01:33 kode54: pixelcluster: are either of your RT branches worth testing at this point on GFX11?
01:33 kode54: (I know the GFX12 one isn't)
13:44 MrCooper: pendingchaos: FYI, I bisected a few piglit cl suite regressions with rusticl on radeonsi to c26851b80b9d ("aco: increase max_const_offset_plus_one for SMEM load_global")
13:53 pendingchaos: MrCooper: what's the test names and gpu?
13:55 MrCooper: I bisected with program@execute@bswap (bswap.cl), running on Navi 14 & Bonaire
13:56 MrCooper: fails with SIGFPE in "aco::(anonymous namespace)::emit_load"
15:03 pendingchaos: MrCooper: so far, it works on my machine
15:03 pendingchaos: what llvm/clang version are you using?
15:04 MrCooper: 19.1.7, what GPU are you testing on?
15:05 pendingchaos: vega10, navi10, navi31, gfx1201
15:08 MrCooper: params.max_const_offset_plus_one is 0
15:08 MrCooper: i.e. it's a division by 0
15:09 pendingchaos: I can see now how it can fail for bonaire, but not navi14
15:13 MrCooper: it's possible that it fails only on Bonaire, I think piglit runs each test on all GPUs it finds
15:22 pendingchaos: https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/34766 likely fixes this
15:27 MrCooper: indeed it seems to, thanks
15:27 glehmann: is there a reason why it's max_const_offset_plus_one and not just max_const_offset?
15:28 pendingchaos: I guess that's what emit_load() wants most of the time
15:31 pendingchaos: I think I do prefer max_const_offset, though