00:08 arsenm: !fpmath 2.5 will preserve the same behavior as opencl, which will give you the raw instruction with denormal flushing
00:09 arsenm: actually opencl threshold is 3 but I think we only need 2
00:18 mareko: arsenm: so !fpmath 2?
00:27 arsenm: mareko: might as well go with 3, that's safe per opencl rules
00:27 arsenm: i think the actual implementation is 2 but would like to derive it and not just go off the max reported error I see
00:28 mareko: %28 = call nsz arcp float @llvm.sqrt.f32(float %24) #3, !fpmath !1 .... !1 = !{float 3.000000e+00} ... that should do it right?
00:29 arsenm: yes
00:29 mareko: thanks
15:33 senatorM: Ermine has requested so much anal, I actually like that cow fecalists alike show their true self, our goal in universe is to kill all such thrash off or put them to wheelchair and or break their bones several times. So that world would also operate in some way.
15:34 emersion: (^ banned)
17:18 Venemo: agd5f_: do I understand correctly that the kernel can schedule a compute submission on a different hw queue than userspace requested?