Oh BTW, drivers that want to implement pipe_screen::finalize_nir to improve cached shader compilation can't use any of those lowering passes, because then things blow up. The only _optional_ lowering you can do in st/mesa is clamp_color and force_persample_in_shader, which covers iris at least (radeonsi doesn't need any). If you need more, you either need to invest a significant amount of time to make driver-specific NIR work with the lowering passes, or give up on good shader cache efficiency.
Marek On Fri, Oct 18, 2019 at 4:17 PM Marek Olšák <mar...@gmail.com> wrote: > Note that none of the lowering passes work if I enable them on radeonsi. > So if you think about enabling them for your driver, I guess you have some > work to do in the driver as well. > > On the other hand, having multiple shader variants in st/mesa is a > performance disadvantage. The lowering should be done at the machine-level > assembly code, so that you don't have to completely recompile from a > higher-level IR. > > Marek > > On Fri, Oct 18, 2019 at 4:08 PM Marek Olšák <mar...@gmail.com> wrote: > >> On Fri, Oct 18, 2019 at 9:07 AM Ilia Mirkin <imir...@alum.mit.edu> wrote: >> >>> On Fri, Oct 18, 2019 at 9:04 AM Erik Faye-Lund >>> <erik.faye-l...@collabora.com> wrote: >>> > >>> > On Fri, 2019-10-18 at 08:57 -0400, Ilia Mirkin wrote: >>> > > On Fri, Oct 18, 2019 at 8:51 AM Erik Faye-Lund >>> > > <erik.faye-l...@collabora.com> wrote: >>> > > > On Thu, 2019-10-17 at 22:24 -0400, Ilia Mirkin wrote: >>> > > > > In the meanwhile (unless you plan on taking up Jason's >>> > > > > suggestion), >>> > > > > might I recommend some assert's for the unhandled cases so that >>> > > > > there >>> > > > > are no surprises? >>> > > > >>> > > > Good idea. I sent a MR for it here: >>> > > > >>> > > > https://gitlab.freedesktop.org/mesa/mesa/merge_requests/2380 >>> > > >>> > > Not sure that approach works, esp with SSO. >>> > >>> > If so, wouldn't that already be a problem with the existing >>> > lower_depth_clamp-stuff, then? I mean, I just lifted the logic from >>> > there... >>> >>> Yeah, that looks bogus. I'm moderately sure that checking >>> "st->vp/gp/tep" at LinkShader time is wrong. Marek can probably >>> confirm. >>> >> >> Don't read any context states at link time. It's wrong. >> >> Marek >> >
_______________________________________________ mesa-dev mailing list mesa-dev@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/mesa-dev
