Messages in this thread | | | Date | Fri, 18 Apr 2008 17:20:16 -0700 (PDT) | From | dean gaudet <> | Subject | Re: Alternative implementation of the generic __ffs |
| |
On Fri, 18 Apr 2008, Harvey Harrison wrote:
> On Fri, 2008-04-18 at 16:46 -0700, dean gaudet wrote: > > On Fri, 18 Apr 2008, Alexander van Heukelum wrote: > > > > > On Mon, Apr 07, 2008 at 12:25:50PM +0200, Alexander van Heukelum wrote: > > > > On Sun, 6 Apr 2008 13:22:58 -0700 (PDT), "dean gaudet" <dean@arctic.org> said: > > > > > On Sun, 6 Apr 2008, Alexander van Heukelum wrote: > > > > > > The current generic implementation of ffz is O(lg(n)) already > > > > > > > > > > it's O(lg(n)) time... the operations all depend on each other. > > > > > > > > > > the implementation i pointed to is O(lg(n)) code space... and the time > > > > > depends on how parallel the machine is, they're not dependent on each > > > > > other. > > > > > > > > Indeed. The worst dependencies are in the sum of all the partial > > > > results in this implementation. And addition is associative, so > > > > partial results can be written as ((a+b)+(c+d))+(e+f). Assuming > > > > perfect parallel execution this would lead to O(ln(ln(n))). Good. > > > > > > Hello all, > > > > > > I've implemented ffs (find first set bit) like it is shown > > > in http://www.hackersdelight.org/ (see revisions, page 21). > > > > sweet! thanks for doing this. > > > > > > > static ATTR int __ffs32_new(unsigned int value) > > > { > > > int x0, x1, x2, x3, x4; > > > > > > value &= -value; > > > x0 = (value & 0x55555555) ? 0 : 1; > > > x1 = (value & 0x33333333) ? 0 : 2; > > > x2 = (value & 0x0f0f0f0f) ? 0 : 4; > > > x3 = (value & 0x00ff00ff) ? 0 : 8; > > > x4 = (value & 0x0000ffff) ? 0 : 16; > > How about: > u8 x; > > value &= -value; > x = (value & 0x55555555) ? 0 : 1; > x |= (value & 0x33333333) ? 0 : 2; > x |= (value & 0x0f0f0f0f) ? 0 : 4; > x |= (value & 0x00ff00ff) ? 0 : 8; > x |= (value & 0x0000ffff) ? 0 : 16;
any reasonable compiler should figure out the two are the same... but i really prefer spelling out the lack of dependencies of the computations by breaking it out per-bit.
-dean
| |