Vdsp(bf561)中的浮点运算（11）：fract16与float的转换

快乐虾

http://blog.csdn.net/lights_joy/

lights@hb165.com

 

本文适用于

ADSP-BF561

Visual DSP++ 5.0 (update 6)

Vdsp dual processor simulate

 

欢迎转载，但请保留作者信息

 

 

vdsp提供了两个函数用以实现fract16与float之间的相互转换：

    fract16 float_to_fr16 (float _x);

    float   fr16_to_float (fract16 _x);

看看这两个转换函数到底做了什么。

1.1    float_to_fr16

这个函数的原始代码在Blackfin/lib/src/libc/runtime/fl2fr.asm中，先看看它的注释：

/***************************************************************************

*

* Function:  FLOAT_TO_FR16 -- Convert a floating-point value to a fract16

*

* Synopsis:

*

*       #include <fract2float_conv.h>

*       fract16 float_to_fr16(float x);

*

* Description:

*

*       The float_to_fr16 function converts a single precision, 32-bit IEEE

*       value into a fract16 number in 1.15 notation. Floating-point values

*       that cannot be converted to fract15 notation are handled as follows:

*

*           Return 0x7fffffff if x >= 1.0 or  NaN or +Inf

*           Return 0x80000000 if x < -1.0 or -NaN or -Inf

*           Return 0          if fabs(x) < 3.0517578125e-5

*

*       (Note that the IEEE single precision, 32-bit, representation

*       contains 24 bits of precision, made up of a hidden bit and 23

*       bits of mantissa, and thus some precision may be lost by converting

*       a float to a fract16).

*

* Algorithm:

*

*       The traditional algorithm to convert a floating-point value to 1.15

*       fractional notation is:

*

*           (fract16) (x * 32768.0)

*

*       However on Blackfin, floating-point multiplication is relatively

*       slow is emulated in software, and this basic algorithm does not

*       handle out of range results.

*

*       This implementation is based on the support routine that converts

*       a float to fract32, and then converts the fract32 into a fract16

*       by performing an arithmetic right shift by 16 bits. (It is possible

*       to avoid the shift by coding the function to "multiply" the input

*       input argument by 2^15 (rather than 2^31) but this approach can lead

*       to the loss of 1-bit precision when handling negative inputs).

*

*       The following is a C implementation of this function and is about

*       a third slower:

 

            #include <fract2float_conv.h>

 

            extern fract16

            float_to_fr16(float x)

            {

 

                int temp;

                fract32 result;

 

                temp = *(int *)(&x);

 

                if ((temp & 0x7f800000) >= 0x3f800000) {

                    result = 0x7fffffff;

                    if (temp < 0)

                        result = 0x80000000;

                } else {

                    temp = temp + 0x0f800000;

                    result = *(float *)(&temp);

                }

 

                return (result >> 16);

 

            }

*

*       WARNING: This algorithm assumes that the floating-point number

*       representation is conformant with IEEE.

*

* Cycle Counts:

*

*       31 cycles when the result is within range

*       30 cycles when the result is out of range

*       28 cycles when the input is 0.0

*

*       These cycle counts were measured using the BF532 cycle accurate

*       simulator and include the overheads involved in calling the function

*       as well as the costs associated with argument passing.

*

* Code Size:

*

*       76 bytes

*

* Registers Used:

*

*       R0 - the input argument

*       R1 - various constants

*       R2 - the exponent of the input argument or a shift amount

*       R3 - the mantissa of the input argument

*

* (c) Copyright 2006 Analog Devices, Inc.  All rights reserved.

