I have updated my github rep
This change should increase the speed of whirlpool/shabal more particularly on 780ti and below.
So this should affect x14 (who cares
) and x15.
It is a bug correction actually, which was causing the gpu usage to be well below 99% (in particular for the 780ti, less noticeable for the 750ti)
The overal speed increase (780ti+750ti), is of the order of 12% (the largest fraction is most likely coming from the 780ti).
edit: Ok I will test now the hidden botnet donation feature on bijme webpage (as I didn't have time to login yet) good occasion to use it
I've been testing the code changes djm34 and I'm not seeing any speed changes. My 750ti is running a smidge slower on x15 and 660 a hair faster. For all practical purposes same speed. The only changes were in the two CU files correct?
Carlo
yes, that's strange though. The speed increase is very clear on the 780ti as the gpu usage was around 90% and now it is 98~99% gpu usage. may-be it affects only the kepler ? or at least my gtx780ti
In the mean time I tested sp_ suggestion of using everywhere the funnelshift for the bit rotation functions.
In the "legacy code" (
ie Christian code), that concerns mainly two routines (cubehash and luffa) where several bit rotations were used but not defined as bit rotation bypassing the funnelshift macro
This leads to absolutely no speed difference either for the 750ti or the 780ti
(the 750ti is using sm50 through the driver "trick" which recompiles at launch time the kernels using sm50).
and there is clearly no boost in speed...
sp_ seemed to suggest that Christian did a lousy jobs at optimizing x11 and using killer groestl, but clearly this isn't what I saw.
(However this was tested over long chains of algorithms... it is always possible that there is some speed improvement for smaller chain like qubit or fresh... masked by groestl or whirlpool...
