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
csdn_franckjun
Paddle
提交
f72ab896
P
Paddle
项目概览
csdn_franckjun
/
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
f72ab896
编写于
8月 22, 2018
作者:
T
tensor-tang
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
refine blas gemm
上级
f5d5d7b2
变更
4
隐藏空白更改
内联
并排
Showing
4 changed file
with
77 addition
and
76 deletion
+77
-76
CMakeLists.txt
CMakeLists.txt
+0
-6
paddle/fluid/operators/math/blas.h
paddle/fluid/operators/math/blas.h
+9
-0
paddle/fluid/operators/math/blas_impl.h
paddle/fluid/operators/math/blas_impl.h
+53
-63
paddle/fluid/operators/math/fc_compute.h
paddle/fluid/operators/math/fc_compute.h
+15
-7
未找到文件。
CMakeLists.txt
浏览文件 @
f72ab896
...
...
@@ -138,12 +138,6 @@ else()
set
(
THIRD_PARTY_BUILD_TYPE Release
)
endif
()
if
(
WITH_MKL
)
option
(
MKL_SPLIT_GEMM
"PaddlePaddle MKL gemm would split to small ones"
OFF
)
if
(
MKL_SPLIT_GEMM
)
add_definitions
(
-DPADDLE_MKL_SPLIT_GEMM
)
endif
()
endif
()
set
(
WITH_MKLML
${
WITH_MKL
}
)
if
(
NOT DEFINED WITH_MKLDNN
)
if
(
WITH_MKL AND AVX2_FOUND
)
...
...
paddle/fluid/operators/math/blas.h
浏览文件 @
f72ab896
...
...
@@ -90,6 +90,11 @@ class Blas {
void
GEMM
(
bool
transA
,
bool
transB
,
int
M
,
int
N
,
int
K
,
T
alpha
,
const
T
*
A
,
int
lda
,
const
T
*
B
,
int
ldb
,
T
beta
,
T
*
C
,
int
ldc
)
const
;
template
<
typename
T
>
void
GEMM
(
CBLAS_TRANSPOSE
transA
,
CBLAS_TRANSPOSE
transB
,
int
M
,
int
N
,
int
K
,
T
alpha
,
const
T
*
A
,
int
lda
,
const
T
*
B
,
int
ldb
,
T
beta
,
T
*
C
,
int
ldc
)
const
;
#ifdef PADDLE_WITH_MKLML
template
<
typename
T
>
T
*
GEMM_ALLOC
(
const
CBLAS_IDENTIFIER
id
,
const
int
M
,
const
int
N
,
...
...
@@ -109,6 +114,10 @@ class Blas {
void
GEMM_FREE
(
T
*
data
)
const
;
#endif
template
<
typename
T
>
void
MatMul
(
const
int
M
,
const
int
N
,
const
int
K
,
const
T
*
A
,
const
T
*
B
,
T
*
C
)
const
;
template
<
typename
T
>
void
MatMul
(
const
framework
::
Tensor
&
mat_a
,
bool
trans_a
,
const
framework
::
Tensor
&
mat_b
,
bool
trans_b
,
T
alpha
,
...
...
paddle/fluid/operators/math/blas_impl.h
浏览文件 @
f72ab896
...
...
