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
马赛克医院
Paddle
提交
266a5d2f
P
Paddle
项目概览
马赛克医院
/
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
0
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
0
Issue
0
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
266a5d2f
编写于
1月 12, 2019
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
implement matmul refer and mkl kernel
上级
c5623c87
变更
6
隐藏空白更改
内联
并排
Showing
6 changed file
with
47 addition
and
6 deletion
+47
-6
paddle/fluid/operators/jit/benchmark.cc
paddle/fluid/operators/jit/benchmark.cc
+1
-1
paddle/fluid/operators/jit/more/mkl/CMakeLists.txt
paddle/fluid/operators/jit/more/mkl/CMakeLists.txt
+1
-1
paddle/fluid/operators/jit/more/mkl/mkl.cc
paddle/fluid/operators/jit/more/mkl/mkl.cc
+21
-0
paddle/fluid/operators/jit/more/mkl/mkl.h
paddle/fluid/operators/jit/more/mkl/mkl.h
+6
-0
paddle/fluid/operators/jit/refer/refer.h
paddle/fluid/operators/jit/refer/refer.h
+14
-1
paddle/fluid/operators/jit/test.cc
paddle/fluid/operators/jit/test.cc
+4
-3
未找到文件。
paddle/fluid/operators/jit/benchmark.cc
浏览文件 @
266a5d2f
...
@@ -213,7 +213,7 @@ void BenchSeqPoolKernel() {
...
@@ -213,7 +213,7 @@ void BenchSeqPoolKernel() {
template
<
paddle
::
operators
::
jit
::
KernelType
KT
,
typename
T
,
typename
PlaceType
>
template
<
paddle
::
operators
::
jit
::
KernelType
KT
,
typename
T
,
typename
PlaceType
>
void
BenchMatMulKernel
()
{
void
BenchMatMulKernel
()
{
for
(
int
m
:
{
1
,
2
,
3
,
4
})
{
for
(
int
m
:
{
1
,
2
,
3
,
4
})
{
for
(
int
n
:
{
1
,
2
,
3
,
4
}
)
{
for
(
int
n
:
TestSizes
()
)
{
for
(
int
k
:
TestSizes
())
{
for
(
int
k
:
TestSizes
())
{
std
::
vector
<
T
>
a
(
m
*
k
),
b
(
k
*
n
),
c
(
m
*
n
);
std
::
vector
<
T
>
a
(
m
*
k
),
b
(
k
*
n
),
c
(
m
*
n
);
RandomVec
<
T
>
(
m
*
k
,
a
.
data
(),
-
2.
f
,
2.
f
);
RandomVec
<
T
>
(
m
*
k
,
a
.
data
(),
-
2.
f
,
2.
f
);
...
...
paddle/fluid/operators/jit/more/mkl/CMakeLists.txt
浏览文件 @
266a5d2f
...
@@ -3,7 +3,7 @@ cc_library(jit_kernel_mkl SRCS mkl.cc DEPS jit_kernel_base dynload_mklml)
...
@@ -3,7 +3,7 @@ cc_library(jit_kernel_mkl SRCS mkl.cc DEPS jit_kernel_base dynload_mklml)
set
(
JIT_KERNEL_DEPS
${
JIT_KERNEL_DEPS
}
dynload_mklml jit_kernel_mkl PARENT_SCOPE
)
set
(
JIT_KERNEL_DEPS
${
JIT_KERNEL_DEPS
}
dynload_mklml jit_kernel_mkl PARENT_SCOPE
)
# use mkl kernels by name and type
# use mkl kernels by name and type
#
USE_JITKERNEL_MORE(kMatMul, mkl)
USE_JITKERNEL_MORE
(
kMatMul, mkl
)
USE_JITKERNEL_MORE
(
kVMul, mkl
)
USE_JITKERNEL_MORE
(
kVMul, mkl
)
USE_JITKERNEL_MORE
(
kVAdd, mkl
)
USE_JITKERNEL_MORE
(
kVAdd, mkl
)
USE_JITKERNEL_MORE
(
kVScal, mkl
)
USE_JITKERNEL_MORE
(
kVScal, mkl
)
...
...
paddle/fluid/operators/jit/more/mkl/mkl.cc
浏览文件 @
266a5d2f
...
@@ -24,6 +24,20 @@ namespace jit {
...
