This is the mail archive of the
mailing list for the GCC project.
Re: numerical instability and estimate-probability
- From: Linus Torvalds <torvalds at transmeta dot com>
- To: jh at suse dot cz, gcc at gcc dot gnu dot org
- Date: Fri, 16 Nov 2001 09:22:12 -0800
- Subject: Re: numerical instability and estimate-probability
- Newsgroups: linux.egcs
- References: <Pine.LNX.email@example.com>
In article <20011116173148.B26573@atrey.karlin.mff.cuni.cz> you write:
>Actualy in this case we are speaking about frequencies (for probabilities
>it is just OK to use 0 ... 10000 integers).
>The frequencies can get very huge, so only sane workaround wihtout FP
>is IMO only inventing the mantisa/exponent at my own :(
Ehh.. Why do you need a mantissa at all?
You're kidding yourself if you think the frequencies are exact, so why
not just do frequencies with logarithmic operations, and always
represent them as powers-of-two - as a simple integer.
I assume that you never do addition of frequencies anyway (what would
that mean?), so you really only ever multiply frequencies with each
In which case a logarithmic expression is by far the most convenient, as
it will turn your existing multiplications into just additions, speeding
the thing up in the process.
And if you select a power-of-two, it's easy enough to do approximations
of the log2() operation for any remaining cases where you _do_ want to
There is another major reason for using a log2 representation. If you
do optimizations, you really don't care whether something is called 100
times or 105 times. You really _only_ care about "orders of magnitude"
differences. Ie you really fundamentally should only care about the
logarithm of the frequency _anyway_.
Let's hear it for orders-of-magnitudes.