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
机器未来
Paddle
提交
2dec25db
P
Paddle
项目概览
机器未来
/
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
未验证
提交
2dec25db
编写于
3月 16, 2022
作者:
Z
Zhang Zheng
提交者:
GitHub
3月 16, 2022
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Optimize the computation of log_softmax (#40612)
* Optimize the computation of log_softmax * modify the var name
上级
a09a93a1
变更
1
隐藏空白更改
内联
并排
Showing
1 changed file
with
20 addition
and
23 deletion
+20
-23
paddle/phi/kernels/gpudnn/softmax_gpudnn.h
paddle/phi/kernels/gpudnn/softmax_gpudnn.h
+20
-23
未找到文件。
paddle/phi/kernels/gpudnn/softmax_gpudnn.h
浏览文件 @
2dec25db
...
...
@@ -121,17 +121,10 @@ struct ReduceMaxFunctor {
};
template
<
typename
Tx
,
typename
Ty
=
Tx
>
struct
ExpSubFunctor
{
HOSTDEVICE
inline
ExpSubFunctor
()
{
y
=
static_cast
<
Tx
>
(
0.0
f
);
}
HOSTDEVICE
explicit
inline
ExpSubFunctor
(
Tx
y
)
:
y
((
Tx
)(
y
))
{}
struct
ExpFunctor
{
HOSTDEVICE
inline
Ty
operator
()(
const
Tx
&
x
)
const
{
return
static_cast
<
Ty
>
(
std
::
exp
(
x
-
y
));
return
static_cast
<
Ty
>
(
std
::
exp
(
x
));
}
private:
Tx
y
;
};
template
<
typename
Tx
,
typename
Ty
=
Tx
>
...
...
@@ -293,10 +286,14 @@ __global__ void WarpSoftmaxForward(T* softmax,
}
// data src
AccT
srcdata
[
kBatchSize
][
kLoopsV
][
kVSize
];
T
src_tmp
[
kBatchSize
][
kLoopsV
][
kVSize
];
kps
::
Init
<
AccT
,
kStep
>
(
&
srcdata
[
0
][
0
][
0
],
kLowInf
);
kps
::
Init
<
T
,
kStep
>
(
&
src_tmp
[
0
][
0
][
0
],
-
std
::
numeric_limits
<
T
>::
infinity
());
// src_data: the raw data form global memory
// sub_data: store the data obtained by (src_data - max), used by log_softmax
// exp_data: store the data obtained by (exp(sub_data)), used by softmax
T
src_data
[
kBatchSize
][
kLoopsV
][
kVSize
];
AccT
sub_data
[
kBatchSize
][
kLoopsV
][
kVSize
];
AccT
exp_data
[
kBatchSize
][
kLoopsV
][
kVSize
];
kps
::
Init
<
AccT
,
kStep
>
(
&
sub_data
[
0
][
0
][
0
],
kLowInf
);
kps
::
Init
<
T
,
kStep
>
(
&
src_data
[
0
][
0
][
0
],
-
std
::
numeric_limits
<
T
>::
infinity
());
// data dst
T
out_tmp
[
kBatchSize
][
kLoopsV
][
kVSize
];
...
...
@@ -313,11 +310,11 @@ __global__ void WarpSoftmaxForward(T* softmax,
for
(
int
i
=
0
;
i
<
kBatchSize
;
++
i
)
{
const
VecT
*
src_v
=
reinterpret_cast
<
const
VecT
*>
(
&
src
[(
first_batch
+
i
)
*
stride
]);
VecT
*
reg_v
=
reinterpret_cast
<
VecT
*>
(
&
src_
tmp
[
i
][
0
][
0
]);
VecT
*
reg_v
=
reinterpret_cast
<
VecT
*>
(
&
src_
data
[
i
][
0
][
0
]);
kps
::
ReadData
<
VecT
,
VecT
,
kLoopsV
,
1
,
1
,
true
>
(
&
reg_v
[
0
],
&
src_v
[
0
],
idx_max_v
[
i
],
0
,
kWarpSize
,
1
);
kps
::
ElementwiseUnary
<
T
,
AccT
,
kVItem
,
1
,
1
,
DataTransFunctor
<
T
,
AccT
>>
(
&
s
rcdata
[
i
][
0
][
0
],
&
src_tmp
[
i
][
0
][
0
],
DataTransFunctor
<
T
,
AccT
>
());
&
s
ub_data
[
i
][
0
][
0
],
&
src_data
[
i
][
0
][
0
],
DataTransFunctor
<
T
,
AccT
>
());
}
// compute max
...
