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
提交
b6260f38
P
Paddle
项目概览
PaddlePaddle
/
Paddle
大约 1 年 前同步成功
通知
2296
Star
20931
Fork
5422
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
1423
列表
看板
标记
里程碑
合并请求
543
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
1,423
Issue
1,423
列表
看板
标记
里程碑
合并请求
543
合并请求
543
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
b6260f38
编写于
10月 31, 2019
作者:
Y
Yihua Xu
提交者:
Tao Luo
10月 31, 2019
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Optimize the kernel implementation of layernorm with openmp (#20895)
上级
8c4573a3
变更
1
隐藏空白更改
内联
并排
Showing
1 changed file
with
114 addition
and
104 deletion
+114
-104
paddle/fluid/operators/jit/more/intrinsic/layer_norm.cc
paddle/fluid/operators/jit/more/intrinsic/layer_norm.cc
+114
-104
未找到文件。
paddle/fluid/operators/jit/more/intrinsic/layer_norm.cc
浏览文件 @
b6260f38
...
@@ -26,131 +26,141 @@ namespace intrinsic {
...
@@ -26,131 +26,141 @@ namespace intrinsic {
void
LayerNorm
(
float
*
x
,
float
*
out
,
float
*
mean
,
float
*
var
,
void
LayerNorm
(
float
*
x
,
float
*
out
,
float
*
mean
,
float
*
var
,
const
float
*
scale
,
const
float
*
bias
,
int
height
,
const
float
*
scale
,
const
float
*
bias
,
int
height
,
const
float
epsilon
,
int
right
)
{
const
float
epsilon
,
int
right
)
{
__m256
sum
;
__m256
mean_vec
,
var_vec
;
__m128
hi
,
lo
;
__m256
tmp
;
size_t
offset
;
size_t
j
;
int
block
=
YMM_FLOAT_BLOCK
;
int
block
=
YMM_FLOAT_BLOCK
;
const
int
rest
=
right
%
block
;
const
int
rest
=
right
%
block
;
const
int
end
=
right
-
rest
;
const
int
end
=
right
-
rest
;
#ifdef PADDLE_WITH_MKLML
#pragma omp parallel
{
#endif
__m256
sum
;
__m256
mean_vec
,
var_vec
;
__m128
hi
,
lo
;
__m256
tmp
;
size_t
offset
;
size_t
j
;
__m256
reverse_num_vec
=
_mm256_div_ps
(
_mm256_set1_ps
(
1.0
),
_mm256_set1_ps
(
right
));
__m256
epsilon_vec
=
_mm256_set1_ps
(
epsilon
);
int
rest_mask
=
((
-
1
)
&
(
~
((
~
0U
)
>>
(
sizeof
(
int
)
*
8
-
(
block
-
rest
)))))
&
0x0ff
;
__m256i
mask_vec
=
_mm256_set_epi32
(
rest_mask
&
0x80
?
0xffffffff
:
0
,
rest_mask
&
0x40
?
0xffffffff
:
0
,
rest_mask
&
0x20
?
0xffffffff
:
0
,
rest_mask
&
0x10
?
0xffffffff
:
0
,
rest_mask
&
0x8
?
0xffffffff
:
0
,
rest_mask
&
0x4
?
0xffffffff
:
0
,
rest_mask
&
0x2
?
0xffffffff
:
0
,
rest_mask
&
0x1
?
0xffffffff
:
0
);
__m256
reverse_num_vec
=
#ifdef PADDLE_WITH_MKLML
_mm256_div_ps
(
_mm256_set1_ps
(
1.0
),
_mm256_set1_ps
(
right
));
#pragma omp for
__m256
epsilon_vec
=
_mm256_set1_ps
(
epsilon
);
#endif
int
rest_mask
=
for
(
int
i
=
0
;
i
<
height
;
++
i
)
{
((
-
1
)
&
(
~
((
~
0U
)
>>
(
sizeof
(
int
)
*
8
-
(
block
-
rest
)))))
&
0x0ff
;
offset
=
i
*
right
;
__m256i
mask_vec
=
_mm256_set_epi32
(
rest_mask
&
0x80
?
0xffffffff
:
0
,
rest_mask
&
0x40
?
0xffffffff
:
0
,
rest_mask
&
0x20
?
0xffffffff
:
0
,
rest_mask
&
0x10
?
