Clever Geek Handbook

Changing the width of the Mantissa in a non-IEEE floating point implementation

I have a gcc cross compiler on an 18 bit soft core target processor which has the following datatypes: Integer 18 bit long 36 bit and float 36 bit (single precision). I am currently focusing on working with floating point. Since the width is non-standard (36 bits), I have the following scheme: 27 bits for Mantissa (value), 8 bits for exponents, and 1 bit for sign. I can see that the width is defined in the float.h file. of interest to me are the following: FLT_MANT_DIG and FLT_DIG. They are defined as:

FLT_MANT_DIG 24 FLT_DIG 6

I changed them to

FLT_MANT_DIG 28 FLT_DIG 9

As per my requirements in float.h and then build the gcc compiler. But still I am getting 32 bit floating point output. Does anyone have experience implementing non-standard single precision floating point numbers and / or know a workaround?

+3


source to share


1 answer


Effective floating point math requires hardware designed to support the exact floating point formats that are used. In the absence of such hardware, subroutines that are designed around a specific floating point format will be much more efficient than subroutines that easily adapt to other formats. The GCC compiler and supplied libraries are designed to work efficiently with IEEE-754 floating point types and cannot be adapted to any others. The aforementioned headers do not allow the programmer to request a specific floating point format, but simply to notify the code which format will be used.



If you don't want 72-bit floating point types, and if the compiler's type double

will do 64-bit math in something akin to sane style, even if it long

's 36 bits instead of 32, you might be able to order things so that the values ​​are float

unboxed in four words double

, doing calculations with that, and then swapping the bits double

to get float

. Also, you can write or find 36 bit floating point libraries. I wouldn't really expect GCC or its libraries to include such a thing, since 36-bit processors are pretty rare these days.

+1


source






More articles:


All Articles