Re: further improved phatk OpenCL kernel (> 2% increase) for Phoenix - 2011-07-06
Thanks, best version yet 
Still not reached the 40 MHash/sec the wiki says my card could do
Did you notice that Ma(x, y, z) is defined exactly the same now whether BFI_INT is enabled or not? Seems more elegant to me if moved out of the #ifdef. Also I tried to replace some #define's with functions, guessing that it would make it easier for a somewhat smart compiler to find repeatedly used terms and put them into registers. No performance improvement, but didn't hurt it either.
Also, OpenCL has a builtin Ch function, not faster for me but maybe for someone else:
#define Ch(x, y, z) bitselect(z, y, x)
Still not reached the 40 MHash/sec the wiki says my card could do
Did you notice that Ma(x, y, z) is defined exactly the same now whether BFI_INT is enabled or not? Seems more elegant to me if moved out of the #ifdef. Also I tried to replace some #define's with functions, guessing that it would make it easier for a somewhat smart compiler to find repeatedly used terms and put them into registers. No performance improvement, but didn't hurt it either.
Also, OpenCL has a builtin Ch function, not faster for me but maybe for someone else:
#define Ch(x, y, z) bitselect(z, y, x)
Thank YOU
another nice hint, even if it not boosts, the code gets cleaner. I'm not sure about the #define as functions. Could you post an example? My problem is, that most variables are defined and declared inside the kernel function. So a function for sharound() for example needs Vals[] and others as passed parameters (copy or pointer).Dia