[music-dsp] 32-bit float error questions

[Mike Marchywka]

> Yes, but setting the denormalization mode can make a huge 
> difference.  I've
> been shocked at how slow calculations can be when the results 
> converge on
> zero.  They sure picked a common number to make unusable!

iirc, there are cases when zero operands produce results more quickly
than non-zero operands. The problem is handling denormals accurately.
Once a denormal becomes zero, you probably have no more denormals to
handle ( slowly through many iterations ) in a given calculation but rather have
more subsequent zeroes ( faster than normal and probably fewer iterations to get to zero). 

http://cedar.intel.com/cgi-bin/ids.dll/content/content.jsp?cntKey=Generic+Editorial%3a%3asolveit_p4_x87SSE&cntType=IDS_EDITORIAL&catCode=CUY

Of course, if you result is supposed to be a denormal and not zero, you may have some problems. 

I remember having some odd timing results when I tried timing something
with random data that didn't match my real data values too closely.
I think the problem was related to this (undocumented) value dependent timing
( rather than exception/denormal handling). 

> -----Original Message-----
> From: Joshua Scholar [mailto:joshscholar at yahoo.com]
> Sent: Monday, July 28, 2003 12:21 AM
> To: music-dsp at aulos.calarts.edu
> Subject: Re: [music-dsp] 32-bit float error questions
> 
> 
> >
> > Since full-resolution FPU asm loops seem "fast enough" so 
> far, I don't
> > see any reason to intentionally degrade accuracy in hopes 
> of gaining a
> > little more speed. But if an algorithm were so complex that full
> > accuracy does NOT run fast enough, guess 32 bit float would 
> be better
> > than non-realtime performance (GRIN).
> >
> > James Chandler Jr.
> >
> >
> 
> Yes, but setting the denormalization mode can make a huge 
> difference.  I've
> been shocked at how slow calculations can be when the results 
> converge on
> zero.  They sure picked a common number to make unusable!
> 
> 
> 
> dupswapdrop -- the music-dsp mailing list and website: 
> subscription info, FAQ, source code archive, list archive, 
> book reviews, dsp links http://shoko.calarts.edu/musicdsp/ 
> http://ceait.calarts.edu/mailman/listinfo/music-dsp
>