0xffffffff
:
0
,
rest_mask
&
0x8
?
0xffffffff
:
0
,
rest_mask
&
0x4
?
0xffffffff
:
0
,
rest_mask
&
0x2
?
0xffffffff
:
0
,
rest_mask
&
0x1
?
0xffffffff
:
0
);
for
(
int
i
=
0
;
i
<
height
;
++
i
)
{
/* get mean */
offset
=
i
*
right
;
sum
=
_mm256_setzero_ps
();
/* get mean */
sum
=
_mm256_setzero_ps
();
for
(
j
=
offset
;
j
<
end
+
offset
;
j
+=
block
)
{
sum
=
_mm256_add_ps
(
sum
,
_mm256_loadu_ps
((
const
float
*
)
x
+
j
));
}
if
(
rest
!=
0
)
{
j
=
offset
+
right
-
block
;
tmp
=
_mm256_loadu_ps
((
const
float
*
)
x
+
j
);
tmp
=
_mm256_blendv_ps
(
_mm256_setzero_ps
(),
tmp
,
*
(
__m256
*
)
&
mask_vec
);
// NOLINT
sum
=
_mm256_add_ps
(
sum
,
tmp
);
}
hi
=
_mm256_extractf128_ps
(
sum
,
1
);
lo
=
_mm256_extractf128_ps
(
sum
,
0
);
sum
=
_mm256_add_ps
(
sum
,
_mm256_insertf128_ps
(
_mm256_insertf128_ps
(
_mm256_setzero_ps
(),
hi
,
0
),
lo
,
1
));
sum
=
_mm256_hadd_ps
(
sum
,
sum
);
sum
=
_mm256_hadd_ps
(
sum
,
sum
);
mean_vec
=
_mm256_mul_ps
(
sum
,
reverse_num_vec
);
mean
[
i
]
=
*
reinterpret_cast
<
float
*>
(
&
mean_vec
);
/* get variance */
sum
=
_mm256_setzero_ps
();
for
(
j
=
offset
;
j
<
end
+
offset
;
j
+=
block
)
{
tmp
=
_mm256_sub_ps
(
_mm256_loadu_ps
((
const
float
*
)
x
+
j
),
mean_vec
);
tmp
=
_mm256_mul_ps
(
tmp
,
tmp
);
sum
=
_mm256_add_ps
(
sum
,
tmp
);
}
if
(
rest
!=
0
)
{
j
=
offset
+
right
-
block
;
tmp
=
_mm256_sub_ps
(
_mm256_loadu_ps
((
const
float
*
)
x
+
j
),
mean_vec
);
tmp
=
_mm256_mul_ps
(
tmp
,
tmp
);
tmp
=
_mm256_blendv_ps
(
_mm256_setzero_ps
(),
tmp
,
*
(
__m256
*
)
&
mask_vec
);
// NOLINT
sum
=
_mm256_add_ps
(
sum
,
tmp
);
}
hi
=
_mm256_extractf128_ps
(
sum
,
1
);
lo
=
_mm256_extractf128_ps
(
sum
,
0
);
sum
=
_mm256_add_ps
(
sum
,
_mm256_insertf128_ps
(
_mm256_insertf128_ps
(
_mm256_setzero_ps
(),
hi
,
0
),
lo
,
1
));
sum
=
_mm256_hadd_ps
(
sum
,
sum
);
sum
=
_mm256_hadd_ps
(
sum
,
sum
);
var_vec
=
_mm256_mul_ps
(
sum
,
reverse_num_vec
);
var
[
i
]
=
*
reinterpret_cast
<
float
*>
(
&
var_vec
);
/* get x_norm and calculate output*/
for
(
j
=
offset
;
j
<
end
+
offset
;
j
+=
block
)
{
tmp
=
_mm256_sub_ps
(
_mm256_loadu_ps
((
const
float
*
)
x
+
j
),
mean_vec
);
tmp
=
_mm256_div_ps
(
tmp
,
_mm256_sqrt_ps
(
_mm256_add_ps
(
var_vec
,
epsilon_vec
)));
_mm256_storeu_ps
(
reinterpret_cast
<
float
*>
(
out
)
+
j
,
tmp
);
}
if
(
rest
!=
0
)
{
j
=
offset
+
right
-
block
;
tmp
=
_mm256_sub_ps
(
_mm256_loadu_ps
((
const
float
*
)
x
+
j
),
mean_vec
);
tmp
