> Does anybody know why truncated IEEE floating-point numbers are used in
> the predition module of MPEG2 AAC decoder? Will it be right if I do not
> convert these numbers to floating-point ones?
> All tips are welcomed.
I'm not familiar with the actual MPEG2 AAC implementation, but one thing
to watch out for in general with floating-point DSP implementations is
that iterative computations that should seemingly go to zero (for
example, x = x*0.1F, computed over and over) might instead go into a
"limit cycle", where x takes on a repeating series of very small floating
point numbers near zero. On some machines, this triggers slow "denormalized"
arthmetic processing, where a software trap or long microcode program is
run for each operation.
One way to avoid limit cycles is to do truncation once the floating-point
values get too small, which might be the reason behind the AAC code you
mention ... just a guess ...
-------------------------------------------------------------------------
John Lazzaro -- Research Specialist -- CS Division -- EECS -- UC Berkeley
lazzaro [at] cs [dot] berkeley [dot] edu www.cs.berkeley.edu/~lazzaro
-------------------------------------------------------------------------
This archive was generated by hypermail 2b29 : Mon Jan 28 2002 - 12:11:07 EST