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
e4733224
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
e4733224
编写于
10月 20, 2017
作者:
K
Kexin Zhao
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
initial commit for float16
上级
0e31d7d7
变更
1
隐藏空白更改
内联
并排
Showing
1 changed file
with
142 addition
and
0 deletion
+142
-0
paddle/math/float16.h
paddle/math/float16.h
+142
-0
未找到文件。
paddle/math/float16.h
0 → 100644
浏览文件 @
e4733224
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
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 <cstdint>
#include <istream>
#include <ostream>
#include <cuda.h>
#ifdef __arm__
#define PADDLE_ARM_32
#endif
#ifdef __aarch64__
#define PADDLE_ARM_64
#endif
#if defined(PADDLE_ARM_32) || defined(PADDLE_ARM_64)
#define PADDLE_ARM
#endif
#if defined(__ARM_NEON) || defined(__ARM_NEON__)
#define PADDLE_NEON
#endif
#if defined(PADDLE_NEON) && defined(PADDLE_ARM_32)
#define PADDLE_NEON_32
#endif
#if defined(PADDLE_NEON) && defined(PADDLE_ARM_64)
#define PADDLE_NEON_64
#endif
#ifdef __CUDA_ARCH__ // use __CUDACC__ instead
#define PADDLE_HOSTDEVICE __host__ __device__
#if CUDA_VERSION >= 7050
#define PADDLE_CUDA_FP16
#include <cuda_fp16.h>
#endif // CUDA_VERSION >= 7050
#else
#define PADDLE_HOSTDEVICE
#endif // __CUDA_ARCH__
#if !defined(__ANDROID__) && !defined(__APPLE__) && !defined(PADDLE_ARM)
#include <immintrin.h>
#else
#ifdef __F16C__
#undef __F16C__
#endif
#endif
#define PADDLE_ALIGNED(x) __attribute__((aligned(x)))
// https://github.com/pytorch/pytorch/blob/master/torch/lib/ATen/Half.h
template
<
typename
To
,
typename
From
>
To
convert
(
From
f
)
{
return
static_cast
<
To
>
(
f
);
}
namespace
paddle
{
class
float16
;
// convert from float to half precision in round-to-nearest-even mode
float16
float2half_rn
(
float
f
);
float
half2float
(
float16
h
);
class
float16
{
public:
uint16_t
val_
;
PADDLE_HOSTDEVICE
inline
explicit
float16
()
:
x
(
0
)
{}
PADDLE_HOSTDEVICE
inline
explicit
float16
(
float
val
)
{
float16
res
=
float2half_rn
(
val
);
x
=
res
.
x
;
}
PADDLE_HOSTDEVICE
inline
explicit
float16
(
int
val
)
{
float16
res
=
cpu_float2half_rn
(
static_cast
<
float
>
(
val
));
x
=
res
.
x
;
}
PADDLE_HOSTDEVICE
inline
explicit
float16
(
double
val
)
{
float16
res
=
cpu_float2half_rn
(
static_cast
<
float
>
(
val
));
x
=
res
.
x
;
}
// Use PADDLE_ALIGNED(2) to ensure that each float16 will be allocated
// and aligned at least on a 2-byte boundary, which leads to efficient
// memory access of float16 struct.
}
PADDLE_ALIGNED
(
2
);
namespace
fp16_impl
{
// Conversion routine adapted from
// http://stackoverflow.com/questions/1659440/32-bit-to-16-bit-floating-point-conversion
Union
Bits
{
float
f
;
int32_t
si
;
uint32_t
ui
;
};
static
const
int
shift
=
13
;
static
const
int
shiftSign
=
16
;
static
const
int32_t
infN
=
0x7F800000
;
static
const
int32_t
maxN
=
0x477FE000
;
// max flt16 as flt32
static
const
int32_t
minN
=
0x38800000
;
// min flt16 normal as flt32
static
const
int32_t
sigN
=
0x80000000
;
// sign bit
static
constexpr
int32_t
infC
=
infN
>>
shift
;
static
constexpr
int32_t
nanN
=
(
infC
+
1
)
<<
shift
;
// minimum flt16 nan as float32
static
constexpr
int32_t
maxC
=
maxN
>>
shift
;
static
constexpr
int32_t
minC
=
minN
>>
shift
;
static
constexpr
int32_t
sigC
=
sigN
>>
shiftSign
;
static
const
int32_t
mulN
=
0x52000000
;
//(1 << 23) / minN
static
const
int32_t
mulC
=
0x33800000
;
// minN / (1 << (23 - shift))
static
const
int32_t
subC
=
0x003FF
;
// max flt32 subnormal downshifted
static
const
int32_t
norC
=
0x00400
;
// min flt32 normal downshifted
static
constexpr
int32_t
maxD
=
infC
-
maxC
-
1
;
static
constexpr
int32_t
minD
=
minC
-
subC
-
1
;
}
// namespace half_impl
}
// 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