Re: [avr-libc-dev] FP formats

[Peeter Vois]

There should be easier representation possible, I have something in my
mind, but have no time to analyze. The format in that paper does not use
3 bits, heh :) I see loss of resources there, but it is ofcourse
secondary as we can use all of them.

If you look at our fplib, then you see that for each operation with
float, the float is transferred into internal representation of floating
point from IEEE format. After the operation, it is transferred back to
IEEE format. It is indeed a overhead for transferring the formats each
time.

Best Regards,
Peeter


Ühel ilusal kenal päeval Sat, 3 Dec 2005 20:44:40 +1100
"Ron" <address@hidden> kirjutas:

> Is the floating point format used by avrgcc full IEEE? I ask because a
> paper by Connors et al proposed a simpler format for embedded systems.
> They suggest that most embedded systems have well defined domains for
> any fp computations, and so sophistry such as positive and negative
> zeroes, underflow and overflow states, and even NaNs just take up
> space and cycles. And their format allows fp comparisons by treating
> both numbers as simple signed 32-bit integers.
> 
> This is an old paper now, but I was looking at it again in the light
> of recent discussions on this list. The exponent representation could
> perhaps do with some discussion in the light of the avr's 8-bit
> registers, but otherwise it might be a candidate for implementation.
> http://www.crhc.uiuc.edu/IMPACT/ftp/conference/case-98-float.pdf.
> 
> Ron
> 
> 
> 
> 
> _______________________________________________
> AVR-libc-dev mailing list
> address@hidden
> http://lists.nongnu.org/mailman/listinfo/avr-libc-dev