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
wmsofts
Paddle
提交
4dbdfa60
P
Paddle
项目概览
wmsofts
/
Paddle
与 Fork 源项目一致
Fork自
PaddlePaddle / Paddle
通知
1
Star
0
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
4dbdfa60
编写于
11月 16, 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
sigmoid and tanh support all size
test=develop
上级
ccb89637
变更
2
隐藏空白更改
内联
并排
Showing
2 changed file
with
54 addition
and
63 deletion
+54
-63
paddle/fluid/operators/math/jit_code.cc
paddle/fluid/operators/math/jit_code.cc
+10
-57
paddle/fluid/operators/math/jit_code.h
paddle/fluid/operators/math/jit_code.h
+44
-6
未找到文件。
paddle/fluid/operators/math/jit_code.cc
浏览文件 @
4dbdfa60
...
...
@@ -132,56 +132,8 @@ const int exp_int_0x7f[] ALIGN32 = {REPEAT_8TIMES(0x7f)};
int
g_tmp_mem
[
16
]
ALIGN32
=
{
0
};
bool
VActJitCode
::
init
(
int
d
,
operand_type
type
)
{
bool
ok
=
MayIUse
(
avx
);
if
(
type
==
operand_type
::
relu
||
type
==
operand_type
::
exp
)
{
// TODO(TJ): implement avx512, avx_exp is slower than mkl when d >= 256
return
ok
;
}
else
{
// TODO(TJ): support more
return
ok
&&
d
%
8
==
0
;
}
}
void
VActJitCode
::
sigmoid_ymm
(
ymm_t
&
ymm_dst
,
ymm_t
&
ymm_src
,
int
fx_idx
,
int
fy_idx
,
int
mask_idx
,
int
tmp_idx
)
{
// y = 1 / (1 + e^-x)
ymm_t
ymm_tmp
=
ymm_t
(
tmp_idx
);
reg64_t
reg_ptr_global
=
rax
;
push
(
reg_ptr_global
);
mov
(
reg_ptr_global
,
reinterpret_cast
<
size_t
>
(
exp_float_consts
));
vmovaps
(
ymm_tmp
,
ptr
[
reg_ptr_global
+
OFFSET_SIGMOID_MAX
]);
vminps
(
ymm_src
,
ymm_src
,
ymm_tmp
);
vmovaps
(
ymm_tmp
,
ptr
[
reg_ptr_global
+
OFFSET_SIGMOID_MIN
]);
vmaxps
(
ymm_src
,
ymm_src
,
ymm_tmp
);
vxorps
(
ymm_tmp
,
ymm_tmp
,
ymm_tmp
);
vsubps
(
ymm_src
,
ymm_tmp
,
ymm_src
);
exp_jmm
<
ymm_t
>
(
ymm_dst
,
ymm_src
,
fx_idx
,
fy_idx
,
mask_idx
,
tmp_idx
);
vmovaps
(
ymm_tmp
,
ptr
[
reg_ptr_global
+
OFFSET_EXP_ONE
]);
vaddps
(
ymm_dst
,
ymm_dst
,
ymm_tmp
);
vdivps
(
ymm_dst
,
ymm_tmp
,
ymm_dst
);
pop
(
reg_ptr_global
);
}
void
VActJitCode
::
tanh_ymm
(
ymm_t
&
ymm_dst
,
ymm_t
&
ymm_src
,
int
fx_idx
,
int
fy_idx
,
int
mask_idx
,
int
tmp_idx
)
{
// y = 2 / (1 + e^(-2x)) - 1
ymm_t
ymm_tmp
=
ymm_t
(
tmp_idx
);
ymm_t
