Put in new opencl kernel for testing. #16
Conversation
|
I remember a performance regression on Intel hardware but gains on AMD/Nvidia so I think that's expected. Will test soon. |
|
In that case if your testing (when you get to it) goes well I may keep the old opencl around and just rename it in a slightly less bad way (maybe in a subdirectory of something). |
|
With an Nvidia 750 ti, my testing results in what the comment at the top of the optimized kernel says (~440Mh/s to ~520Mh/s). AMD was pretty inconclusive. It's not slower but not much faster either. I did try adjusting the intensity of both kernels but performance was still pretty close. I don't have a wide range of AMD cards to test with so since I don't see any regression, I think this can go in. OK by me since #17 was merged and I successfully solo mined a few blocks on testnet. |
Minor modifications of the return format was needed to work with gominer. Keep old blake256 kernel as it performs better on intel hardware but not others. The improvement is mainly for nvidia cards. AMD cards seem to have little difference. For testing with Nvidia 750 ti it went from ~440Mh/s to ~520Mh/s.
We need to do some more testing (not just in benchmark more). @jolan maybe at some point when you get a chance.
For the new kernel:
For the old one;
In benchmark mode I'm getting worse performance with the new kernel but I didn't try to tune the intensity and I'm really not on production useful hardware.