29 Jan
2011
29 Jan
'11
12:46 p.m.
On Sat, Jan 29, 2011 at 12:07:12PM +0100, Philipp Überbacher wrote:
Everything floating point 'suffers from the same problem'.
Even addition. (1/3.0 + 1/3.0 + 1/3.0 == 1.0) doesn't have
to result in true.
It's a single CPU cycle on most processors, and on some (e.g. ARM)
it can even be combined with other instructions.
Ciao,
--
FA
There are three of them, and Alleline.
Took me a while to figure out that libm is part of
glibc :)
It isn't. But it lives in /lib, not /usr/lib
k =
(int)(log2(x) + 1e-6)
log2() suffers from the same problem? I somewhat dislike the idea of
adding a constant. int m, k;
for (k = 0, m = 1; m < x; k++, m <<= 1);
which will round up if x is not a power of 2.
Neat. I thought about it myself yesterday but my ideas weren't exactly
brilliant. One idea was to divide by 2, the other to use a small
lookup table for powers of 2. I don't really know about efficiency, but
I guess bit shifting is as efficient as it gets?