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
d1d614b9
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
提交
d1d614b9
编写于
1月 17, 2018
作者:
Y
Yibing Liu
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Refine the GPU kernel for sequence_erase_op
上级
7d3b2e4b
变更
1
隐藏空白更改
内联
并排
Showing
1 changed file
with
41 addition
and
28 deletion
+41
-28
paddle/operators/sequence_erase_op.cu
paddle/operators/sequence_erase_op.cu
+41
-28
未找到文件。
paddle/operators/sequence_erase_op.cu
浏览文件 @
d1d614b9
...
...
@@ -42,8 +42,8 @@ __global__ void LabelErasedIdx(const T* in_dat, const int in_len,
}
template
<
typename
T
>
__global__
void
GetOutLod
(
const
T
*
num_erased
,
const
in
t
*
in_lod
,
const
int
lod_len
,
in
t
*
out_lod0
)
{
__global__
void
GetOutLod
(
const
T
*
num_erased
,
const
size_
t
*
in_lod
,
const
int
lod_len
,
size_
t
*
out_lod0
)
{
int
index
=
blockIdx
.
x
*
blockDim
.
x
+
threadIdx
.
x
;
if
(
index
<
lod_len
)
{
out_lod0
[
index
]
=
in_lod
[
index
]
-
num_erased
[
in_lod
[
index
]];
...
...
@@ -61,6 +61,26 @@ __global__ void SetOutput(const T* in_dat, const int in_len,
}
}
template
<
typename
T
,
typename
Vector
>
thrust
::
device_vector
<
T
>
set_device_vector
(
Vector
&
vector
)
{
thrust
::
host_vector
<
T
>
host_vec
(
vector
.
size
());
for
(
size_t
i
=
0
;
i
<
vector
.
size
();
++
i
)
{
host_vec
[
i
]
=
vector
[
i
];
}
thrust
::
device_vector
<
T
>
dev_vec
=
host_vec
;
return
dev_vec
;
}
template
<
typename
T
>
std
::
vector
<
T
>
get_std_vector
(
thrust
::
device_vector
<
T
>&
dev_vec
)
{
thrust
::
host_vector
<
T
>
host_vec
=
dev_vec
;
std
::
vector
<
T
>
std_vec
(
host_vec
.
size
(),
0
);
for
(
size_t
i
=
0
;
i
<
host_vec
.
size
();
++
i
)
{
std_vec
[
i
]
=
host_vec
[
i
];
}
return
std_vec
;
}
template
<
typename
T
>
class
SequenceEraseOpCUDAKernel
:
public
framework
::
OpKernel
<
T
>
{
public:
...
...
@@ -73,52 +93,45 @@ class SequenceEraseOpCUDAKernel : public framework::OpKernel<T> {
PADDLE_ENFORCE_EQ
(
lod
[
0
].
back
(),
(
size_t
)
in
->
numel
(),
"The actual size mismatches with the LoD information."
);
auto
tokens
=
ctx
.
Attr
<
std
::
vector
<
T
>>
(
"tokens"
);
auto
tokens_len
=
tokens
.
size
();
auto
in_len
=
in
->
numel
();
auto
in_dat
=
in
->
data
<
T
>
();
auto
lod0
=
lod
[
0
];
// Copy tokens to GPU
thrust
::
device_vector
<
T
>
dev_tokens
=
set_device_vector
<
T
,
std
::
vector
<
T
>>
(
tokens
);
T
*
dev_tokens_ptr
=
thrust
::
raw_pointer_cast
(
dev_tokens
.
data
());
thrust
::
host_vector
<
T
>
host_tokens
(
tokens_len
);
for
(
size_t
i
=
0
;
i
<
tokens
.
size
();
++
i
)
{
host_tokens
[
i
]
=
tokens
[
i
];
}
thrust
::
device_vector
<
T
>
dev_tokens
=
host_tokens
;
// Count number of elements to be erased
thrust
::
device_vector
<
int
>
num_erased
(
in_len
+
1
);
T
*
dev_tokens_ptr
=
thrust
::
raw_pointer_cast
(
dev_tokens
.
data
());
int
*
num_erased_ptr
=
thrust
::
raw_pointer_cast
(
num_erased
.
data
());
auto
stream
=
ctx
.
cuda_device_context
().
stream
();
LabelErasedIdx
<<<
(
in_len
-
1
)
/
PADDLE_CUDA_NUM_THREADS
+
1
,
PADDLE_CUDA_NUM_THREADS
,
0
,
stream
>>>
(
in_dat
,
in_len
,
dev_tokens_ptr
,
tokens
_len
,
num_erased_ptr
);
in_dat
,
in_len
,
dev_tokens_ptr
,
tokens
.
size
()
,
num_erased_ptr
);
thrust
::
inclusive_scan
(
num_erased
.
begin
()
+
1
,
num_erased
.
end
(),
num_erased
.
begin
()
+
1
);
// Calc LoD
// Copy LoD to GPU
auto
lod0
=
lod
[
0
];
auto
lod_len
=
lod0
.
size
();
thrust
::
host_vector
<
int
>
host_lod
(
lod_len
);
for
(
size_t
i
=
0
;
i
<
lod_len
;
++
i
)
{
host_lod
[
i
]
=
lod0
[
i
];
}
thrust
::
device_vector
<
int
>
dev_in_lod
=
host_lod
;
thrust
::
device_vector
<
int
>
dev_out_lod
(
lod_len
);
int
*
dev_in_lod_ptr
=
thrust
::
raw_pointer_cast
(
dev_in_lod
.
data
());
int
*
dev_out_lod_ptr
=
thrust
::
raw_pointer_cast
(
dev_out_lod
.
data
());
thrust
::
device_vector
<
size_t
>
dev_in_lod
=
set_device_vector
<
size_t
,
paddle
::
framework
::
Vector
<
size_t
>>
(
lod0
);
size_t
*
dev_in_lod_ptr
=
thrust
::
raw_pointer_cast
(
dev_in_lod
.
data
());
// Calc output LoD
thrust
::
device_vector
<
size_t
>
dev_out_lod
(
lod_len
);
size_t
*
dev_out_lod_ptr
=
thrust
::
raw_pointer_cast
(
dev_out_lod
.
data
());
GetOutLod
<<<
(
lod_len
-
1
)
/
PADDLE_CUDA_NUM_THREADS
+
1
,
PADDLE_CUDA_NUM_THREADS
,
0
,
stream
>>>
(
num_erased_ptr
,
dev_in_lod_ptr
,
lod_len
,
dev_out_lod_ptr
);
thrust
::
host_vector
<
int
>
host_out_lod
=
dev_out_lod
;
std
::
vector
<
int
>
out_lod0
(
lod_len
,
0
);
for
(
size_t
i
=
0
;
i
<
lod_len
;
i
++
)
{
out_lod0
[
i
]
=
host_out_lod
[
i
];
}
// Set LoD for output
std
::
vector
<
size_t
>
out_lod0
=
get_std_vector
<
size_t
>
(
dev_out_lod
);
framework
::
LoD
out_lod
;
out_lod
.
push_back
(
out_lod0
);
out
->
set_lod
(
out_lod
);
// Set output
out
->
Resize
({
out_lod0
.
back
(
),
1
});
out
->
Resize
({
static_cast
<
int64_t
>
(
out_lod0
.
back
()
),
1
});
auto
out_dat
=
out
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
SetOutput
<<<
(
in_len
-
1
)
/
PADDLE_CUDA_NUM_THREADS
+
1
,
PADDLE_CUDA_NUM_THREADS
,
0
,
stream
>>>
(
in_dat
,
in_len
,
...
...
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