MR, and NC is multiple of NR
int
mblock_num
=
(
m
+
MC
-
1
)
/
MC
;
MC
=
(
m
+
mblock_num
-
1
)
/
mblock_num
;
MC
=
(
MC
+
MR
-
1
)
/
MR
*
MR
;
if
(
MC
==
0
)
{
MC
=
MR
;
}
else
{
int
mblock_num
=
(
m
+
MC
-
1
)
/
MC
;
MC
=
(
m
+
mblock_num
-
1
)
/
mblock_num
;
MC
=
(
MC
+
MR
-
1
)
/
MR
*
MR
;
}
// DLOG << "mblock_num = " << mblock_num << ", MC = " << MC << "\n";
int
nblock_num
=
(
n
+
NC
-
1
)
/
NC
;
NC
=
(
n
+
nblock_num
-
1
)
/
nblock_num
;
NC
=
(
NC
+
NR
-
1
)
/
NR
*
NR
;
if
(
NC
==
0
)
{
NC
=
NR
;
}
else
{
int
nblock_num
=
(
n
+
NC
-
1
)
/
NC
;
NC
=
(
n
+
nblock_num
-
1
)
/
nblock_num
;
NC
=
(
NC
+
NR
-
1
)
/
NR
*
NR
;
}
// DLOG << "nblock_num = " << nblock_num << ", NC = " << NC << "\n";
packedA
=
static_cast
<
float
*>
(
...
...
@@ -3067,14 +3074,21 @@ void SgemmWithBn(int m, int n, int k, float alpha, const float *A, int lda,
NC
=
L2
/
(
KC
*
sizeof
(
float
));
// make sure MC is multiple of MR, and NC is multiple of NR
int
mblock_num
=
(
m
+
MC
-
1
)
/
MC
;
MC
=
(
m
+
mblock_num
-
1
)
/
mblock_num
;
MC
=
(
MC
+
MR
-
1
)
/
MR
*
MR
;
if
(
MC
==
0
)
{
MC
=
MR
;
}
else
{
int
mblock_num
=
(
m
+
MC
-
1
)
/
MC
;
MC
=
(
m
+
mblock_num
-
1
)
/
mblock_num
;
MC
=
(
MC
+
MR
-
1
)
/
MR
*
MR
;
}
// DLOG << "mblock_num = " << mblock_num << ", MC = " << MC << "\n";
int
nblock_num
=
(
n
+
NC
-
1
)
/
NC
;
NC
=
(
n
+
nblock_num
-
1
)
/
nblock_num
;
NC
=
(
NC
+
NR
-
1
)
/
NR
*
NR
;
if
(
NC
==
0
)
{
NC
=
NR
;
}
else
{
int
nblock_num
=
(
n
+
NC
-
1
)
/
NC
;
NC
=
(
n
+
nblock_num
-
1
)
/
nblock_num
;
NC
=
(
NC
+
NR
-
1
)
/
NR
*
NR
;
}
// DLOG << "nblock_num = " << nblock_num << ", NC = " << NC << "\n";
packedA
=
static_cast
<
float
*>
(
...
...
@@ -3133,14 +3147,21 @@ void SgemmWithPRelu(int m, int n, int k, const float *A, int lda,
NC
=
L2
/
(
KC
*
sizeof
(
float
));
// make sure MC is multiple of MR, and NC is multiple of NR
int
mblock_num
=
(
m
+
MC
-
1
)
/
MC
;
MC
=
(
m
+
mblock_num
-
1
)
/
mblock_num
;
MC
=
(
MC
+
MR
-
1
)
/
MR
*
MR
;
if
(
MC
==
0
)
{
MC
=
MR
;
}
else
{
int
mblock_num
=
(
m
+
MC
-
1
)
/
MC
;
MC
=
(
m
+
mblock_num
-
1
)
/
mblock_num
;
MC
=
(
MC
+
MR
-
1
)
/
MR
*
MR
;
}
// DLOG << "mblock_num = " << mblock_num << ", MC = " << MC << "\n";
int
nblock_num
=
(
n
+
NC
-
1
)
/
NC
;
NC
=
(
n
+
nblock_num
-
1
)
/
nblock_num
;
NC
=
(
NC
+
NR
-
1
)
/
NR
*
NR
;
if
(
NC
==
0
)
{
NC
=
NR
;
}
else
{
int
nblock_num
=
(
n
+
NC
-
1
)
/
NC
;
NC
=
(
n
+
nblock_num
-
1
)
/
nblock_num
;
NC
=
(
NC
+
NR
-
1
)
/
NR
*
NR
;
}
// DLOG << "nblock_num = " << nblock_num << ", NC = " << NC << "\n";
packedA
=
static_cast
<
float
*>
(
...