=
_mm256_div_ps
(
tmp
,
_mm256_sqrt_ps
(
_mm256_add_ps
(
var_vec
,
epsilon_vec
)));
_mm256_storeu_ps
(
reinterpret_cast
<
float
*>
(
out
)
+
j
,
tmp
);
}
if
(
scale
)
{
if
(
rest
!=
0
)
{
j
=
offset
+
right
-
block
;
tmp
=
_mm256_loadu_ps
((
const
float
*
)
out
+
j
);
}
for
(
j
=
offset
;
j
<
end
+
offset
;
j
+=
block
)
{
for
(
j
=
offset
;
j
<
end
+
offset
;
j
+=
block
)
{
_mm256_storeu_ps
(
sum
=
_mm256_add_ps
(
sum
,
_mm256_loadu_ps
((
const
float
*
)
x
+
j
));
reinterpret_cast
<
float
*>
(
out
)
+
j
,
_mm256_mul_ps
(
_mm256_loadu_ps
((
const
float
*
)
out
+
j
),
_mm256_loadu_ps
((
const
float
*
)
scale
+
j
-
offset
)));
}
}
if
(
rest
!=
0
)
{
if
(
rest
!=
0
)
{
j
=
offset
+
right
-
block
;
j
=
offset
+
right
-
block
;
_mm256_storeu_ps
(
tmp
=
_mm256_loadu_ps
((
const
float
*
)
x
+
j
);
reinterpret_cast
<
float
*>
(
out
)
+
j
,
tmp
=
_mm256_blendv_ps
(
_mm256_setzero_ps
(),
tmp
,
_mm256_mul_ps
(
tmp
,
*
(
__m256
*
)
&
mask_vec
);
// NOLINT
_mm256_loadu_ps
((
const
float
*
)
scale
+
j
-
offset
))
);
sum
=
_mm256_add_ps
(
sum
,
tmp
);
}
}
}
hi
=
_mm256_extractf128_ps
(
sum
,
1
);
lo
=
_mm256_extractf128_ps
(
sum
,
0
);
sum
=
_mm256_add_ps
(
sum
,
_mm256_insertf128_ps
(
_mm256_insertf128_ps
(
_mm256_setzero_ps
(),
hi
,
0
),
lo
,
1
));
sum
=
_mm256_hadd_ps
(
sum
,
sum
);
sum
=
_mm256_hadd_ps
(
sum
,
sum
);
mean_vec
=
_mm256_mul_ps
(
sum
,
reverse_num_vec
);
mean
[
i
]
=
*
reinterpret_cast
<
float
*>
(
&
mean_vec
);
if
(
bias
)
{
/* get variance */
sum
=
_mm256_setzero_ps
();
for
(
j
=
offset
;
j
<
end
+
offset
;
j
+=
block
)
{
tmp
=
_mm256_sub_ps
(
_mm256_loadu_ps
((
const
float
*
)
x
+
j
),
mean_vec
);
tmp
=
_mm256_mul_ps
(
tmp
,
tmp
);
sum
=
_mm256_add_ps
(
sum
,
tmp
);
}
if
(
rest
!=
0
)
{
if
(
rest
!=
0
)
{
j
=
offset
+
right
-
block
;
j
=
offset
+
right
-
block
;
tmp
=
_mm256_loadu_ps
((
const
float
*
)
out
+
j
);
tmp
=
_mm256_sub_ps
(
_mm256_loadu_ps
((
const
float
*
)
x
+
j
),
mean_vec
);
tmp
=
_mm256_mul_ps
(
tmp
,
tmp
);
tmp
=
_mm256_blendv_ps
(
_mm256_setzero_ps
(),
tmp
,
*
(
__m256
*
)
&
mask_vec
);
// NOLINT
sum
=
_mm256_add_ps
(
sum
,
tmp
);
}
}
hi
=
_mm256_extractf128_ps
(
sum
,
1
);
lo
=
_mm256_extractf128_ps
(
sum
,
0
);
sum
=
_mm256_add_ps
(
sum
,
_mm256_insertf128_ps
(
_mm256_insertf128_ps
(
_mm256_setzero_ps
(),
hi
,
0
),
lo
,
1
));
sum
=
_mm256_hadd_ps
(
sum
,
sum
);
sum
=
_mm256_hadd_ps
(
sum
,
sum
);
var_vec
=
_mm256_mul_ps
(
sum
,
reverse_num_vec
);
