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
7b2dc4e6
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
7b2dc4e6
编写于
12月 21, 2020
作者:
W
wangchaochaohu
提交者:
GitHub
12月 21, 2020
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
optimization for fp16 elementwise add (#29744)
上级
27bdbec7
变更
1
隐藏空白更改
内联
并排
Showing
1 changed file
with
38 addition
and
1 deletion
+38
-1
paddle/fluid/operators/elementwise/elementwise_add_op.h
paddle/fluid/operators/elementwise/elementwise_add_op.h
+38
-1
未找到文件。
paddle/fluid/operators/elementwise/elementwise_add_op.h
浏览文件 @
7b2dc4e6
...
@@ -19,6 +19,7 @@ limitations under the License. */
...
@@ -19,6 +19,7 @@ limitations under the License. */
#include "paddle/fluid/operators/elementwise/elementwise_op_function.cu.h"
#include "paddle/fluid/operators/elementwise/elementwise_op_function.cu.h"
#include "paddle/fluid/operators/elementwise/elementwise_op_function.h"
#include "paddle/fluid/operators/elementwise/elementwise_op_function.h"
#include "paddle/fluid/operators/math/blas.h"
#include "paddle/fluid/operators/math/blas.h"
#include "paddle/fluid/operators/math/math_function.h"
#ifdef PADDLE_WITH_CUDA
#ifdef PADDLE_WITH_CUDA
#ifdef __NVCC__
#ifdef __NVCC__
#include "cub/cub.cuh"
#include "cub/cub.cuh"
...
@@ -176,6 +177,25 @@ __global__ void MatrixColReduce(const T *__restrict__ in, T *__restrict__ out,
...
@@ -176,6 +177,25 @@ __global__ void MatrixColReduce(const T *__restrict__ in, T *__restrict__ out,
}
}
}
}
template
<
int
SIZE
>
__global__
void
VecFP16MatrixColReduce
(
const
__half2
*
__restrict__
in
,
__half2
*
__restrict__
out
,
size_t
width
,
size_t
height
)
{
int
idx
=
threadIdx
.
x
+
blockIdx
.
x
*
blockDim
.
x
;
int
by
=
blockIdx
.
y
;
__half2
zero
=
__half2half2
(
static_cast
<
__half
>
(
0
));
const
int
cols
=
width
/
2
;
for
(;
idx
<
cols
;
idx
+=
blockDim
.
x
*
gridDim
.
x
)
{
__half2
sum
=
zero
;
for
(
int
row
=
0
;
row
<
SIZE
;
row
++
)
{
int
index
=
idx
+
(
row
+
by
*
SIZE
)
*
cols
;
sum
=
__hadd2
(
sum
,
in
[
index
]);
}
atomicAdd
(
&
(
out
[
idx
]),
sum
);
}
}
template
<
typename
T
>
template
<
typename
T
>
__global__
void
MatrixReduceLongWidth
(
const
T
*
__restrict__
in
,
T
*
out
,
__global__
void
MatrixReduceLongWidth
(
const
T
*
__restrict__
in
,
T
*
out
,
size_t
width
,
size_t
height
)
{
size_t
width
,
size_t
height
)
{
...
@@ -198,7 +218,7 @@ __global__ void VecMatrixReduceLongWidth(const T *__restrict__ in, T *out,
...
@@ -198,7 +218,7 @@ __global__ void VecMatrixReduceLongWidth(const T *__restrict__ in, T *out,
int
idx
=
threadIdx
.
x
+
blockIdx
.
x
*
blockDim
.
x
;
int
idx
=
threadIdx
.
x
+
blockIdx
.
x
*
blockDim
.
x
;
int
w
=
idx
*
VEC_SIZE
;
int
w
=
idx
*
VEC_SIZE
;
int
width_stride
=
blockDim
.
x
*
gridDim
.
x
*
VEC_SIZE
;
int
width_stride
=
blockDim
.
x
*
gridDim
.
x
*
VEC_SIZE
;
for
(;
w
<
width
;
w
+=
width
)
{
for
(;
w
<
width
;
w
+=
width
_stride
)
{
T
zero
=
static_cast
<
T
>
(
0
);
T
zero
=
static_cast
<
T
>
(
0
);
T
sum
[
VEC_SIZE
]
=
{
zero
};
T
sum
[
VEC_SIZE
]
=
{
zero
};
T
tmp_vec
[
VEC_SIZE
]
=
{
zero
};
T
tmp_vec
[
VEC_SIZE
]
=
{
zero
};
...
@@ -341,6 +361,23 @@ class ElementwiseAddGradKernel : public ElemwiseGradKernel<T> {
...
@@ -341,6 +361,23 @@ class ElementwiseAddGradKernel : public ElemwiseGradKernel<T> {
int
max_blocks
=
std
::
max
(
max_physical_threads
/
(
block_x
*
block_y
),
1
);
int
max_blocks
=
std
::
max
(
max_physical_threads
/
(
block_x
*
block_y
),
1
);
int
theory_block
=
(
width
+
blocks
.
x
-
1
)
/
blocks
.
x
;
int
theory_block
=
(
width
+
blocks
.
x
-
1
)
/
blocks
.
x
;
dim3
grids
(
std
::
min
(
theory_block
,
max_blocks
));
dim3
grids
(
std
::
min
(
theory_block
,
max_blocks
));
if
(
std
::
is_same
<
T
,
paddle
::
platform
::
float16
>::
value
&&
width
<
2048
&&
width
%
2
==
0
&&
height
%
64
==
0
)
{
auto
&
dev_ctx
=
ctx
.
template
device_context
<
platform
::
CUDADeviceContext
>();
math
::
SetConstant
<
platform
::
CUDADeviceContext
,
T
>
functor
;
if
(
dout
->
dims
()
==
dx
->
dims
())
functor
(
dev_ctx
,
dy
,
static_cast
<
T
>
(
0
));
else
functor
(
dev_ctx
,
dx
,
static_cast
<
T
>
(
0
));
const
__half2
*
ptr1
=
reinterpret_cast
<
const
__half2
*>
(
dout_data
);
__half2
*
ptr2
=
reinterpret_cast
<
__half2
*>
(
out_data
);
const
int
threads
=
128
;
dim3
grid
(
1
,
(
height
+
64
-
1
)
/
64
);
VecFP16MatrixColReduce
<
64
><<<
grid
,
threads
,
0
,
stream
>>>
(
ptr1
,
ptr2
,
width
,
height
);
return
;
}
if
(
width
/
height
<
32
)
{
if
(
width
/
height
<
32
)
{
MatrixColReduce
<
T
,
block_x
,
block_y
><<<
grids
,
blocks
,
0
,
stream
>>>
(
MatrixColReduce
<
T
,
block_x
,
block_y
><<<
grids
,
blocks
,
0
,
stream
>>>
(
...
...
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