14:51 fdobridge: <e​sdrastarsis> Nice, Ben updated the gsp branch
18:45 fdobridge: <g​eorgeouzou> @airlied using "NIR_PASS(progress, nir, nir_opt_peephole_select, 2, true, true) " makes all tests (tessellation, clipping) happy
18:46 fdobridge: <g​eorgeouzou> A limit above 2 makes some tests fail...
18:46 fdobridge: <a​irlied> that is very narrow peephole 🙂
18:47 fdobridge: <a​irlied> @gfxstrand any ideas why peephole selection might blow stuff up later?
19:12 fdobridge: <e​sdrastarsis> @gfxstrand I think the vk_format_get_plane_count definition is duplicated in the last nvk commit
19:22 fdobridge: <g​fxstrand> Looks like. I messed up the merge. I'll fix it.
19:22 fdobridge: <g​fxstrand> Gotta wait for my internet to come back first, though. 🙄
21:08 fdobridge: <g​fxstrand> Fixed. Sorry about that.
21:08 fdobridge: <g​fxstrand> Is it blowing stuff up? I mean, it can increase pressure a bit in theory.
21:31 fdobridge: <e​sdrastarsis> np
21:47 fdobridge: <a​irlied> @gfxstrand tests start to fail that didn't fail before peephole, but also other tests don't pass unless we do it 🙂
21:49 fdobridge: <g​fxstrand> Whoops! 🙃
21:51 fdobridge: <a​irlied> setting peephole size to 2 seems to win, we could do that with a big revisit this once NAK lands 🙂
21:52 fdobridge: <a​irlied> though I suppose I should try and figure out what the backend does wrong
21:52 fdobridge: <a​irlied> or even if it is the backend
22:08 fdobridge: <m​henning> there are some known issues in nv50/codegen that I could imagine affecting things with peephole_select
22:09 fdobridge: <m​henning> one is that we never reconverge threads on gv100+, so some shaders might need peephole_select to prevent divergence for correctness
22:09 fdobridge: <m​henning> on the other hand, if you're trying to do a select on a 64-bit value, I think that will become a SPLIT, SLCT, SLCT, MERGE, and RA doesn't handle merge/select correctly
22:09 fdobridge: <m​henning> so that could be suspect
22:10 fdobridge: <m​henning> those are just off the top of my head though
22:10 fdobridge: <g​fxstrand> Yeah, I'm wholesale replacing the NIR optimization loop when NAK lands.
22:12 fdobridge: <k​arolherbst🐧🦀> mhh.. I think I have some patches around to converge those threads...
22:12 fdobridge: <k​arolherbst🐧🦀> https://gitlab.freedesktop.org/karolherbst/mesa/-/commits/nv_volta_barriers/
22:12 fdobridge: <a​irlied> also it mostly seems to affect tess eval shaders
22:12 fdobridge: <k​arolherbst🐧🦀> some of it already landed though
22:13 fdobridge: <a​irlied> if that helps narrow it down
22:13 fdobridge: <k​arolherbst🐧🦀> do those use barriers?
22:14 fdobridge: <k​arolherbst🐧🦀> but maybe tess shader need more advanced barrier stuff.. dunno
22:15 fdobridge: <a​irlied> yeah I'd guess there's barriers involved somewhere
22:17 fdobridge: <k​arolherbst🐧🦀> mhhh
22:17 fdobridge: <k​arolherbst🐧🦀> mind dumping the generated IR?
22:21 fdobridge: <a​irlied> in a meeting, will leave it here when I get a chance
22:31 fdobridge: <g​fxstrand> barriers are something codegen doesn't really do right now
22:38 fdobridge: <k​arolherbst🐧🦀> maybe I should really just clean up the patches and upstream it, but there was something funky going on with how converging threads works inside codegen
22:38 fdobridge: <k​arolherbst🐧🦀> at least for older gens
22:47 fdobridge: <m​henning> @karolherbst glancing at those patches, "nv50/ir/nir: insert joins for loops as well " isn't what we want on older gens - the break/prebreak and cont/precont stuff already reconverge so there's no need for a join/joinat (at least, this is what I observe on kepler)
22:48 fdobridge: <m​henning> I have some reconvergence stuff for older gens that I also never got to a point where I could upstream it
22:49 fdobridge: <k​arolherbst🐧🦀> yeah...
22:51 fdobridge: <m​henning> granted, I'm not sure codegen is worth the effort to fix in general - at this point I'm mostly just trying to avoid regressions
22:53 fdobridge: <a​irlied> yeah we should probably just land tess with workarounds then let @faith sort it out 😛
22:56 fdobridge: <k​arolherbst🐧🦀> yeah...
23:17 fdobridge: <g​fxstrand> I'm fine with that plan.
23:17 fdobridge: <g​fxstrand> If the only thing that's failing is barriers, oh well.