...
@@ -3203,9 +3224,13 @@ void Sgemm_omp(int m, int n, int k, float alpha, const float *A, int lda,
if
(
m
>
n
)
{
// 对 A 分块
MC
=
L1
/
(
KC
*
sizeof
(
float
));
int
mblock_num
=
(
m
+
MC
-
1
)
/
MC
;
MC
=
(
m
+
mblock_num
-
1
)
/
mblock_num
;
MC
=
(
MC
+
MR
-
1
)
/
MR
*
MR
;
if
(
MC
==
0
)
{
MC
=
MR
;
}
else
{
int
mblock_num
=
(
m
+
MC
-
1
)
/
MC
;
MC
=
(
m
+
mblock_num
-
1
)
/
mblock_num
;
MC
=
(
MC
+
MR
-
1
)
/
MR
*
MR
;
}
// 补齐 B
NC
=
(
n
+
NR
-
1
)
/
NR
*
NR
;
...
...
@@ -3227,9 +3252,13 @@ void Sgemm_omp(int m, int n, int k, float alpha, const float *A, int lda,
}
else
{
// 对 B 分块
NC
=
L1
/
(
KC
*
sizeof
(
float
));
int
nblock_num
=
(
n
+
NC
-
1
)
/
NC
;
NC
=
(
n
+
nblock_num
-
1
)
/
nblock_num
;
NC
=
(
NC
+
NR
-
1
)
/
NR
*
NR
;
if
(
NC
==
0
)
{
NC
=
NR
;
}
else
{
int
nblock_num
=
(
n
+
NC
-
1
)
/
NC
;
NC
=
(
n
+
nblock_num
-
1
)
/
nblock_num
;
NC
=
(
NC
+
NR
-
1
)
/
NR
*
NR
;
}
// 补齐 A
MC
=
(
m
+
MR
-
1
)
/
MR
*
MR
;
...
...
@@ -3311,9 +3340,13 @@ void SgemmWithBn_omp(int m, int n, int k, float alpha, const float *A, int lda,
if
(
m
>
n
)
{
// 对 A 分块
MC
=
L1
/
(
KC
*
sizeof
(
float
));
int
mblock_num
=
(
m
+
MC
-
1
)
/
MC
;
MC
=
(
m
+
mblock_num
-
1
)
/
mblock_num
;
MC
=
(
MC
+
MR
-
1
)
/
MR
*
MR
;
if
(
MC
==
0
)
{
MC
=
MR
;
}
else
{
int
mblock_num
=
(
m
+
MC
-
1
)
/
MC
;
MC
=
(
m
+
mblock_num
-
1
)
/
mblock_num
;
MC
=
(
MC
+
MR
-
1
)
/
MR
*
MR
;
}
// 补齐 B
NC
=
(
n
+
NR
-
1
)
/
NR
*
NR
;
...
...
@@ -3335,9 +3368,13 @@ void SgemmWithBn_omp(int m, int n, int k, float alpha, const float *A, int lda,
}
else
{
// 对 B 分块
NC
=
L1
/
(
KC
*
sizeof
(
float
));
int
nblock_num
=
(
n
+
NC
-
1
)
/
NC
;
NC
=
(
n
+
nblock_num
-
1
)
/
nblock_num
;
NC
=
(
NC
+
NR
-
1
)
/
NR
*
NR
;
if
(
NC
==
0
)
{
NC
==
NR
;
}
else
{
int
nblock_num
=
(
n
+
NC
-
1
)
/
NC
;
NC
=
(
n
+
nblock_num
-
1
)
/
nblock_num
;
NC
=
(
NC
+
NR
-
1
)
/
NR
*
NR
;
}
// 补齐 A
MC
=
(
m
+
MR
-
1
)
/
MR
*
MR
;
...
...
@@ -3430,9 +3467,13 @@ void SgemmWithPRelu_omp(int m, int n, int k, const float *A, int lda,
if
(
m
>
n
)
{
// 对 A 分块
MC
=
L1
/
(
KC
*
sizeof
(
float
));
int
mblock_num
=
(
m
+
MC
-
1
)
/
MC
;
MC
=
(
m
+
mblock_num
-
1
)
/
mblock_num
;
MC
=
(
MC
+
MR
-
1
)
/
MR
*
MR
;
if
(
MC
==
0
)
{
MC
=
MR
;
}
else
{
int
mblock_num
=
(
m
+
MC
-
1
)
/
MC
;
MC
=
(
m
+
mblock_num
-
1
)
/
mblock_num
;
MC
=
(
MC
+
MR
-
1
)
/
MR
*
MR
;
}
// 补齐 B
NC
=
(
n
+
NR
-
1
)
/
NR
*
NR
;
...
