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
PaddlePaddle
Paddle
提交
c8adc2c6
P
Paddle
项目概览
PaddlePaddle
/
Paddle
12 个月 前同步成功
通知
2292
Star
20931
Fork
5422
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
1423
列表
看板
标记
里程碑
合并请求
543
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
1,423
Issue
1,423
列表
看板
标记
里程碑
合并请求
543
合并请求
543
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
c8adc2c6
编写于
10月 29, 2018
作者:
D
dzhwinter
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
cudnn version. staged.
上级
7141debe
变更
3
显示空白变更内容
内联
并排
Showing
3 changed file
with
70 addition
and
36 deletion
+70
-36
paddle/fluid/operators/top_k_op.cc
paddle/fluid/operators/top_k_op.cc
+1
-1
paddle/fluid/operators/top_k_op.cu
paddle/fluid/operators/top_k_op.cu
+68
-31
paddle/fluid/operators/top_k_op.h
paddle/fluid/operators/top_k_op.h
+1
-4
未找到文件。
paddle/fluid/operators/top_k_op.cc
浏览文件 @
c8adc2c6
...
...
@@ -50,7 +50,7 @@ class TopkOpMaker : public framework::OpProtoAndCheckerMaker {
public:
void
Make
()
override
{
AddInput
(
"X"
,
"(Tensor) The input of Topk op"
);
AddOutput
(
"Out"
,
"(Tensor) The output tensor of Topk op"
)
.
Reuse
(
"X"
)
;
AddOutput
(
"Out"
,
"(Tensor) The output tensor of Topk op"
);
AddOutput
(
"Indices"
,
"(Tensor) The indices of Topk elements of input"
);
AddComment
(
R"DOC(
Top K operator
...
...
paddle/fluid/operators/top_k_op.cu
浏览文件 @
c8adc2c6
...
...
@@ -256,36 +256,65 @@ __device__ __forceinline__ void BlockReduce(Pair<T>* sh_topk, int* maxid,
* 3. go to the second setp, until one thread's topk value is null;
* 4. go to the first setp, until get the topk value.
*/
template
<
typename
T
,
int
MaxLength
,
int
BlockSize
>
__global__
void
KeMatrixTopK
(
T
*
output
,
int
output_stride
,
int64_t
*
indices
,
const
T
*
src
,
int
lds
,
int
dim
,
int
k
)
{
const
T
*
src
,
int
lds
,
int
dim
,
int
k
,
int
grid_dim
,
int
num
)
{
__shared__
Pair
<
T
>
sh_topk
[
BlockSize
];
__shared__
int
maxid
[
BlockSize
/
2
];
const
int
tid
=
threadIdx
.
x
;
const
int
warp
=
threadIdx
.
x
/
32
;
output
+=
blockIdx
.
x
*
output_stride
;
indices
+=
blockIdx
.
x
*
k
;
const
int
bid
=
blockIdx
.
x
;
for
(
int
i
=
bid
;
i
<
num
;
i
+=
grid_dim
)
{
int
top_num
=
k
;
__shared__
int
maxid
[
BlockSize
/
2
];
T
*
out
=
output
+
i
*
output_stride
;
int64_t
*
inds
=
indices
+
i
*
k
;
Pair
<
T
>
topk
[
MaxLength
];
int
beam
=
MaxLength
;
Pair
<
T
>
max
;
bool
is_empty
=
false
;
bool
firststep
=
true
;
for
(
int
k
=
0
;
k
<
MaxLength
;
k
++
)
{
topk
[
k
].
set
(
-
INFINITY
,
-
1
);
for
(
int
j
=
0
;
j
<
MaxLength
;
j
++
)
{
topk
[
j
].
set
(
-
INFINITY
,
-
1
);
}
while
(
k
)
{
ThreadGetTopK
<
T
,
MaxLength
,
BlockSize
>
(
topk
,
&
beam
,
k
,
src
+
blockIdx
.
x
*
lds
,
&
firststep
,
&
is_empty
,
&
max
,
dim
,
tid
);
while
(
top_num
)
{
ThreadGetTopK
<
T
,
MaxLength
,
BlockSize
>
(
topk
,
&
beam
,
k
,
src
+
i
*
lds
,
&
firststep
,
&
is_empty
,
&
max
,
dim
,
tid
);
sh_topk
[
tid
]
=
topk
[
0
];
BlockReduce
<
T
,
MaxLength
,
BlockSize
>
(
sh_topk
,
maxid
,
topk
,
&
output
,
&
indices
,
&
beam
,
&
k
,
tid
,
warp
);
BlockReduce
<
T
,
MaxLength
,
BlockSize
>
(
sh_topk
,
maxid
,
topk
,
&
out
,
&
inds
,
&
beam
,
&
top_num
,
tid
,
warp
);
}
}
}
inline
static
int
GetDesiredBlockDim
(
int
dim
)
{
if
(
dim
>
128
)
{
return
256
;
}
else
if
(
dim
>
64
)
{
return
128
;
}
else
if
(
dim
>
32
)
{
return
64
;
}
else
{
return
32
;
}
}
#define FIXED_BLOCK_DIM_BASE(dim, ...) \
case (dim): { \
constexpr auto kBlockDim = (dim); \
__VA_ARGS__; \
} break
#define FIXED_BLOCK_DIM(...) \
FIXED_BLOCK_DIM_BASE(256, ##__VA_ARGS__); \
FIXED_BLOCK_DIM_BASE(128, ##__VA_ARGS__); \
FIXED_BLOCK_DIM_BASE(64, ##__VA_ARGS__); \
FIXED_BLOCK_DIM_BASE(32, ##__VA_ARGS__)
template
<
typename
T
>
class
TopkOpCUDAKernel
:
public
framework
::
OpKernel
<
T
>
{
public:
...