ymm_zero
=
ymm_t
(
mask_idx
);
reg64_t
reg_ptr_global
=
rax
;
push
(
reg_ptr_global
);
mov
(
reg_ptr_global
,
reinterpret_cast
<
size_t
>
(
exp_float_consts
));
vmovaps
(
ymm_tmp
,
ptr
[
reg_ptr_global
+
OFFSET_EXP_TWO
]);
vxorps
(
ymm_zero
,
ymm_zero
,
ymm_zero
);
vsubps
(
ymm_tmp
,
ymm_zero
,
ymm_tmp
);
vmulps
(
ymm_src
,
ymm_src
,
ymm_tmp
);
exp_jmm
<
ymm_t
>
(
ymm_dst
,
ymm_src
,
fx_idx
,
fy_idx
,
mask_idx
,
tmp_idx
);
vmovaps
(
ymm_tmp
,
ptr
[
reg_ptr_global
+
OFFSET_EXP_ONE
]);
vaddps
(
ymm_dst
,
ymm_dst
,
ymm_tmp
);
vmovaps
(
ymm_tmp
,
ptr
[
reg_ptr_global
+
OFFSET_EXP_TWO
]);
vdivps
(
ymm_dst
,
ymm_tmp
,
ymm_dst
);
vmovaps
(
ymm_tmp
,
ptr
[
reg_ptr_global
+
OFFSET_EXP_ONE
]);
vsubps
(
ymm_dst
,
ymm_dst
,
ymm_tmp
);
pop
(
reg_ptr_global
);
// TODO(TJ): implement avx512, avx_exp is slower than mkl when d >= 256
return
MayIUse
(
avx
);
}
void
VActJitCode
::
generate
()
{
...
...
@@ -201,10 +153,10 @@ void VActJitCode::generate() {
exp_jmm
<
ymm_t
>
(
ymm_dst
,
ymm_src
,
2
,
3
,
4
,
5
);
break
;
case
operand_type
::
sigmoid
:
sigmoid_
ymm
(
ymm_dst
,
ymm_src
,
2
,
3
,
4
,
5
);
sigmoid_
jmm
<
ymm_t
>
(
ymm_dst
,
ymm_src
,
2
,
3
,
4
,
5
);
break
;
case
operand_type
::
tanh
:
tanh_
ymm
(
ymm_dst
,
ymm_src
,
2
,
3
,
4
,
5
);
tanh_
jmm
<
ymm_t
>
(
ymm_dst
,
ymm_src
,
2
,
3
,
4
,
5
);
break
;
case
operand_type
::
identity
:
break
;
...
...
@@ -214,11 +166,6 @@ void VActJitCode::generate() {
vmovups
(
ptr
[
param2
+
offset
],
ymm_dst
);
offset
+=
sizeof
(
float
)
*
YMM_FLOAT_BLOCK
;
}
if
(
type_
!=
operand_type
::
relu
&&
type_
!=
operand_type
::
exp
)
{
// TODO(TJ): remove me
ret
();
return
;
}
int
rest
=
num_
%
YMM_FLOAT_BLOCK
;
int
block
=
XMM_FLOAT_BLOCK
;
while
(
rest
>
0
)
{
...
...
@@ -236,6 +183,12 @@ void VActJitCode::generate() {
case
operand_type
::
exp
:
exp_jmm
<
xmm_t
>
(
xmm_dst
,
xmm_src
,
2
,
3
,
4
,
5
);
break
;
case
operand_type
::
sigmoid
:
sigmoid_jmm
<
xmm_t
>
(
xmm_dst
,
xmm_src
,
2
,
3
,
4
,
5
);
break
;
case
operand_type
::
tanh
:
tanh_jmm
<
xmm_t
>
(
xmm_dst
,
xmm_src
,
2
,
3
,
4
,
5
);
break
;
default:
break
;
}
...
...
paddle/fluid/operators/math/jit_code.h
浏览文件 @
4dbdfa60
...
...
