#
✕ nick:alyssa
#asahi-gpu · 2024-08-08 15:47
alyssawas slightly more complicated than that but yeah, added the test and indeed we're failing for large inputs
#asahi-gpu · 2024-08-08 15:44
alyssareal
#asahi-gpu · 2024-08-08 15:43
alyssaI would like math_bruteforce to test the full 2^32 input space because otherwise I have no confidence in writing the lowering correctly
#asahi-gpu · 2024-08-08 15:42
alyssa:=/
#asahi-gpu · 2024-08-08 15:42
alyssareciprocal is suspiciously absent from math_bruteforce
#asahi-gpu · 2024-08-08 15:41
alyssaah. but does the CL CTS test it.
#asahi-gpu · 2024-08-08 15:40
alyssaI guess libclc doesn't do divide?
#asahi-gpu · 2024-08-08 15:38
alyssaalthough i dont actually see where we call libclc frcp
#asahi-gpu · 2024-08-08 15:37
alyssaso that seems.. bad
#asahi-gpu · 2024-08-08 15:36
alyssabut you need exact frcp! in the backend *anyway* for correct idiv
#asahi-gpu · 2024-08-08 15:36
alyssakarolherbst: ook. so what I realized is real funny is that CL CTS is going to hit libclc for frcp
#asahi-gpu · 2024-08-08 15:05
alyssamaybe not
#asahi-gpu · 2024-08-08 15:04
alyssaI'm also wondering if there's a way to make lower idiv tolerate the error, for faster overall
#asahi-gpu · 2024-08-08 15:04
alyssaneeds to be in-tree for idiv lowering anyhow
#asahi-gpu · 2024-08-08 15:02
alyssamy aborted math degree came in handy!
#asahi-gpu · 2024-08-08 15:02
alyssa:P
#asahi-gpu · 2024-08-08 15:01
alyssagot it in 2
#asahi-gpu · 2024-08-08 15:01
alyssaok
#asahi-gpu · 2024-08-08 15:01
alyssalolz
#asahi-gpu · 2024-08-08 15:01
alyssa----wait is that a newton raphson step
#asahi-gpu · 2024-08-08 15:01
alyssarcp := fma(fma(-input, rcp, 1.0), rcp, rcp)
#asahi-gpu · 2024-08-08 15:01
alyssaunder a certain condition it'll adjust the output as
#asahi-gpu · 2024-08-08 15:00
alyssathe other part though I don't quite understand yet
#asahi-gpu · 2024-08-08 15:00
alyssa*denominators
#asahi-gpu · 2024-08-08 15:00
alyssaone part is a scaling like scale_fdiv, to reduce the domain for large denomintaors
#asahi-gpu · 2024-08-08 15:00
alyssathere are two parts to the lowering
#asahi-gpu · 2024-08-08 15:00
alyssathis is... subtle
#asahi-gpu · 2024-08-08 14:50
alyssamore complex, gimme a few mins
#asahi-gpu · 2024-08-08 14:48
alyssacute.
#asahi-gpu · 2024-08-08 14:48
alyssakarolherbst: frcp with -no-fast-math has lowering indeed. ughhh.
#asahi-gpu · 2024-08-08 14:48
alyssaoh, jeez, ugh!
#asahi-gpu · 2024-08-08 14:45
alyssabut AGX gives 32A60001
#asahi-gpu · 2024-08-08 14:45
alyssashould be 32A60000
#asahi-gpu · 2024-08-08 14:44
alyssafrcp(0x4C4565C8)
#asahi-gpu · 2024-08-08 14:44
alyssayeah..
#asahi-gpu · 2024-08-08 14:42
alyssahm no that should be ok
#asahi-gpu · 2024-08-08 14:41
alyssaer wait
#asahi-gpu · 2024-08-08 14:40
alyssayes
#asahi-gpu · 2024-08-08 14:40
alyssaalthough Metal's docs claim that 1.0 / x is correctly rounded
#asahi-gpu · 2024-08-08 14:39
alyssaI mean I'm just plumbing through the hardware frcp here
#asahi-gpu · 2024-08-08 14:39
alyssaplausible
#asahi-gpu · 2024-08-08 14:37
alyssaIDK if this is an AGX backend bug or a nir_lower_idiv bug, strictly
#asahi-gpu · 2024-08-08 14:37
alyssaand that off-by-1 propagates to the totally wrong final rsult
#asahi-gpu · 2024-08-08 14:36
alyssaagx has an off-by-1 compared to my cpu
#asahi-gpu · 2024-08-08 14:36
alyssaint((1.0 / float(b)) * (4294966784.0))
#asahi-gpu · 2024-08-08 14:36
alyssathe problem is the value of `rcp` at the start
#asahi-gpu · 2024-08-08 14:36
alyssakarolherbst: ok, this is interesting
#asahi-gpu · 2024-08-08 14:31
alyssapoking at idiv
#asahi-gpu · 2024-08-08 14:16
alyssabuilds rusticl
#asahi-gpu · 2024-08-08 13:36
alyssarebuilds piglit

plus ›