@@ -24,6 +24,20 @@ namespace jit {
namespace
more
{
namespace
more
{
namespace
mkl
{
namespace
mkl
{
template
<
>
void
MatMul
<
float
>
(
const
float
*
a
,
const
float
*
b
,
float
*
c
,
int
m
,
int
n
,
int
k
)
{
platform
::
dynload
::
cblas_sgemm
(
CblasRowMajor
,
CblasNoTrans
,
CblasNoTrans
,
m
,
n
,
k
,
1.
f
,
a
,
k
,
b
,
n
,
0.
f
,
c
,
n
);
}
template
<
>
void
MatMul
<
double
>
(
const
double
*
a
,
const
double
*
b
,
double
*
c
,
int
m
,
int
n
,
int
k
)
{
platform
::
dynload
::
cblas_dgemm
(
CblasRowMajor
,
CblasNoTrans
,
CblasNoTrans
,
m
,
n
,
k
,
1.0
,
a
,
k
,
b
,
n
,
0.0
,
c
,
n
);
}
template
<
>
template
<
>
void
VMul
<
float
>
(
const
float
*
x
,
const
float
*
y
,
float
*
z
,
int
n
)
{
void
VMul
<
float
>
(
const
float
*
x
,
const
float
*
y
,
float
*
z
,
int
n
)
{
platform
::
dynload
::
vsMul
(
n
,
x
,
y
,
z
);
platform
::
dynload
::
vsMul
(
n
,
x
,
y
,
z
);
...
@@ -93,6 +107,11 @@ void VAXPY<double>(double a, const double* x, double* y, int n) {
...
@@ -93,6 +107,11 @@ void VAXPY<double>(double a, const double* x, double* y, int n) {
}
}
// TODO(TJ): tuning me carefully on AVX, AVX2 and AVX512
// TODO(TJ): tuning me carefully on AVX, AVX2 and AVX512
template
<
>
bool
MatMulKernel
<
float
>::
UseMe
(
const
int
&
d
)
const
{
return
platform
::
MayIUse
(
platform
::
avx
);
}
template
<
>
template
<
>
bool
VMulKernel
<
float
>::
UseMe
(
const
int
&
d
)
const
{
bool
VMulKernel
<
float
>::
UseMe
(
const
int
&
d
)
const
{
return
platform
::
MayIUse
(
platform
::
avx512f
)
&&
d
>
512
;
return
platform
::
MayIUse
(
platform
::
avx512f
)
&&
d
>
512
;
...
@@ -139,6 +158,7 @@ bool SeqPoolKernel<double>::UseMe(const seq_pool_attr_t& attr) const {
...
@@ -139,6 +158,7 @@ bool SeqPoolKernel<double>::UseMe(const seq_pool_attr_t& attr) const {
return true; \
return true; \
}
}
AWALYS_USE_ME_WITH_DOUBLE
(
MatMul
);
AWALYS_USE_ME_WITH_DOUBLE
(
VMul
);
AWALYS_USE_ME_WITH_DOUBLE
(
VMul
);
AWALYS_USE_ME_WITH_DOUBLE
(
VAdd
);
AWALYS_USE_ME_WITH_DOUBLE
(
VAdd
);
AWALYS_USE_ME_WITH_DOUBLE
(
VScal
);
AWALYS_USE_ME_WITH_DOUBLE
(
VScal
);
...
@@ -159,6 +179,7 @@ namespace mkl = paddle::operators::jit::more::mkl;
...
@@ -159,6 +179,7 @@ namespace mkl = paddle::operators::jit::more::mkl;
REGISTER_JITKERNEL_MORE(key, mkl, mkl::func##Kernel<float>, \
REGISTER_JITKERNEL_MORE(key, mkl, mkl::func##Kernel<float>, \
mkl::func##Kernel<double>)
mkl::func##Kernel<double>)
REGISTER_MKL_KERNEL
(
kMatMul
,
MatMul
);
REGISTER_MKL_KERNEL
(
kVMul
,
VMul
);
REGISTER_MKL_KERNEL
(
kVMul
,
VMul
);
REGISTER_MKL_KERNEL
(
kVAdd
,
VAdd
);
REGISTER_MKL_KERNEL
(
kVAdd
,
VAdd
);
REGISTER_MKL_KERNEL
(
kVScal
,
VScal
);
REGISTER_MKL_KERNEL
(
kVScal
,
VScal
);
...
...
paddle/fluid/operators/jit/more/mkl/mkl.h
浏览文件 @
266a5d2f
...
@@ -24,6 +24,9 @@ namespace jit {
...
@@ -24,6 +24,9 @@ namespace jit {
namespace
more
{
namespace
more
{
namespace
mkl
{
namespace
mkl
{
template
<
typename
T
>
void
MatMul
(
const
T
*
a
,
const
T
*
b
,
T
*
c
,
int
m
,
int
n
,
int
k
);
template
<
typename
T
>
template
<
typename
T
>
void
VMul
(
const
T
*
x
,
const
T
*
y
,
T
*
z
,
int
n
);
void
VMul
(
const
T
*
x
,
const
T
*
y
,
T
*
z
,
int
n
);
...
@@ -93,6 +96,9 @@ void SeqPool(const T* x, T* y, const seq_pool_attr_t* attr) {
...
@@ -93,6 +96,9 @@ void SeqPool(const T* x, T* y, const seq_pool_attr_t* attr) {
const char* ImplType() const override { return "MKL"; } \
const char* ImplType() const override { return "MKL"; } \
}
}
// ABCMNK
DECLARE_MKL_KERNEL
(
MatMul
,
MatMulTuples
);
// XYZN
// XYZN
DECLARE_MKL_KERNEL
(
VMul
,
XYZNTuples
);
DECLARE_MKL_KERNEL
(
VMul
,
XYZNTuples
);
DECLARE_MKL_KERNEL
(
VAdd
,
XYZNTuples
);
DECLARE_MKL_KERNEL
(
VAdd
,
XYZNTuples
);
...