@@ -217,64 +217,6 @@ struct CBlas<platform::float16> {
#endif
};
template
<
typename
T
>
inline
bool
UseXSMM
(
const
int
&
m
,
const
int
&
n
,
const
int
&
k
,
bool
transa
,
bool
transb
,
const
T
&
alpha
,
const
T
&
beta
)
{
#ifdef PADDLE_WITH_LIBXSMM
// Refer to https://github.com/hfp/libxsmm/blob/master/README.md
// But the threshold is custom
constexpr
int
LIBXSMM_THRESHOLD
=
20
*
20
*
20
;
if
(
m
*
n
*
k
>
LIBXSMM_THRESHOLD
||
transa
||
transb
||
std
::
abs
<
T
>
(
alpha
-
static_cast
<
T
>
(
1
)
>
std
::
numeric_limits
<
T
>::
epsilon
())
||
std
::
abs
<
T
>
(
beta
)
>
std
::
numeric_limits
<
T
>::
epsilon
())
{
return
false
;
}
else
{
return
true
;
}
#endif
return
false
;
}
template
<
>
inline
bool
UseXSMM
<
platform
::
float16
>
(
const
int
&
m
,
const
int
&
n
,
const
int
&
k
,
bool
transa
,
bool
transb
,
const
platform
::
float16
&
alpha
,
const
platform
::
float16
&
beta
)
{
return
false
;
}
template
<
typename
T
>
inline
void
GEMM_WARP
(
CBLAS_ORDER
order
,
CBLAS_TRANSPOSE
transA
,
CBLAS_TRANSPOSE
transB
,
int
M
,
int
N
,
int
K
,
T
alpha
,
const
T
*
A
,
int
lda
,
const
T
*
B
,
int
ldb
,
T
beta
,
T
*
C
,
int
ldc
)
{
#ifdef PADDLE_WITH_LIBXSMM
if
(
UseXSMM
<
T
>
(
M
,
N
,
K
,
transA
!=
CblasNoTrans
,
transB
!=
CblasNoTrans
,
alpha
,
beta
))
{
// Note: SMM use ColMajor
const
char
transa
=
'N'
;
const
char
transb
=
'N'
;
CBlas
<
T
>::
SMM_GEMM
(
&
transa
,
&
transb
,
&
N
,
&
M
,
&
K
,
&
alpha
,
B
,
&
ldb
,
A
,
&
lda
,
&
beta
,
C
,
&
ldc
);
return
;
}
#endif
#ifdef PADDLE_MKL_SPLIT_GEMM
constexpr
int
bs
=
2
;
if
(
M
%
bs
==
0
&&
transA
==
CblasNoTrans
&&
transB
==
CblasNoTrans
)
{
for
(
int
off
=
0
;
off
<
M
;
off
+=
bs
)
{
CBlas
<
T
>::
GEMM
(
CblasRowMajor
,
CblasNoTrans
,
CblasNoTrans
,
bs
,
N
,
K
,
alpha
,
A
+
off
*
lda
,
lda
,
B
,
ldb
,
beta
,
C
+
off
*
ldb
,
ldc
);
}
return
;
}
#endif
CBlas
<
T
>::
GEMM
(
CblasRowMajor
,
transA
,
transB
,
M
,
N
,
K
,
alpha
,
A
,
lda
,
B
,
ldb
,
beta
,
C
,
ldc
);
}
#ifdef PADDLE_WITH_MKLML
template
<
>
template
<
typename
T
>
...
...
@@ -319,8 +261,8 @@ void Blas<platform::CPUDeviceContext>::GEMM(CBLAS_TRANSPOSE transA,
int
lda
=
(
transA
==
CblasNoTrans
)
?
K
:
M
;
int
ldb
=
(
transB
==
CblasNoTrans
)
?
N
:
K
;
int
ldc
=
N
;
GEMM_WARP
<
T
>
(
CblasRowMajor
,
transA
,
transB
,
M
,
N
,
K
,
alpha
,
A
,
lda
,
B
,
ldb
,
beta
,
C
,
ldc
);
CBlas
<
T
>::
GEMM
(
CblasRowMajor
,
transA
,
transB
,
M
,
N
,
K
,
alpha
,
A
,
lda
,
B
,
ldb
,
beta
,
C
,
ldc
);
}
template
<
>
...
...
@@ -329,9 +271,20 @@ void Blas<platform::CPUDeviceContext>::GEMM(bool transA, bool transB, int M,
int
N
,
int
K
,
T
alpha
,
const
T
*
A
,
int
lda
,
const
T
*
B
,
int
ldb
,
T
beta
,
T
*
C
,
int
ldc
)
const
{
GEMM_WARP
<
T
>
(
CblasRowMajor
,
transA
==
false
?
CblasNoTrans
:
CblasTrans
,
transB
==
false
?
CblasNoTrans
:
CblasTrans
,
M
,
N
,
K
,
alpha
,
A
,
lda
,
B
,
ldb
,
beta
,
C
,
ldc
);
CBlas
<
T
>::
GEMM
(
CblasRowMajor
,
transA
==
false
?
CblasNoTrans
:
CblasTrans
,
transB
==
false
?
CblasNoTrans
:
CblasTrans
,
M
,
N
,
K
,
alpha
,
A
,
lda
,
B
,
ldb
,
beta
,
C
,
ldc
);
}
template
<
>
template
<
typename
T
>
void
Blas
<
platform
::
CPUDeviceContext
>::
GEMM
(
CBLAS_TRANSPOSE
transA
,
CBLAS_TRANSPOSE
transB
,
int
M
,
int
N
,
int
K
,
T
alpha
,
const
T
*
A
,
int
lda
,
const
T
*
B
,
int
ldb
,
T
beta
,
T
*
C
,
int
ldc
)
const
{
CBlas
<
T
>::
GEMM
(
CblasRowMajor
,
transA
,
transB
,
M
,
N
,
K
,
alpha
,
A
,
lda
,
B
,
ldb
,
beta
,
C
,
ldc
);
}
template
<
typename
DeviceContext
>
...