...
@@ -298,30 +327,38 @@ class TopkOpCUDAKernel : public framework::OpKernel<T> {
size_t
k
=
static_cast
<
int
>
(
ctx
.
Attr
<
int
>
(
"k"
));
const
T
*
input_data
=
input
->
data
<
T
>
();
T
*
output_data
=
output
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
// FIXME(typhoonzero): data is always converted to type T?
int64_t
*
indices_data
=
indices
->
mutable_data
<
int64_t
>
(
ctx
.
GetPlace
());
size_t
input_height
=
input
->
dims
()[
0
];
size_t
input_width
=
input
->
dims
()[
1
];
framework
::
DDim
inputdims
=
input
->
dims
();
const
size_t
input_height
=
framework
::
product
(
framework
::
slice_ddim
(
inputdims
,
0
,
inputdims
.
size
()
-
1
));
const
size_t
input_width
=
inputdims
[
inputdims
.
size
()
-
1
];
if
(
k
>
input_width
)
k
=
input_width
;
// NOTE: pass lds and dim same to input width.
// NOTE: old matrix implementation of stride is different to eigen.
// TODO(typhoonzero): refine this kernel.
dim3
threads
(
256
,
1
);
dim3
grid
(
input_height
,
1
);
KeMatrixTopK
<
T
,
5
,
256
><<<
grid
,
threads
,
0
,
reinterpret_cast
<
const
platform
::
CUDADeviceContext
&>
(
ctx
.
device_context
())
.
stream
()
>>>
(
output_data
,
output
->
dims
()[
1
],
indices_data
,
input_data
,
input_width
,
input_width
,
static_cast
<
int
>
(
k
));
const
int
kMaxHeight
=
2048
;
int
gridx
=
input_height
<
kMaxHeight
?
input_height
:
kMaxHeight
;
auto
&
dev_ctx
=
ctx
.
cuda_device_context
();
switch
(
GetDesiredBlockDim
(
input_width
))
{
FIXED_BLOCK_DIM
(
KeMatrixTopK
<
T
,
5
,
kBlockDim
><<<
gridx
,
kBlockDim
,
0
,
dev_ctx
.
stream
()
>>>
(
output_data
,
k
,
indices_data
,
input_data
,
input_width
,
input_width
,
static_cast
<
int
>
(
k
),
gridx
,
input_height
));
default:
PADDLE_THROW
(
"Error"
);
}
}
};
#undef FIXED_BLOCK_DIM_BASE
#undef FIXED_BLOCK_DIM
}
// namespace operators
}
// namespace paddle
...
...
paddle/fluid/operators/top_k_op.h
浏览文件 @
c8adc2c6
...
...
@@ -34,7 +34,6 @@ class TopkKernel : public framework::OpKernel<T> {
public:
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
// Get the top k elements of each row of input tensor
// FIXME: only deal with matrix(2d tensor).
auto
*
input
=
ctx
.
Input
<
Tensor
>
(
"X"
);
auto
*
output
=
ctx
.
Output
<
Tensor
>
(
"Out"
);
auto
*
indices
=
ctx
.
Output
<
Tensor
>
(
"Indices"
);
...
...
@@ -44,8 +43,6 @@ class TopkKernel : public framework::OpKernel<T> {
T
*
output_data
=
output
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
int64_t
*
indices_data
=
indices
->
mutable_data
<
int64_t
>
(
ctx
.
GetPlace
());
auto
eg_input
=
EigenMatrix
<
T
>::
From
(
*
input
);
// reshape input to a flattern matrix(like flat_inner_dims)
framework
::
DDim
inputdims
=
input
->
dims
();
const
size_t
row
=
framework
::
product
(
...
...
@@ -53,7 +50,7 @@ class TopkKernel : public framework::OpKernel<T> {
const
size_t
col
=
inputdims
[
inputdims
.
size
()
-
1
];
Eigen
::
DSizes
<
int
,
2
>
flat2dims
(
row
,
col
);
// NOTE: eigen shape doesn't affect paddle tensor.
eg_input
.
reshape
(
flat2dims
);
auto
eg_input
=
EigenMatrix
<
T
>::
Reshape
(
*
input
,
inputdims
.
size
()
-
1
);
#ifdef PADDLE_WITH_MKLML
#pragma omp parallel for
...
...
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