@@ -263,13 +263,51 @@ class VActJitCode : public JitCode {
pop
(
reg_ptr_global
);
}
// compute sigmoid with ymm
void
sigmoid_ymm
(
const
Xbyak
::
Ymm
&
dst
,
const
Xbyak
::
Ymm
&
src
,
int
fx_idx
=
2
,
int
fy_idx
=
3
,
int
mask_idx
=
4
,
int
tmp_idx
=
5
);
// compute sigmoid with ymm, xmm
template
<
typename
JMM
>
void
sigmoid_jmm
(
JMM
&
dst
,
JMM
&
src
,
int
fx_idx
=
2
,
// NOLINT
int
fy_idx
=
3
,
int
mask_idx
=
4
,
int
tmp_idx
=
5
)
{
// y = 1 / (1 + e^-x)
JMM
jmm_tmp
=
JMM
(
tmp_idx
);
reg64_t
reg_ptr_global
=
rax
;
push
(
reg_ptr_global
);
mov
(
reg_ptr_global
,
reinterpret_cast
<
size_t
>
(
exp_float_consts
));
vmovaps
(
jmm_tmp
,
ptr
[
reg_ptr_global
+
OFFSET_SIGMOID_MAX
]);
vminps
(
src
,
src
,
jmm_tmp
);
vmovaps
(
jmm_tmp
,
ptr
[
reg_ptr_global
+
OFFSET_SIGMOID_MIN
]);
vmaxps
(
src
,
src
,
jmm_tmp
);
vxorps
(
jmm_tmp
,
jmm_tmp
,
jmm_tmp
);
vsubps
(
src
,
jmm_tmp
,
src
);
exp_jmm
<
JMM
>
(
dst
,
src
,
fx_idx
,
fy_idx
,
mask_idx
,
tmp_idx
);
vmovaps
(
jmm_tmp
,
ptr
[
reg_ptr_global
+
OFFSET_EXP_ONE
]);
vaddps
(
dst
,
dst
,
jmm_tmp
);
vdivps
(
dst
,
jmm_tmp
,
dst
);
pop
(
reg_ptr_global
);
}
// compute tanh with ymm
void
tanh_ymm
(
const
Xbyak
::
Ymm
&
dst
,
const
Xbyak
::
Ymm
&
src
,
int
fx_idx
=
2
,
int
fy_idx
=
3
,
int
mask_idx
=
4
,
int
tmp_idx
=
5
);
// compute tanh with ymm, xmm
template
<
typename
JMM
>
void
tanh_jmm
(
JMM
&
dst
,
JMM
&
src
,
int
fx_idx
=
2
,
int
fy_idx
=
3
,
// NOLINT
int
mask_idx
=
4
,
int
tmp_idx
=
5
)
{
// y = 2 / (1 + e^(-2x)) - 1
JMM
jmm_tmp
=
JMM
(
tmp_idx
);
JMM
jmm_zero
=
JMM
(
mask_idx
);
reg64_t
reg_ptr_global
=
rax
;
push
(
reg_ptr_global
);
mov
(
reg_ptr_global
,
reinterpret_cast
<
size_t
>
(
exp_float_consts
));
vmovaps
(
jmm_tmp
,
ptr
[
reg_ptr_global
+
OFFSET_EXP_TWO
]);
vxorps
(
jmm_zero
,
jmm_zero
,
jmm_zero
);
vsubps
(
jmm_tmp
,
jmm_zero
,
jmm_tmp
);
vmulps
(
src
,
src
,
jmm_tmp
);
exp_jmm
<
JMM
>
(
dst
,
src
,
fx_idx
,
fy_idx
,
mask_idx
,
tmp_idx
);
vmovaps
(
jmm_tmp
,
ptr
[
reg_ptr_global
+
OFFSET_EXP_ONE
]);
vaddps
(
dst
,
dst
,
jmm_tmp
);
vmovaps
(
jmm_tmp
,
ptr
[
reg_ptr_global
+
OFFSET_EXP_TWO
]);
vdivps
(
dst
,
jmm_tmp
,
dst
);
vmovaps
(
jmm_tmp
,
ptr
[
reg_ptr_global
+
OFFSET_EXP_ONE
]);
vsubps
(
dst
,
dst
,
jmm_tmp
);
pop
(
reg_ptr_global
);
}
protected:
int
num_
;
...
...
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