...
@@ -440,6 +393,43 @@ void Blas<platform::CPUDeviceContext>::BatchedGEMM(
#endif
}
template
<
typename
DeviceContext
>
template
<
typename
T
>
void
Blas
<
DeviceContext
>::
MatMul
(
const
int
M
,
const
int
N
,
const
int
K
,
const
T
*
A
,
const
T
*
B
,
T
*
C
)
const
{
this
->
template
GEMM
<
T
>(
CblasRowMajor
,
CblasNoTrans
,
CblasNoTrans
,
M
,
N
,
K
,
static_cast
<
T
>
(
1
),
A
,
K
,
B
,
N
,
static_cast
<
T
>
(
0
),
C
,
N
);
}
template
<
>
template
<
typename
T
>
void
Blas
<
platform
::
CPUDeviceContext
>::
MatMul
(
const
int
M
,
const
int
N
,
const
int
K
,
const
T
*
A
,
const
T
*
B
,
T
*
C
)
const
{
#ifdef PADDLE_WITH_LIBXSMM
// Refer to https://github.com/hfp/libxsmm/blob/master/README.md
// But the threshold is custom constexpr int LIBXSMM_THRESHOLD = 20 * 20 * 20;
// Since the matrix is very small,
// so the unit of calculation is already very fast,
// and the if( M*N*K < LIBXSMM_THRESHOLD) would be overhead,
// use xsmm directly.
// Note: SMM use ColMajor
const
char
transa
=
'N'
;
const
char
transb
=
'N'
;
const
T
alpha
=
static_cast
<
T
>
(
1
);
const
T
beta
=
static_cast
<
T
>
(
0
);
CBlas
<
T
>::
SMM_GEMM
(
&
transa
,
&
transb
,
&
N
,
&
M
,
&
K
,
&
alpha
,
B
,
&
N
,
A
,
&
K
,
&
beta
,
C
,
&
N
);
return
;
#endif
CBlas
<
T
>::
GEMM
(
CblasRowMajor
,
CblasNoTrans
,
CblasNoTrans
,
M
,
N
,
K
,
static_cast
<
T
>
(
1
),
A
,
K
,
B
,
N
,
static_cast
<
T
>
(
0
),
C
,
N
);
}
template
<
typename
DeviceContext
>
template
<
typename
T
>
void
Blas
<
DeviceContext
>::
MatMul
(
const
framework
::
Tensor
&
mat_a
,
...
...
paddle/fluid/operators/math/fc_compute.h
浏览文件 @
f72ab896
...
...
@@ -25,17 +25,25 @@ namespace math {
template
<
typename
DeviceContext
,
typename
T
>
inline
void
FCCompute
(
const
BlasT
<
DeviceContext
,
T
>&
blas
,
const
int
M
,
const
int
N
,
const
int
K
,
const
T
*
X
,
const
T
*
W
,
T
*
Y
,
const
T
*
B
=
NULL
)
{
blas
.
GEMM
(
CblasNoTrans
,
CblasNoTrans
,
M
,
N
,
K
,
static_cast
<
T
>
(
1
),
X
,
W
,
static_cast
<
T
>
(
0
),
Y
);
if
(
B
)
{
const
T
*
B
=
NULL
,
bool
relu
=
false
)
{
blas
.
MatMul
(
M
,
N
,
K
,
X
,
W
,
Y
);
if
(
B
==
NULL
)
{
return
;
}
#ifdef PADDLE_WITH_MKLML
#pragma omp parallel for if (FLAGS_paddle_num_threads > 1)
#endif
for
(
int
i
=
0
;
i
<
M
;
i
++
)
{
blas
.
AXPY
(
N
,
static_cast
<
T
>
(
1
),
B
,
Y
+
i
*
N
);
}
for
(
int
i
=
0
;
i
<
M
;
i
++
)
{
blas
.
AXPY
(
N
,
static_cast
<
T
>
(
1
),
B
,
Y
+
i
*
N
);
}
if
(
!
relu
)
{
return
;
}
// TODO(TJ): fuse relu
LOG
(
FATAL
)
<<
"Not implemented!"
;
}
}
// namespace math
...
...
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