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
afeixing77
Paddle
提交
40402d5e
P
Paddle
项目概览
afeixing77
/
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
40402d5e
编写于
2月 15, 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 emb seqpool jitcode
test=develop
上级
2ccbcb15
变更
4
显示空白变更内容
内联
并排
Showing
4 changed file
with
231 addition
and
1 deletion
+231
-1
paddle/fluid/operators/jit/gen/CMakeLists.txt
paddle/fluid/operators/jit/gen/CMakeLists.txt
+1
-0
paddle/fluid/operators/jit/gen/embseqpool.cc
paddle/fluid/operators/jit/gen/embseqpool.cc
+148
-0
paddle/fluid/operators/jit/gen/embseqpool.h
paddle/fluid/operators/jit/gen/embseqpool.h
+81
-0
paddle/fluid/operators/jit/gen/seqpool.h
paddle/fluid/operators/jit/gen/seqpool.h
+1
-1
未找到文件。
paddle/fluid/operators/jit/gen/CMakeLists.txt
浏览文件 @
40402d5e
...
...
@@ -31,3 +31,4 @@ USE_JITKERNEL_GEN(kNCHW16CMulNC)
USE_JITKERNEL_GEN
(
kSeqPool
)
USE_JITKERNEL_GEN
(
kHMax
)
USE_JITKERNEL_GEN
(
kHSum
)
USE_JITKERNEL_GEN
(
kEmbSeqPool
)
paddle/fluid/operators/jit/gen/embseqpool.cc
0 → 100644
浏览文件 @
40402d5e
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License. */
#include "paddle/fluid/operators/jit/gen/embseqpool.h"
#include <stddef.h> // offsetof
#include <vector>
#include "paddle/fluid/operators/jit/gen/act.h" // for exp_float_consts ones
#include "paddle/fluid/operators/jit/registry.h"
#include "paddle/fluid/platform/cpu_info.h"
namespace
paddle
{
namespace
operators
{
namespace
jit
{
namespace
gen
{
void
EmbSeqPoolJitCode
::
genCode
()
{
preCode
();
constexpr
int
block
=
YMM_FLOAT_BLOCK
;
constexpr
int
max_num_regs
=
8
;
const
int
num_block
=
tbl_w_
/
block
;
const
int
num_groups
=
num_block
/
max_num_regs
;
const
size_t
block_size
=
sizeof
(
float
)
*
block
;
std
::
vector
<
int
>
groups
(
num_groups
,
max_num_regs
);
int
rest_num_regs
=
num_block
%
max_num_regs
;
if
(
rest_num_regs
>
0
)
{
groups
.
push_back
(
rest_num_regs
);
}
// protect param_dst
mov
(
reg_ptr_param_dst
,
param_dst
);
mov
(
reg_idx_width_in_byte
,
qword
[
param_attr
+
offsetof
(
emb_seq_pool_attr_t
,
index_width
)]);
mov
(
reg_idx_height
,
qword
[
param_attr
+
offsetof
(
emb_seq_pool_attr_t
,
index_height
)]);
mov
(
rax
,
sizeof
(
int64_t
));
mul
(
reg_idx_width_in_byte
);
mov
(
reg_idx_width_in_byte
,
rax
);
const
size_t
tbl_width_in_byte
=
sizeof
(
float
)
*
tbl_w_
;
int
acc_num_regs
=
0
;
for
(
int
num_regs
:
groups
)
{
Label
l_next_idx_w
,
l_next_idx_h
,
l_save_now
;
xor_
(
reg_idx_w_i_in_byte
,
reg_idx_w_i_in_byte
);
mov
(
reg_ptr_dst_i
,
reg_ptr_param_dst
);
add
(
reg_ptr_dst_i
,