...
@@ -327,14 +324,16 @@ __global__ void WarpSoftmaxForward(T* softmax,
1
,
ReduceMaxFunctor
<
AccT
>
,
kMode
::
kLocalMode
>
(
&
max
[
0
],
&
s
rc
data
[
0
][
0
][
0
],
ReduceMaxFunctor
<
AccT
>
(),
true
);
&
max
[
0
],
&
s
ub_
data
[
0
][
0
][
0
],
ReduceMaxFunctor
<
AccT
>
(),
true
);
WarpReduceMax
<
AccT
,
kBatchSize
,
kWarpSize
>
(
max
);
// compute sum
#pragma unroll
for
(
int
i
=
0
;
i
<
kBatchSize
;
++
i
)
{
kps
::
ElementwiseUnary
<
AccT
,
AccT
,
kVItem
,
1
,
1
,
ExpSubFunctor
<
AccT
>>
(
&
srcdata
[
i
][
0
][
0
],
&
srcdata
[
i
][
0
][
0
],
ExpSubFunctor
<
AccT
>
(
max
[
i
]));
kps
::
ElementwiseUnary
<
AccT
,
AccT
,
kVItem
,
1
,
1
,
UnarySubFunctor
<
AccT
>>
(
&
sub_data
[
i
][
0
][
0
],
&
sub_data
[
i
][
0
][
0
],
UnarySubFunctor
<
AccT
>
(
max
[
i
]));
kps
::
ElementwiseUnary
<
AccT
,
AccT
,
kVItem
,
1
,
1
,
ExpFunctor
<
AccT
>>
(
&
exp_data
[
i
][
0
][
0
],
&
sub_data
[
i
][
0
][
0
],
ExpFunctor
<
AccT
>
());
}
kps
::
Reduce
<
AccT
,
kVItem
,
...
...
@@ -342,7 +341,7 @@ __global__ void WarpSoftmaxForward(T* softmax,
1
,
kps
::
AddFunctor
<
AccT
>
,
kMode
::
kLocalMode
>
(
&
sum
[
0
],
&
src
data
[
0
][
0
][
0
],
kps
::
AddFunctor
<
AccT
>
(),
true
);
&
sum
[
0
],
&
exp_
data
[
0
][
0
][
0
],
kps
::
AddFunctor
<
AccT
>
(),
true
);
WarpReduceSum
<
AccT
,
kBatchSize
,
kWarpSize
>
(
sum
);
// write data to global memory
...
...
@@ -352,15 +351,13 @@ __global__ void WarpSoftmaxForward(T* softmax,
reinterpret_cast
<
VecT
*>
(
&
softmax
[(
first_batch
+
i
)
*
stride
]);
VecT
*
reg_v
=
reinterpret_cast
<
VecT
*>
(
&
out_tmp
[
i
][
0
][
0
]);
if
(
LogMode
)
{
kps
::
ElementwiseUnary
<
AccT
,
AccT
,
kVItem
,
1
,
1
,
UnaryLogFunctor
<
AccT
>>
(
&
srcdata
[
i
][
0
][
0
],
&
srcdata
[
i
][
0
][
0
],
UnaryLogFunctor
<
AccT
>
());
kps
::
ElementwiseUnary
<
AccT
,
T
,
kVItem
,
1
,
1
,
UnarySubFunctor
<
AccT
>>
(
&
out_tmp
[
i
][
0
][
0
],
&
s
rc
data
[
i
][
0
][
0
],
&
s
ub_
data
[
i
][
0
][
0
],
UnarySubFunctor
<
AccT
>
(
std
::
log
(
sum
[
i
])));
}
else
{
kps
::
ElementwiseUnary
<
AccT
,
T
,
kVItem
,
1
,
1
,
UnaryDivFunctor
<
AccT
>>
(
&
out_tmp
[
i
][
0
][
0
],
&
src
data
[
i
][
0
][
0
],
UnaryDivFunctor
<
AccT
>
(
sum
[
i
]));
&
out_tmp
[
i
][
0
][
0
],
&
exp_
data
[
i
][
0
][
0
],
UnaryDivFunctor
<
AccT
>
(
sum
[
i
]));
}
kps
::
WriteData
<
VecT
,
VecT
,
kLoopsV
,
1
,
1
,
true
>
(
&
softmax_v
[
0
],
&
reg_v
[
0
],
idx_max_v
[
i
],
0
,
kWarpSize
,
1
);
...
...
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