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
sven357
Paddle
提交
a9c18241
P
Paddle
项目概览
sven357
/
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
a9c18241
编写于
11月 05, 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 jit vmul code supporting multiple of 2
上级
61fdc38e
变更
3
隐藏空白更改
内联
并排
Showing
3 changed file
with
34 addition
and
15 deletion
+34
-15
paddle/fluid/operators/math/jit_code.cc
paddle/fluid/operators/math/jit_code.cc
+29
-8
paddle/fluid/operators/math/jit_code.h
paddle/fluid/operators/math/jit_code.h
+4
-6
paddle/fluid/operators/math/jit_kernel_test.cc
paddle/fluid/operators/math/jit_kernel_test.cc
+1
-1
未找到文件。
paddle/fluid/operators/math/jit_code.cc
浏览文件 @
a9c18241
...
...
@@ -25,10 +25,10 @@ namespace gen {
using
namespace
platform
::
jit
;
// NOLINT
bool
VMulJitCode
::
init
(
int
d
)
{
//
TODO(TJ): maybe one AVX is enough, AVX above would slow down freq
//
try more with avx2 or avx512
if
(
MayIUse
(
avx
)
||
MayIUse
(
avx2
)
)
{
return
d
%
AVX_FLOAT_BLOCK
==
0
;
//
It's not necessary to use avx512 since it would slow down the frequency
//
and this kernel is not compute bound.
if
(
MayIUse
(
avx
))
{
return
d
%
2
==
0
;
}
else
{
return
false
;
}
...
...
@@ -36,12 +36,33 @@ bool VMulJitCode::init(int d) {
void
VMulJitCode
::
generate
()
{
// do not need push stack, and do not need save avx512reg if do not use avx512
int
stride
=
sizeof
(
float
)
*
AVX_FLOAT_BLOCK
;
int
offset
=
0
;
for
(
int
i
=
0
;
i
<
num_
/
AVX_FLOAT_BLOCK
;
++
i
)
{
vmovups
(
ymm_src1
,
ptr
[
param1
+
i
*
stride
]);
vmovups
(
ymm_src2
,
ptr
[
param2
+
i
*
stride
]);
vmovups
(
ymm_src1
,
ptr
[
param1
+
offset
]);
vmovups
(
ymm_src2
,
ptr
[
param2
+
offset
]);
vmulps
(
ymm_dst
,
ymm_src1
,
ymm_src2
);
vmovups
(
ptr
[
param3
+
stride
*
i
],
ymm_dst
);
vmovups
(
ptr
[
param3
+
offset
],
ymm_dst
);
offset
+=
sizeof
(
float
)
*
AVX_FLOAT_BLOCK
;
}
int
rest
=
num_
%
AVX_FLOAT_BLOCK
;
if
(
rest
>=
4
)
{
vmovups
(
xmm_src1
,
ptr
[
param1
+
offset
]);
vmovups
(
xmm_src2
,
ptr
[
param2
+
offset
]);
vmulps
(
xmm_dst
,
xmm_src1
,
xmm_src2
);
vmovups
(
ptr
[
param3
+
offset
],
xmm_dst
);
offset
+=
sizeof
(
float
)
*
4
;
rest
-=
4
;
}
if
(
rest
>=
2
)
{
mov
(
tmp
,
qword
[
param1
+
offset
]);
vmovq
(
xmm_src1
,
tmp
);
mov
(
tmp
,
qword
[
param2
+
offset
]);
vmovq
(
xmm_src2
,
tmp
);
vmulps
(
xmm_dst
,
xmm_src1
,
xmm_src2
);
vmovq
(
tmp
,
xmm_dst
);
mov
(
ptr
[
param3
+
offset
],
tmp
);
offset
+=
sizeof
(
float
)
*
2
;
rest
-=
2
;
}
ret
();
}
...
...
paddle/fluid/operators/math/jit_code.h
浏览文件 @
a9c18241
...
...
@@ -43,17 +43,15 @@ class VMulJitCode : public JitCode {
reg64_t
param1
{
abi_param1
};
reg64_t
param2
{
abi_param2
};
reg64_t
param3
{
abi_param3
};
reg64_t
tmp
=
rax
;
xmm_t
xmm_src1
=
xmm_t
(
0
);
ymm_t
ymm_src1
=
ymm_t
(
0
);
zmm_t
zmm_src1
=
zmm_t
(
0
);
xmm_t
xmm_src2
=
xmm_t
(
1
);
ymm_t
ymm_src2
=
ymm_t
(
1
);
zmm_t
zmm_src2
=
zmm_t
(
1
);
xmm_t
xmm_dst
=
xmm_t
(
2
);
ymm_t
ymm_src1
=
ymm_t
(
0
);
ymm_t
ymm_src2
=
ymm_t
(
1
);
ymm_t
ymm_dst
=
ymm_t
(
2
);
zmm_t
zmm_dst
=
zmm_t
(
2
);
};
}
// namespace gen
...
...
paddle/fluid/operators/math/jit_kernel_test.cc
浏览文件 @
a9c18241
...
...
@@ -578,7 +578,7 @@ void vmul_mkl(const int n, const float* x, const float* y, float* z) {
TEST
(
JitKernel
,
vmul
)
{
namespace
jit
=
paddle
::
operators
::
math
::
jitkernel
;
for
(
int
d
:
{
7
,
8
,
15
,
16
,
30
,
256
,
512
,
1000
,
1024
})
{
for
(
int
d
:
{
7
,
8
,
15
,
16
,
20
,
30
,
256
,
512
,
1000
,
1024
})
{
std
::
vector
<
float
>
x
(
d
),
y
(
d
);
std
::
vector
<
float
>
zref
(
d
),
ztgt
(
d
);
RandomVec
<
float
>
(
d
,
x
.
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