acc_num_regs
*
block_size
);
add
(
param_tbl
,
acc_num_regs
*
block_size
);
L
(
l_next_idx_w
);
{
// h == 0
mov
(
reg_ptr_idx_i
,
param_idx
);
add
(
reg_ptr_idx_i
,
reg_idx_w_i_in_byte
);
mov
(
reg_idx
,
qword
[
reg_ptr_idx_i
]);
mov
(
rax
,
tbl_width_in_byte
);
mul
(
reg_idx
);
mov
(
reg_ptr_tbl_i
,
rax
);
// reg is offset now
add
(
reg_ptr_tbl_i
,
param_tbl
);
// reg is ptr_i now
size_t
w_offset
=
0
;
for
(
int
reg_i
=
0
;
reg_i
<
num_regs
;
++
reg_i
)
{
vmovups
(
ymm_t
(
reg_i
+
num_regs
),
ptr
[
reg_ptr_tbl_i
+
w_offset
]);
w_offset
+=
block_size
;
}
add
(
reg_ptr_idx_i
,
reg_idx_width_in_byte
);
// end condition of idx h
mov
(
reg_idx_h_end
,
reg_idx_height
);
mov
(
rax
,
reg_idx_width_in_byte
);
mul
(
reg_idx_h_end
);
mov
(
reg_idx_h_end
,
rax
);
add
(
reg_idx_h_end
,
reg_idx_w_i_in_byte
);
add
(
reg_idx_h_end
,
param_idx
);
cmp
(
reg_ptr_idx_i
,
reg_idx_h_end
);
jge
(
l_save_now
,
T_NEAR
);
L
(
l_next_idx_h
);
{
mov
(
reg_idx
,
qword
[
reg_ptr_idx_i
]);
mov
(
reg_ptr_tbl_i
,
reg_idx
);
mov
(
rax
,
tbl_width_in_byte
);
mul
(
reg_idx
);
mov
(
reg_ptr_tbl_i
,
rax
);
add
(
reg_ptr_tbl_i
,
param_tbl
);
size_t
w_offset
=
0
;
for
(
int
reg_i
=
0
;
reg_i
<
num_regs
;
++
reg_i
)
{
vmovups
(
ymm_t
(
reg_i
),
ptr
[
reg_ptr_tbl_i
+
w_offset
]);
vaddps
(
ymm_t
(
reg_i
+
num_regs
),
ymm_t
(
reg_i
+
num_regs
),
ymm_t
(
reg_i
));
w_offset
+=
block_size
;
}
add
(
reg_ptr_idx_i
,
reg_idx_width_in_byte
);
cmp
(
reg_ptr_idx_i
,
reg_idx_h_end
);
jl
(
l_next_idx_h
,
T_NEAR
);
}
// end of idx h
L
(
l_save_now
);
// avg or sqrt here, if needed
w_offset
=
0
;
for
(
int
reg_i
=
0
;
reg_i
<
num_regs
;
++
reg_i
)
{
vmovups
(
ptr
[
reg_ptr_dst_i
+
w_offset
],
ymm_t
(
reg_i
+
num_regs
));
w_offset
+=
block_size
;
}
add
(
reg_ptr_dst_i
,
tbl_width_in_byte
);
add
(
reg_idx_w_i_in_byte
,
sizeof
(
int64_t
));
cmp
(
reg_idx_w_i_in_byte
,
reg_idx_width_in_byte
);
jl
(
l_next_idx_w
,
T_NEAR
);
}
// end of idx w
acc_num_regs
+=
num_regs
;
}
// end of groups
postCode
();
}
class
EmbSeqPoolCreator
:
public
JitCodeCreator
<
emb_seq_pool_attr_t
>
{
public:
bool
UseMe
(
const
emb_seq_pool_attr_t
&
attr
)
const
override
{
return
platform
::
MayIUse
(
platform
::
avx
)
&&
attr
.
table_width
%
YMM_FLOAT_BLOCK
==
0
;
}
size_t
CodeSize
(
const
emb_seq_pool_attr_t
&
attr
)
const
override
{
return
96
+
(
attr
.
table_width
/
YMM_FLOAT_BLOCK
)
*
96
*
8
;
}
std
::
unique_ptr
<
GenBase
>
CreateJitCode
(
const
emb_seq_pool_attr_t
&
attr
)
const
override
{
PADDLE_ENFORCE_GT
(
attr
.
table_height
,
0
);
PADDLE_ENFORCE_GT
(
attr
.
table_width
,
0
);
PADDLE_ENFORCE_GT
(
attr
.
