Pekka Jääskeläinen
Pekka Jääskeläinen
Just shooting in the dark: Could this be another of those additional '_' in the symbol name prefix differences we've seen with MacOSs. If you look at the parallel.so using...
@maleadt After those cases are solid again, maybe we could add a core set of Julia tests and run them with the PoCL CIs as well, to keep them working?...
Thanks for the report. We'll test with the new compiler that has a lot of bugs fixed, but quickly looking at your CL: [Early exits with barriers in the remaining...
We have a new version of the kernel compiler which is very much reworked. We will upstream it after it's finished and we get an academic publication of it accepted....
@franz it is still true. The main difference with the new kernel compiler is that the work-group functions usually get vectorized better (more often run WIs in multiple vector lanes)...
> @pjaaskel could this be fallbacking CBS failing to execute some Julia test cases? @tneval, most likely yes. > The failing tests are the same as mentioned by @maleadt here...
@maleadt what do you think: Is the fallback behavior for 'unreachables' better than the brutal one we have now? I haven't yet come up with a solution for unreachables that...
Yes, the CUDA driver is a bit undermaintained, still experimental, and improved with best effort and community contributions. Patches more than welcome!
@franz I think you looked at this a bit before, can you give pointers for best approaches here?
If you can, please create a small reproducer and we can test it with the new compiler branch as well.