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
s920243400
PaddleDetection
提交
3713d08d
P
PaddleDetection
项目概览
s920243400
/
PaddleDetection
与 Fork 源项目一致
Fork自
PaddlePaddle / PaddleDetection
通知
2
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
PaddleDetection
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
提交
3713d08d
编写于
12月 16, 2018
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
enable jitcode gru
上级
7c1f3ad6
变更
3
隐藏空白更改
内联
并排
Showing
3 changed file
with
235 addition
and
0 deletion
+235
-0
paddle/fluid/operators/jit/gen/CMakeLists.txt
paddle/fluid/operators/jit/gen/CMakeLists.txt
+3
-0
paddle/fluid/operators/jit/gen/gru.cc
paddle/fluid/operators/jit/gen/gru.cc
+116
-0
paddle/fluid/operators/jit/gen/gru.h
paddle/fluid/operators/jit/gen/gru.h
+116
-0
未找到文件。
paddle/fluid/operators/jit/gen/CMakeLists.txt
浏览文件 @
3713d08d
...
...
@@ -22,3 +22,6 @@ USE_JITKERNEL_GEN(vsigmoid)
USE_JITKERNEL_GEN
(
vtanh
)
USE_JITKERNEL_GEN
(
lstmctht
)
USE_JITKERNEL_GEN
(
lstmc1h1
)
USE_JITKERNEL_GEN
(
gruh1
)
USE_JITKERNEL_GEN
(
gruhtpart1
)
USE_JITKERNEL_GEN
(
gruhtpart2
)
paddle/fluid/operators/jit/gen/gru.cc
0 → 100644
浏览文件 @
3713d08d
/* 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/gru.h"
#include <stddef.h> // offsetof
#include "paddle/fluid/operators/jit/registry.h"
#include "paddle/fluid/platform/cpu_info.h"
namespace
paddle
{
namespace
operators
{
namespace
jit
{
namespace
gen
{
void
GRUJitCode
::
genCode
()
{
reg64_t
reg_ptr_gates
=
rax
;
reg64_t
reg_ptr_ht_1
=
r9
;
reg64_t
reg_ptr_ht
=
r10
;
mov
(
reg_ptr_gates
,
ptr
[
param1
+
offsetof
(
gru_t
,
gates
)]);
mov
(
reg_ptr_ht_1
,
ptr
[
param1
+
offsetof
(
gru_t
,
ht_1
)]);
mov
(
reg_ptr_ht
,
ptr
[
param1
+
offsetof
(
gru_t
,
ht
)]);
ymm_t
ymm_one
=
ymm_t
(
0
);
if
(
id_
==
2
)
{
reg64_t
reg_ptr_tmp
=
r11
;
mov
(
reg_ptr_tmp
,
reinterpret_cast
<
size_t
>
(
exp_float_consts
));
vmovaps
(
ymm_one
,
ptr
[
reg_ptr_tmp
+
OFFSET_EXP_ONE
]);
}
int
offset
=
0
;
int
d
=
num_
*
sizeof
(
float
);
for
(
int
i
=
0
;
i
<
num_
/
YMM_FLOAT_BLOCK
;
++
i
)
{
ymm_t
ymm_u
=
ymm_t
(
1
);
ymm_t
ymm_r
=
ymm_t
(
2
);
ymm_t
ymm_s
=
ymm_t
(
3
);
ymm_t
ymm_ht_1
=
ymm_t
(
4
);
// W: {W_update, W_reset; W_state}
if
(
id_
==
0
||
id_
==
2
)
{
vmovups
(
ymm_u
,
ptr
[
reg_ptr_gates
+
offset
]);
vmovups
(
ymm_s
,
ptr
[
reg_ptr_gates
+
offset
+
2
*
d
]);
}
if
(
id_
==
1
)
{
vmovups
(
ymm_r
,
ptr
[
reg_ptr_gates
+
offset
+
d
]);
}
if
(
id_
==
1
||
id_
==
2
)
{
vmovups
(
ymm_ht_1
,
ptr
[
reg_ptr_ht_1
+
offset
]);
}
if
(
id_
==
0
)
{
// ht = act_gate(u) * act_cand(s)
act
<
ymm_t
>
(
ymm_u
,
ymm_u
,
act_gate_
);
act
<
ymm_t
>
(
ymm_s
,
ymm_s
,
act_cand_
);
vmulps
(
ymm_s
,
ymm_s
,
ymm_u
);
vmovups
(
ptr
[
reg_ptr_ht
+
offset
],
ymm_s
);
}
else
if
(
id_
==
1
)
{
// ht = act_gate(r) * ht_1
act
<
ymm_t
>
(
ymm_r
,
ymm_r
,
act_gate_
);
vmulps
(
ymm_r
,
ymm_r
,
ymm_ht_1
);
vmovups
(
ptr
[
reg_ptr_ht
+
offset
],
ymm_r
);
}
else
if
(
id_
==
2
)
{
// ht = act_gate(u) * act_cand(s) + (1-act_gate(u)) * ht_1
ymm_t
ymm_one_inner
=
ymm_t
(
ymm_one
.