index_height
,
0
);
PADDLE_ENFORCE_GT
(
attr
.
index_width
,
0
);
PADDLE_ENFORCE_GT
(
attr
.
out_width
,
0
);
return
make_unique
<
EmbSeqPoolJitCode
>
(
attr
,
CodeSize
(
attr
));
}
};
}
// namespace gen
}
// namespace jit
}
// namespace operators
}
// namespace paddle
namespace
gen
=
paddle
::
operators
::
jit
::
gen
;
REGISTER_JITKERNEL_GEN
(
kEmbSeqPool
,
gen
::
EmbSeqPoolCreator
);
paddle/fluid/operators/jit/gen/embseqpool.h
0 → 100644
浏览文件 @
40402d5e
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License. */
#pragma once
#include <string>
#include "glog/logging.h"
#include "paddle/fluid/operators/jit/gen/jitcode.h"
#include "paddle/fluid/platform/enforce.h"
namespace
paddle
{
namespace
operators
{
namespace
jit
{
namespace
gen
{
class
EmbSeqPoolJitCode
:
public
JitCode
{
public:
explicit
EmbSeqPoolJitCode
(
const
emb_seq_pool_attr_t
&
attr
,
size_t
code_size
=
256
*
1024
,
void
*
code_ptr
=
nullptr
)
:
JitCode
(
code_size
,
code_ptr
),
tbl_w_
(
attr
.
table_width
),
type_
(
attr
.
pool_type
)
{
if
(
type_
!=
SeqPoolType
::
kSum
)
{
LOG
(
FATAL
)
<<
"Only support sum pool yet "
;
}
this
->
genCode
();
}
std
::
string
name
()
const
override
{
std
::
string
base
=
"EmbSeqPoolJitCode"
;
if
(
type_
==
SeqPoolType
::
kSum
)
{
base
+=
"_Sum"
;
}
else
if
(
type_
==
SeqPoolType
::
kAvg
)
{
base
+=
"_Avg"
;
}
else
if
(
type_
==
SeqPoolType
::
kSqrt
)
{
base
+=
"_Sqrt"
;
}
base
+=
(
"_W"
+
std
::
to_string
(
tbl_w_
));
return
base
;
}
void
genCode
()
override
;
private:
int
tbl_w_
;
SeqPoolType
type_
;
reg64_t
param_tbl
{
abi_param1
};
reg64_t
param_idx
{
abi_param2
};
reg64_t
param_dst
{
abi_param3
};
reg64_t
param_attr
{
abi_param4
};
reg64_t
reg_tmp
{
rax
};
reg64_t
reg_idx_width_in_byte
{
r8
};
reg64_t
reg_idx_height
{
r9
};
reg64_t
reg_ptr_tbl_i
{
r10
};
reg64_t
reg_idx
{
r10
};
// could use same of reg_ptr_tbl_i
reg64_t
reg_ptr_idx_i
{
r11
};
reg64_t
reg_ptr_dst_i
{
r12
};
reg64_t
reg_ptr_param_dst
{
r13
};
// rdx is used in mul so protect param_dst
reg64_t
reg_idx_w_i_in_byte
{
r14
};
reg64_t
reg_idx_h_end
{
r15
};
};
}
// namespace gen
}
// namespace jit
}
// namespace operators
}
// namespace paddle
paddle/fluid/operators/jit/gen/seqpool.h
浏览文件 @
40402d5e
...
...
@@ -32,7 +32,7 @@ class SeqPoolJitCode : public JitCode {
:
JitCode
(
code_size
,
code_ptr
),
w_
(
attr
.
w
),
type_
(
attr
.
type
)
{
if
(
!
(
type_
==
SeqPoolType
::
kSum
||
type_
==
SeqPoolType
::
kAvg
||
type_
==
SeqPoolType
::
kSqrt
))
{
LOG
(
FATAL
)
<<
"Only support
sum pool yet
"
;
LOG
(
FATAL
)
<<
"Only support
ed pool type: sum, avg and sqrt.
"
;
}
fp_h_
[
0
]
=
1.
f
;
this
->
genCode
();
...
...
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