...
@@ -3454,9 +3495,13 @@ void SgemmWithPRelu_omp(int m, int n, int k, const float *A, int lda,
}
else
{
// 对 B 分块
NC
=
L1
/
(
KC
*
sizeof
(
float
));
int
nblock_num
=
(
n
+
NC
-
1
)
/
NC
;
NC
=
(
n
+
nblock_num
-
1
)
/
nblock_num
;
NC
=
(
NC
+
NR
-
1
)
/
NR
*
NR
;
if
(
NC
==
0
)
{
NC
=
NR
;
}
else
{
int
nblock_num
=
(
n
+
NC
-
1
)
/
NC
;
NC
=
(
n
+
nblock_num
-
1
)
/
nblock_num
;
NC
=
(
NC
+
NR
-
1
)
/
NR
*
NR
;
}
// 补齐 A
MC
=
(
m
+
MR
-
1
)
/
MR
*
MR
;
...
...
src/operators/math/math_function.cpp
浏览文件 @
412ee03a
...
...
@@ -28,16 +28,9 @@ void matmul<float>(const framework::Tensor &matrix_a, bool trans_a,
auto
dim_a
=
matrix_a
.
dims
();
auto
dim_b
=
matrix_b
.
dims
();
auto
dim_out
=
matrix_out
->
dims
();
// PADDLE_ENFORCE(dim_a.size() == 2 && dim_b.size() == 2 &&
// dim_out.size() ==
// 2,
// "The input and output of matmul be matrix");
//
// PADDLE_ENFORCE(platform::is_cpu_place(matrix_a.place()) &&
// platform::is_cpu_place(matrix_b.place())
// &&
// platform::is_cpu_place(matrix_out->place()),
// "Matrix must all be in CPUPlace");
PADDLE_MOBILE_ENFORCE
(
dim_a
.
size
()
==
2
&&
dim_b
.
size
()
==
2
&&
dim_out
.
size
()
==
2
,
"The input and output of matmul be matrix"
);
int
M
=
dim_out
[
0
];
int
N
=
dim_out
[
1
];
...
...
@@ -61,16 +54,9 @@ void matmulWithBn<float>(const framework::Tensor &matrix_a, bool trans_a,
auto
dim_a
=
matrix_a
.
dims
();
auto
dim_b
=
matrix_b
.
dims
();
auto
dim_out
=
matrix_out
->
dims
();
// PADDLE_ENFORCE(dim_a.size() == 2 && dim_b.size() == 2 &&
// dim_out.size() ==
// 2,
// "The input and output of matmul be matrix");
//
// PADDLE_ENFORCE(platform::is_cpu_place(matrix_a.place()) &&
// platform::is_cpu_place(matrix_b.place())
// &&
// platform::is_cpu_place(matrix_out->place()),
// "Matrix must all be in CPUPlace");
PADDLE_MOBILE_ENFORCE
(
dim_a
.
size
()
==
2
&&
dim_b
.
size
()
==
2
&&
dim_out
.
size
()
==
2
,
"The input and output of matmul be matrix"
);
int
M
=
dim_out
[
0
];
int
N
=
dim_out
[
1
];
...
...
@@ -95,16 +81,9 @@ void matmulWithPRelu(const framework::Tensor &matrix_a, bool trans_a,
auto
dim_a
=
matrix_a
.
dims
();
auto
dim_b
=
matrix_b
.
dims
();
auto
dim_out
=
matrix_out
->
dims
();
// PADDLE_ENFORCE(dim_a.size() == 2 && dim_b.size() == 2 &&
// dim_out.size() ==
// 2,
// "The input and output of matmul be matrix");
//
// PADDLE_ENFORCE(platform::is_cpu_place(matrix_a.place()) &&
// platform::is_cpu_place(matrix_b.place())
// &&
// platform::is_cpu_place(matrix_out->place()),
// "Matrix must all be in CPUPlace");
PADDLE_MOBILE_ENFORCE
(
dim_a
.
size
()
==
2
&&
dim_b
.
size
()
==
2
&&
dim_out
.
size
()
==
2
,
"The input and output of matmul be matrix"
);
int
M
=
dim_out
[
0
];
int
N
=
dim_out
[
1
];
...
...
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