getIdx
());
act
<
ymm_t
>
(
ymm_u
,
ymm_u
,
act_gate_
);
act
<
ymm_t
>
(
ymm_s
,
ymm_s
,
act_cand_
);
vmulps
(
ymm_s
,
ymm_s
,
ymm_u
);
vsubps
(
ymm_u
,
ymm_one_inner
,
ymm_u
);
vmulps
(
ymm_u
,
ymm_ht_1
,
ymm_u
);
vaddps
(
ymm_u
,
ymm_s
,
ymm_u
);
vmovups
(
ptr
[
reg_ptr_ht
+
offset
],
ymm_u
);
}
offset
+=
sizeof
(
float
)
*
YMM_FLOAT_BLOCK
;
}
ret
();
}
#define DECLARE_GRU_CREATOR(name) \
class name##Creator : public JitCodeCreator<gru_attr_t> { \
public: \
/* TODO(TJ): enable more */
\
bool UseMe(const gru_attr_t& attr) const override { \
return platform::MayIUse(platform::avx) && attr.d % 8 == 0; \
} \
size_t CodeSize(const gru_attr_t& attr) const override { \
return 96 + attr.d / YMM_FLOAT_BLOCK * 96 * 2 * 8; \
} \
std::unique_ptr<GenBase> CreateJitCode( \
const gru_attr_t& attr) const override { \
return make_unique<name##JitCode>(attr, CodeSize(attr)); \
} \
}
DECLARE_GRU_CREATOR
(
GRUH1
);
DECLARE_GRU_CREATOR
(
GRUHtPart1
);
DECLARE_GRU_CREATOR
(
GRUHtPart2
);
#undef DECLARE_GRU_CREATOR
}
// namespace gen
}
// namespace jit
}
// namespace operators
}
// namespace paddle
namespace
gen
=
paddle
::
operators
::
jit
::
gen
;
REGISTER_JITKERNEL_GEN
(
gruh1
,
gen
::
GRUH1Creator
);
REGISTER_JITKERNEL_GEN
(
gruhtpart1
,
gen
::
GRUHtPart1Creator
);
REGISTER_JITKERNEL_GEN
(
gruhtpart2
,
gen
::
GRUHtPart2Creator
);
paddle/fluid/operators/jit/gen/gru.h
0 → 100644
浏览文件 @
3713d08d
/* 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/act.h"
#include "paddle/fluid/operators/jit/gen/jitcode.h"
namespace
paddle
{
namespace
operators
{
namespace
jit
{
namespace
gen
{
class
GRUJitCode
:
public
VActJitCode
{
public:
explicit
GRUJitCode
(
int
id
,
const
gru_attr_t
&
attr
,
size_t
code_size
,
void
*
code_ptr
=
nullptr
)
:
VActJitCode
(
attr
.
d
,
operand_type
::
sigmoid
/* this is bugy*/
,
code_size
,
code_ptr
),
id_
(
id
)
{
auto
typeExchange
=
[](
KernelType
type
)
->
gen
::
operand_type
{
if
(
type
==
KernelType
::
vsigmoid
)
{
return
operand_type
::
sigmoid
;
}
else
if
(
type
==
KernelType
::
vrelu
)
{
return
operand_type
::
relu
;
}
else
if
(
type
==
KernelType
::
vtanh
)
{
return
operand_type
::
tanh
;
}
else
if
(
type
==
KernelType
::
videntity
)
{
return
operand_type
::
identity
;
}
else
{
LOG
(
FATAL
)
<<
"Do not support this jit::KernelType: "
<<
type
;
}
return
operand_type
::
identity
;
};
num_
=
attr
.
d
;
act_gate_
=
typeExchange
(
attr
.
act_gate
);
act_cand_
=
typeExchange
(
attr
.
act_cand
);
this
->
genCode
();
}
const
char
*
name
()
const
override
{
std
::
string
base
=
"GRUJitCode"
;
if
(
id_
==
0
)
{
base
+=
"_H1"
;
}
else
if
(
id_
==
1
)
{
base
+=
"_HtPart1"
;
}
else
if
(
id_
==
2
)
{
base
+=
"_HtPart2"
;
}
auto
AddTypeStr
=
[
&
](
operand_type
type
)
{
switch
(
type
)
{
case
operand_type
::
relu
:
base
+=
"_Relu"
;
break
;
case
operand_type
::
exp
:
base
+=
"_Exp"
;
break
;
case
operand_type
::
sigmoid
:
base
+=
"_Sigmoid"
;
break
;
case
operand_type
::
tanh
:
base
+=
"_Tanh"
;
break
;
case
operand_type
::
identity
:
base
+=
"_Identity"
;
break
;
default:
break
;
}
};
AddTypeStr
(
act_gate_
);
AddTypeStr
(
act_cand_
);
return
base
.
c_str
();
}
void
genCode
()
override
;
protected:
int
id_
;
int
num_
;
operand_type
act_gate_
;
operand_type
act_cand_
;
reg64_t
param1
{
abi_param1
};
};
#define DECLARE_GRU_JITCODE(name, id) \
class name##JitCode : public GRUJitCode { \
public: \
explicit name##JitCode(const gru_attr_t& attr, size_t code_size, \
void* code_ptr = nullptr) \
: GRUJitCode(id, attr, code_size, code_ptr) {} \
};
DECLARE_GRU_JITCODE
(
GRUH1
,
0
);
DECLARE_GRU_JITCODE
(
GRUHtPart1
,
1
);
DECLARE_GRU_JITCODE
(
GRUHtPart2
,
2
);
#undef DECLARE_GRU_JITCODE
}
// namespace gen
}
// namespace jit
}
// namespace operators
}
// namespace paddle
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