On Monday, July 18, 2016 10:58:31 PM PDT Ben Widawsky wrote: > On Mon, Jul 18, 2016 at 07:08:46PM -0700, Kenneth Graunke wrote: > > Fixes a 10-20% performance regression in OglCSDof caused by commit > > 5a8c89038abab0184ea72664ab390ec6ca58b4d6, which made images (in the > > image load/store sense) use BDW_MOCS_PTE instead of BDW_MOCS_WB. > > > > This seems sketchy, as the default PTE value is supposed to be > > WB LLC eLLC, which is the same as our MOCS WB setting. It's only > > supposed to change when using a surface for display, which won't > > ever happen for images. Something may be wrong in the kernel... > > Which platform was the regression on? Broadwell has some weirdness if you > disable PPGTT which might cause PTE MOCS to fall back to UC. I think at least > aliasing PPGTT has been enabled since pretty early for Broadwell. If there is > no > ppgtt overrides here, then I'm not sure what would be going on - it sound > sketchy to me too.
My Broadwell GT2 laptop running drm-intel-nightly from 2016y-05m-27d-12h-32m-45s (4.6.0 based), and Mark's Broadwell GT3e boxes in Jenkins. I tried to verify the kernel PTE settings, but I got totally lost by the _PAGE_PAT stuff. Earlier platforms I can follow easily... This patch fixes the observed problem, but I still would like to understand why the PTE isn't good enough...for renderbuffers, we use the "Use PTE" setting...and could be missing out on some performance there...
signature.asc
Description: This is a digitally signed message part.
_______________________________________________ mesa-dev mailing list mesa-dev@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/mesa-dev