Steve wrote:
Since the sprintf function is slowing down on the log10 due to
floating point arithmetic, I wonder if a further optimization could be
made by rewriting the log10 function in assembler to take advantage of
the floating point registers?
I found that writing an integer-based log10 function sped up my itoa
function considerably. I don't think you could do anything faster in
ASM, especially since floating point math is not strictly required anyway.
Anyway, I revised my implementation today to not *need* a log10 function
by writing the ascii value into the *end* of the buffer, right-to-left,
least-significant to most-significant, and then it returns a pointer to
the beginning of the ascii string it creates. That way, I didn't need to
calculate the length of the string up front (or ever), and it cut my
runtime in half.
I did, however, find another special case (in addition to zero), and
that is INT_MIN. If you can figure out why INT_MIN is a special case,
you get bonus Plug Karma points. The answer lies in the result of the
following C expression:
abs( INT_MIN ); // What does this return? :)
(Note: you'll need to include <limits.h> and <stdlib.h>)
I found this special case by running 2^32 test cases for every 32 bit
integer, and 48 minutes later, it reported that it passes (ie, returns
the same as sprintf()) for all 2^32 values. Also, my code performed
about 3.3 million itoa()'s per second. Not too shabby. :)
I wonder how Java would do, and Qt's QString::number() for that matter.
--Dave
/*
PLUG: http://plug.org, #utah on irc.freenode.net
Unsubscribe: http://plug.org/mailman/options/plug
Don't fear the penguin.
*/
