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
wmsofts
Paddle
提交
a3a3d3d8
P
Paddle
项目概览
wmsofts
/
Paddle
与 Fork 源项目一致
Fork自
PaddlePaddle / Paddle
通知
1
Star
0
Fork
0
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
0
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
0
Issue
0
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 >>
提交
a3a3d3d8
编写于
2月 14, 2019
作者:
T
tensor-tang
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
add embseqpool jitkernel mkl impl and use it
test=develop
上级
15da2f9a
变更
4
隐藏空白更改
内联
并排
Showing
4 changed file
with
50 addition
and
32 deletion
+50
-32
paddle/fluid/operators/fused/fused_embedding_seq_pool_op.h
paddle/fluid/operators/fused/fused_embedding_seq_pool_op.h
+9
-32
paddle/fluid/operators/jit/more/mkl/CMakeLists.txt
paddle/fluid/operators/jit/more/mkl/CMakeLists.txt
+1
-0
paddle/fluid/operators/jit/more/mkl/mkl.cc
paddle/fluid/operators/jit/more/mkl/mkl.cc
+11
-0
paddle/fluid/operators/jit/more/mkl/mkl.h
paddle/fluid/operators/jit/more/mkl/mkl.h
+29
-0
未找到文件。
paddle/fluid/operators/fused/fused_embedding_seq_pool_op.h
浏览文件 @
a3a3d3d8
...
...
@@ -21,6 +21,7 @@ limitations under the License. */
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/selected_rows.h"
#include "paddle/fluid/operators/jit/kernels.h"
#include "paddle/fluid/operators/math/blas.h"
namespace
paddle
{
...
...
@@ -31,35 +32,6 @@ using LoDTensor = framework::LoDTensor;
using
SelectedRows
=
framework
::
SelectedRows
;
using
DDim
=
framework
::
DDim
;
template
<
typename
T
>
void
emb_seqpool
(
const
framework
::
ExecutionContext
&
context
,
const
T
*
table
,
const
int64_t
*
idx
,
T
*
out
,
int64_t
table_height
,
int64_t
table_width
,
int64_t
idx_height
,
int64_t
idx_width
,
int64_t
out_width
)
{
// pool type == sum
PADDLE_ENFORCE_EQ
(
table_width
*
idx_width
,
out_width
);
auto
check_idx_value_valid
=
[
&
](
int
i
)
{
PADDLE_ENFORCE_LT
(
idx
[
i
],
table_height
,
"idx value: %d, i: %d"
,
idx
[
i
],
i
);
PADDLE_ENFORCE_GE
(
idx
[
i
],
0
,
"idx value: %d, i: %d"
,
idx
[
i
],
i
);
};
auto
blas
=
math
::
GetBlas
<
platform
::
CPUDeviceContext
,
T
>
(
context
);
for
(
int
w
=
0
;
w
!=
idx_width
;
++
w
)
{
check_idx_value_valid
(
w
);
blas
.
VCOPY
(
table_width
,
table
+
idx
[
w
]
*
table_width
,
out
+
w
*
table_width
);
}
for
(
int
h
=
1
;
h
<
idx_height
;
++
h
)
{
for
(
int
w
=
0
;
w
<
idx_width
;
++
w
)
{
int
i
=
h
*
idx_width
+
w
;
check_idx_value_valid
(
i
);
blas
.
AXPY
(
table_width
,
static_cast
<
T
>
(
1
),
table
+
idx
[
i
]
*
table_width
,
out
+
w
*
table_width
);
}
}
}
template
<
typename
T
>
struct
EmbeddingVSumFunctor
{
void
operator
()(
const
framework
::
ExecutionContext
&
context
,
...
...
@@ -75,10 +47,15 @@ struct EmbeddingVSumFunctor {
auto
*
output
=
output_t
->
mutable_data
<
T
>
(
context
.
GetPlace
());
PADDLE_ENFORCE_LE
(
table_width
*
idx_width
,
out_width
);
jit
::
emb_seq_pool_attr_t
attr
(
table_height
,
table_width
,
0
,
idx_width
,
out_width
,
jit
::
SeqPoolType
::
kSum
);
for
(
int64_t
i
=
0
;
i
!=
ids_lod
.
size
()
-
1
;
++
i
)
{
emb_seqpool
(
context
,
table
,
ids
+
ids_lod
[
i
]
*
idx_width
,
output
+
i
*
out_width
,
table_height
,
table_width
,
ids_lod
[
i
+
1
]
-
ids_lod
[
i
],
idx_width
,
out_width
);
attr
.
index_height
=
ids_lod
[
i
+
1
]
-
ids_lod
[
i
];
auto
emb_seqpool
=
jit
::
Get
<
jit
::
kEmbSeqPool
,
jit
::
EmbSeqPoolTuples
<
T
>
,
platform
::
CPUPlace
>
(
attr
);
emb_seqpool
(
table
,
ids
+
ids_lod
[
i
]
*
idx_width
,
output
+
i
*
out_width
,
&
attr
);
}
}
};
...