var
[
i
]
=
*
reinterpret_cast
<
float
*>
(
&
var_vec
);
/* get x_norm and calculate output*/
for
(
j
=
offset
;
j
<
end
+
offset
;
j
+=
block
)
{
for
(
j
=
offset
;
j
<
end
+
offset
;
j
+=
block
)
{
_mm256_storeu_ps
(
tmp
=
_mm256_sub_ps
(
_mm256_loadu_ps
((
const
float
*
)
x
+
j
),
mean_vec
);
reinterpret_cast
<
float
*>
(
out
)
+
j
,
tmp
=
_mm256_div_ps
(
_mm256_add_ps
(
_mm256_loadu_ps
((
const
float
*
)
out
+
j
),
tmp
,
_mm256_sqrt_ps
(
_mm256_add_ps
(
var_vec
,
epsilon_vec
)));
_mm256_loadu_ps
((
const
float
*
)
bias
+
j
-
offset
))
);
_mm256_storeu_ps
(
reinterpret_cast
<
float
*>
(
out
)
+
j
,
tmp
);
}
}
if
(
rest
!=
0
)
{
if
(
rest
!=
0
)
{
j
=
offset
+
right
-
block
;
j
=
offset
+
right
-
block
;
_mm256_storeu_ps
(
reinterpret_cast
<
float
*>
(
out
)
+
j
,
tmp
=
_mm256_sub_ps
(
_mm256_loadu_ps
((
const
float
*
)
x
+
j
),
mean_vec
);
_mm256_add_ps
(
tmp
,
_mm256_loadu_ps
((
const
float
*
)
bias
+
tmp
=
_mm256_div_ps
(
j
-
offset
)));
tmp
,
_mm256_sqrt_ps
(
_mm256_add_ps
(
var_vec
,
epsilon_vec
)));
_mm256_storeu_ps
(
reinterpret_cast
<
float
*>
(
out
)
+
j
,
tmp
);
}
if
(
scale
)
{
if
(
rest
!=
0
)
{
j
=
offset
+
right
-
block
;
tmp
=
_mm256_loadu_ps
((
const
float
*
)
out
+
j
);
}
for
(
j
=
offset
;
j
<
end
+
offset
;
j
+=
block
)
{
_mm256_storeu_ps
(
reinterpret_cast
<
float
*>
(
out
)
+
j
,
_mm256_mul_ps
(
_mm256_loadu_ps
((
const
float
*
)
out
+
j
),
_mm256_loadu_ps
((
const
float
*
)
scale
+
j
-
offset
)));
}
if
(
rest
!=
0
)
{
j
=
offset
+
right
-
block
;
_mm256_storeu_ps
(
reinterpret_cast
<
float
*>
(
out
)
+
j
,
_mm256_mul_ps
(
tmp
,
_mm256_loadu_ps
((
const
float
*
)
scale
+
j
-
offset
)));
}
}
if
(
bias
)
{
if
(
rest
!=
0
)
{
j
=
offset
+
right
-
block
;
tmp
=
_mm256_loadu_ps
((
const
float
*
)
out
+
j
);
}
for
(
j
=
offset
;
j
<
end
+
offset
;
j
+=
block
)
{
_mm256_storeu_ps
(
reinterpret_cast
<
float
*>
(
out
)
+
j
,
_mm256_add_ps
(
_mm256_loadu_ps
((
const
float
*
)
out
+
j
),
_mm256_loadu_ps
((
const
float
*
)
bias
+
j
-
offset
)));
}
if
(
rest
!=
0
)
{
j
=
offset
+
right
-
block
;
_mm256_storeu_ps
(
reinterpret_cast
<
float
*>
(
out
)
+
j
,
_mm256_add_ps
(
tmp
,
_mm256_loadu_ps
((
const
float
*
)
bias
+
j
-
offset
)));
}
}
}
}
}
#ifdef PADDLE_WITH_MKLML
}
}
#endif
}
}
bool
LayerNormKernel
::
CanBeUsed
(
const
int
&
d
)
const
{
bool
LayerNormKernel
::
CanBeUsed
(
const
int
&
d
)
const
{
...
...
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