...
paddle/fluid/operators/jit/refer/refer.h
浏览文件 @
266a5d2f
...
@@ -356,7 +356,20 @@ void SeqPool(const T* x, T* y, const seq_pool_attr_t* attr) {
...
@@ -356,7 +356,20 @@ void SeqPool(const T* x, T* y, const seq_pool_attr_t* attr) {
// A(M,K) * B(K,N) = C(M,N)
// A(M,K) * B(K,N) = C(M,N)
template
<
typename
T
>
template
<
typename
T
>
void
MatMul
(
const
T
*
A
,
const
T
*
B
,
T
*
C
,
int
M
,
int
N
,
int
K
)
{}
void
MatMul
(
const
T
*
A
,
const
T
*
B
,
T
*
C
,
int
M
,
int
N
,
int
K
)
{
for
(
int
m
=
0
;
m
<
M
;
++
m
)
{
const
T
*
pa
=
A
+
m
*
K
;
T
*
pc
=
C
+
m
*
N
;
for
(
int
n
=
0
;
n
<
N
;
++
n
)
{
const
T
*
pb
=
B
+
n
;
T
sum
=
static_cast
<
T
>
(
0
);
for
(
int
k
=
0
;
k
<
K
;
++
k
)
{
sum
+=
(
pa
[
k
]
*
pb
[
k
*
N
]);
}
*
(
pc
+
n
)
=
sum
;
}
}
}
#define DECLARE_REFER_KERNEL(name, tuples) \
#define DECLARE_REFER_KERNEL(name, tuples) \
template <typename T> \
template <typename T> \
...
...
paddle/fluid/operators/jit/test.cc
浏览文件 @
266a5d2f
...
@@ -230,7 +230,8 @@ struct TestFuncWithRefer<jit::SeqPoolTuples<T>, std::vector<T>,
...
@@ -230,7 +230,8 @@ struct TestFuncWithRefer<jit::SeqPoolTuples<T>, std::vector<T>,
};
};
template
<
typename
T
>
template
<
typename
T
>
struct
TestFuncWithRefer
<
jit
::
MatMulTuples
<
T
>
,
std
::
vector
<
T
>
,
std
::
vector
<
T
>>
{
struct
TestFuncWithRefer
<
jit
::
MatMulTuples
<
T
>
,
std
::
vector
<
T
>
,
std
::
vector
<
T
>
,
std
::
vector
<
T
>
,
int
,
int
,
int
>
{
void
operator
()(
const
typename
jit
::
MatMulTuples
<
T
>::
func_type
tgt
,
void
operator
()(
const
typename
jit
::
MatMulTuples
<
T
>::
func_type
tgt
,
const
std
::
vector
<
T
>&
a
,
const
std
::
vector
<
T
>&
b
,
const
std
::
vector
<
T
>&
a
,
const
std
::
vector
<
T
>&
b
,
const
std
::
vector
<
T
>&
cref
,
int
m
,
int
n
,
int
k
)
{
const
std
::
vector
<
T
>&
cref
,
int
m
,
int
n
,
int
k
)
{
...
@@ -486,8 +487,8 @@ void TestMatMulKernel() {
...
@@ -486,8 +487,8 @@ void TestMatMulKernel() {
auto
ref
=
jit
::
GetRefer
<
KT
,
jit
::
MatMulTuples
<
T
>>
();
auto
ref
=
jit
::
GetRefer
<
KT
,
jit
::
MatMulTuples
<
T
>>
();
EXPECT_TRUE
(
ref
!=
nullptr
);
EXPECT_TRUE
(
ref
!=
nullptr
);
std
::
vector
<
T
>
a
(
m
*
k
),
b
(
k
*
n
),
c
(
m
*
n
);
std
::
vector
<
T
>
a
(
m
*
k
),
b
(
k
*
n
),
c
(
m
*
n
);
RandomVec
<
T
>
(
m
*
k
,
a
.
data
(),
-
2.
f
,
2.
f
);
RandomVec
<
T
>
(
m
*
k
,
a
.
data
(),
-
0.2
f
,
0.2
f
);
RandomVec
<
T
>
(
k
*
n
,
b
.
data
(),
-
2.
f
,
2.
f
);
RandomVec
<
T
>
(
k
*
n
,
b
.
data
(),
-
0.2
f
,
0.2
f
);
const
T
*
a_data
=
a
.
data
();
const
T
*
a_data
=
a
.
data
();
const
T
*
b_data
=
b
.
data
();
const
T
*
b_data
=
b
.
data
();
T
*
c_data
=
c
.
data
();
T
*
c_data
=
c
.
data
();
...
...
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