...
paddle/fluid/operators/jit/more/mkl/CMakeLists.txt
浏览文件 @
a3a3d3d8
...
...
@@ -13,3 +13,4 @@ USE_JITKERNEL_MORE(kVSigmoid, mkl)
USE_JITKERNEL_MORE
(
kVTanh, mkl
)
USE_JITKERNEL_MORE
(
kSeqPool, mkl
)
USE_JITKERNEL_MORE
(
kSoftmax, mkl
)
USE_JITKERNEL_MORE
(
kEmbSeqPool, mkl
)
paddle/fluid/operators/jit/more/mkl/mkl.cc
浏览文件 @
a3a3d3d8
...
...
@@ -174,6 +174,16 @@ bool SeqPoolKernel<double>::UseMe(const seq_pool_attr_t& attr) const {
return
true
;
}
template
<
>
bool
EmbSeqPoolKernel
<
float
>::
UseMe
(
const
emb_seq_pool_attr_t
&
attr
)
const
{
return
true
;
}
template
<
>
bool
EmbSeqPoolKernel
<
double
>::
UseMe
(
const
emb_seq_pool_attr_t
&
attr
)
const
{
return
true
;
}
template
<
>
bool
MatMulKernel
<
float
>::
UseMe
(
const
matmul_attr_t
&
attr
)
const
{
return
platform
::
MayIUse
(
platform
::
avx
);
...
...
@@ -227,6 +237,7 @@ REGISTER_MKL_KERNEL(kVSquare, VSquare);
REGISTER_MKL_KERNEL
(
kVSigmoid
,
VSigmoid
);
REGISTER_MKL_KERNEL
(
kVTanh
,
VTanh
);
REGISTER_MKL_KERNEL
(
kSeqPool
,
SeqPool
);
REGISTER_MKL_KERNEL
(
kEmbSeqPool
,
EmbSeqPool
);
REGISTER_MKL_KERNEL
(
kSoftmax
,
Softmax
);
#undef REGISTER_MKL_KERNEL
paddle/fluid/operators/jit/more/mkl/mkl.h
浏览文件 @
a3a3d3d8
...
...
@@ -18,6 +18,7 @@
#include <type_traits>
#include <vector>
#include "paddle/fluid/operators/jit/kernel_base.h"
#include "paddle/fluid/platform/enforce.h"
namespace
paddle
{
namespace
operators
{
...
...
@@ -91,6 +92,32 @@ void SeqPool(const T* x, T* y, const seq_pool_attr_t* attr) {
}
}
template
<
typename
T
>
void
EmbSeqPool
(
const
T
*
table
,
const
int64_t
*
idx
,
T
*
out
,
const
emb_seq_pool_attr_t
*
attr
)
{
PADDLE_ENFORCE_EQ
(
attr
->
table_width
*
attr
->
index_width
,
attr
->
out_width
);
auto
check_idx_value_valid
=
[
&
](
int64_t
i
)
{
PADDLE_ENFORCE_LT
(
idx
[
i
],
attr
->
table_height
,
"idx value: %d, i: %d"
,
idx
[
i
],
i
);
PADDLE_ENFORCE_GE
(
idx
[
i
],
0
,
"idx value: %d, i: %d"
,
idx
[
i
],
i
);
};
for
(
int64_t
w
=
0
;
w
!=
attr
->
index_width
;
++
w
)
{
check_idx_value_valid
(
w
);
VCopy
<
T
>
(
table
+
idx
[
w
]
*
attr
->
table_width
,
out
+
w
*
attr
->
table_width
,
attr
->
table_width
);
}
for
(
int64_t
h
=
1
;
h
<
attr
->
index_height
;
++
h
)
{
for
(
int64_t
w
=
0
;
w
<
attr
->
index_width
;
++
w
)
{
int64_t
i
=
h
*
attr
->
index_width
+
w
;
check_idx_value_valid
(
i
);
VAXPY
<
T
>
(
static_cast
<
T
>
(
1
),
table
+
idx
[
i
]
*
attr
->
table_width
,
out
+
w
*
attr
->
table_width
,
attr
->
table_width
);
}
}
}
template
<
typename
T
>
void
ASum
(
const
T
*
x
,
T
*
res
,
int
n
);
...
...
@@ -142,6 +169,8 @@ DECLARE_MKL_KERNEL(VSquare, XYNTuples);
DECLARE_MKL_KERNEL
(
SeqPool
,
SeqPoolTuples
);
DECLARE_MKL_KERNEL
(
EmbSeqPool
,
EmbSeqPoolTuples
);
DECLARE_MKL_KERNEL
(
Softmax
,
SoftmaxTuples
);
#undef DECLARE_MKL_KERNEL
...
...
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