*     $Revision: 1.3 $

*

***************************************************************************/

这段注释和原始代码比VDSP提供的文档清晰多了，从这里可以知道其转换过程是先将其转换为fract32类型，在最后将转换结果右移16位得到fract16类型，且由于float类型有23位的尾数，而fract16则只有16位，因此不可避免地会引起精度丢失。

l         当输入值>=1、为nan或者为inf时

返回0x7fff，也就是fract16能表示的最大值0.999969482421875。

此时需要29个cycle

l         当输入值<-1或者为-inf时

返回0x8000，也就是fract16能表示的最小值-1。

此时需要29个cycle。

l         范围内的值

此时需要30个cycle。

1.2    fr16_to_float

这个转换由于是从小精度的数转换为大精度的数，过程比较简单，也没有精度丢失的问题，其实现代码在Blackfin/lib/src/libc/runtime/fr2fl.asm中，看其注释：

/***************************************************************************

*

* Function:  FR16_TO_FLOAT -- Convert a fract16 to a floating-point value

*

* Synopsis:

*

*       #include <fract2float_conv.h>

*       float fr16_to_float(fract16 x);

*

* Description:

*

*       The fr16_to_float converts a fixed-point, 16-bit fractional number

*       in 1.15 notation into a single precision, 32-bit IEEE floating-point

*       value; no precision is lost during the conversion.

*

* Algorithm:

*

*       The traditional algorithm to convert a 1.15 fractional numbers to

*       floating-point value is:

*

*           (float)(x) / 32768.0

*

*       However on Blackfin, floating-point division is relatively slow,

*       and one can alternatively adapt the algorithm for converting from

*       a short int to a float and then subtracting 15 from the exponent

*       to simulate a division by 32768.0.

*

*       The following is a slower C implementation of this function:

 

            #include <fract2float_conv.h>

 

            extern float

            fr16_to_float(fract16 x)

            {

 

                float result = fabsf(x);

                int *presult = (int *)(&result);

 

                if (result != 0.0) {

                    *presult = *ptemp - 0x07800000;

                    if (x < 0)

                       result = -result;

                }

 

                return result;

 

            }

*

*       WARNING: This algorithm assumes that the floating-point number

*       representation is conformant with IEEE.

*

* Cycle Counts:

*

*       22 cycles when the input is 0

*       25 cycles for all other input

*

*       These cycle counts were measured using the BF532 cycle accurate

*       simulator and include the overheads involved in calling the function

*       as well as the costs associated with argument passing.

*

* Code Size:

*

*       38 bytes

*

* Registers Used:

*

*       R0 - the input argument and result

*       R1 - various

*       R2 - various

*

* (c) Copyright 2006 Analog Devices, Inc.  All rights reserved.

*     $Revision: 1.2 $

*

***************************************************************************/

这个转换过程需要24个cycle。

1.3    fract16常量赋值

vdsp虽然没有将fract16当成内置类型，但是对于常量，编译器还是网开一面，使用r16或者r32后缀，编译器会自动将这个常量正确转换为fract16类型，如

fract16 r = 0.2r16;

编译器自动计算r的值为0x1999。

 

1.4    不幸的事件

由于fract16不是内置类型，编译器将不会自动完成两种类型之间的转换，如果不小心写上：

       float r;

       fract16 r1 = 0.2r16;

       r = r1;

那么很不幸，r的值不是期望的0.2，而是6553！

Fract的其它运算也一样，务必通过函数调用来完成。

 

 

2       参考资料

Vdsp(bf561)中的浮点运算（10）：fract16类型表示(2009-8-17)

Vdsp(bf561)中的浮点运算（9）：long double和float的比较(2009-8-14)

Vdsp(bf561)中的浮点运算（8）：float除法运算(2009-8-14)

Vdsp(bf561)中的浮点运算（7）：float乘法运算(2009-8-13)

Vdsp(bf561)中的浮点运算（6）：float加减运算(2009-8-13)

Vdsp(bf561)中的浮点运算（5）：float类型表示总结(2009-8-12)

Vdsp(bf561)中的浮点运算（4）：FLT_MAX(2009-8-12)

Vdsp(bf561)中的浮点运算（3）：FLT_MIN(2008-12-19)

Vdsp(bf561)中的浮点运算（2）：float的疑问(2008-12-18)

Vdsp(bf561)中的浮点运算（1）：文档的说法(2008-12-16)