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
csdn_franckjun
Paddle
提交
acef85b2
P
Paddle
项目概览
csdn_franckjun
/
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 >>
未验证
提交
acef85b2
编写于
12月 26, 2021
作者:
Z
Zhang Ting
提交者:
GitHub
12月 26, 2021
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
improve forward performace (#38279)
上级
e6c3f64f
变更
1
隐藏空白更改
内联
并排
Showing
1 changed file
with
19 addition
and
20 deletion
+19
-20
paddle/fluid/operators/dropout_impl.cu.h
paddle/fluid/operators/dropout_impl.cu.h
+19
-20
未找到文件。
paddle/fluid/operators/dropout_impl.cu.h
浏览文件 @
acef85b2
...
...
@@ -34,6 +34,7 @@ limitations under the License. */
#include "paddle/fluid/operators/dropout_op.h"
#include "paddle/fluid/platform/aligned_vector.h"
#include "paddle/fluid/platform/device/gpu/gpu_launch_config.h"
#include "paddle/pten/kernels/hybird/cuda/elementwise/elementwise_no_broadcast.cu.h"
namespace
paddle
{
namespace
operators
{
...
...
@@ -180,9 +181,6 @@ void DropoutFwGPUKernelDriver(const platform::CUDADeviceContext& dev_ctx,
return
;
}
platform
::
GpuLaunchConfig
config
=
platform
::
GetGpuLaunchConfig1D
(
dev_ctx
,
size
);
// increment is used to set the args(offset) of curand_init, which defines
// offset in subsequence.
// The detail:
...
...
@@ -192,11 +190,15 @@ void DropoutFwGPUKernelDriver(const platform::CUDADeviceContext& dev_ctx,
// same as the previous calls.
uint64_t
seed_data
;
uint64_t
increment
;
int
vec_size
=
platform
::
GetVectorizedSize
<
T
>
(
x_data
);
auto
offset
=
((
x_numel
-
1
)
/
(
config
.
block_per_grid
.
x
*
config
.
thread_per_block
.
x
*
vec_size
)
+
1
)
*
vec_size
;
// VectorizedRandomGenerator use curand_uniform4, so we only support
// vec_size is 4;
int
vec_size
=
(
platform
::
GetVectorizedSize
<
T
>
(
x_data
)
==
4
)
?
4
:
1
;
int
block_size
=
pten
::
GetThreadsConfig
(
dev_ctx
,
x_numel
,
vec_size
);
int
grid_size
=
((
x_numel
+
vec_size
-
1
)
/
vec_size
+
block_size
-
1
)
/
block_size
;
auto
offset
=
((
x_numel
-
1
)
/
(
grid_size
*
block_size
*
vec_size
)
+
1
)
*
vec_size
;
GetSeedDataAndIncrement
(
dev_ctx
,
seed
,
is_fix_seed
,
seed_val
,
offset
,
&
seed_data
,
&
increment
);
...
...
@@ -204,26 +206,23 @@ void DropoutFwGPUKernelDriver(const platform::CUDADeviceContext& dev_ctx,
#ifdef __HIPCC__
if
(
vec_size
==
4
&&
size
%
4
==
0
)
{
hipLaunchKernelGGL
(
HIP_KERNEL_NAME
(
VectorizedRandomGenerator
<
T
,
uint8_t
,
4
>
),
config
.
block_per_grid
,
config
.
thread_per_block
,
0
,
stream
,
size
,
seed_data
,
dropout_prob
,
x_data
,
mask_data
,
y_data
,
upscale_in_train
,
increment
);
HIP_KERNEL_NAME
(
VectorizedRandomGenerator
<
T
,
uint8_t
,
4
>
),
grid_size
,
block_size
,
0
,
stream
,
size
,
seed_data
,
dropout_prob
,
x_data
,
mask_data
,
y_data
,
upscale_in_train
,
increment
);
}
else
{
hipLaunchKernelGGL
(
HIP_KERNEL_NAME
(
RandomGenerator
<
T
,
uint8_t
>
),
config
.
block_per_grid
,
config
.
thread_per_block
,
0
,
stream
,
size
,
seed_data
,
dropout_prob
,
x
_data
,
mask_data
,
y_data
,
upscale_in_train
,
increment
);
grid_size
,
block_size
,
0
,
stream
,
size
,
seed_data
,
dropout_prob
,
x_data
,
mask_data
,
y
_data
,
upscale_in_train
,
increment
);
}
#else
if
(
vec_size
==
4
&&
size
%
4
==
0
)
{
VectorizedRandomGenerator
<
T
,
uint8_t
,
4
><<<
config
.
block_per_grid
,
config
.
thread_per_block
,
0
,
stream
>>>
(
VectorizedRandomGenerator
<
T
,
uint8_t
,
4
><<<
grid_size
,
block_size
,
0
,
stream
>>>
(
size
,
seed_data
,
dropout_prob
,
x_data
,
mask_data
,
y_data
,
upscale_in_train
,
increment
);
}
else
{
RandomGenerator
<
T
,
uint8_t
><<<
config
.
block_per_grid
,
config
.
thread_per_block
,
0
,
stream
>>>
(
RandomGenerator
<
T
,
uint8_t
><<<
grid_size
,
block_size
,
0
,
stream
>>>
(
size
,
seed_data
,
dropout_prob
,
x_data
,
mask_data
,
y_data
,
upscale_in_train
,
increment
);
}
...
...
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