00:24 gfxstrand[d]: Yeah. I'm fine with landing it bindless only for now.
00:24 gfxstrand[d]: But we need a long term plan because I suspect that's important for perf somewhere.
00:33 gfxstrand[d]: It's gonna have to be roughly the same thing we do for textures but we have to do it early or move textures way late. It's gonna be annoying.
00:39 gfxstrand[d]: It may be that we can demote stuff later but I'm not sure. If it changes that texture data layout too much that might get tricky
01:08 mhenning[d]: gfxstrand[d]: I assume you mean "the same thing we do for cbuf handles"
01:08 mhenning[d]: but yeah
01:10 mhenning[d]: I'm also wondering about more general ways to let us use ugprs in nonuniform control flow eg. we could probably extend the live ranges of ugprs to cover the entire nonuniform section during ra
01:11 mhenning[d]: but then we need to be able to pick if it's a ugpr or not very late and I'm not sure that works for cx[] or tex handles
01:13 gfxstrand[d]: That's basically what pinning does
01:14 gfxstrand[d]: Well, pinning does two things. One is to extend live ranges. The other (and this is important for CBuf handles) is to ensure they get allocated to consecutive registers so we don't get to a CBuf reference and have to shuffle.
01:17 gfxstrand[d]: But yes, we could theoretically even write UGPRs from non-uniform control flow given enough analysis but ugh...
02:14 mhenning[d]: gfxstrand[d]: I'm aware of that, but we're slowly accumulating cases where we would benefit from that - cx[], redux, texture handles
02:14 gfxstrand[d]: Yeah
02:15 gfxstrand[d]: Pinning was always a hack
02:15 mhenning[d]: also pinning requires us to hoist the ugpr write out of the non uniform section, which we don't actually need to do
02:15 mhenning[d]: we're getting to a place where something more general might be appropriate
02:15 mhenning[d]: and I'm not totally sure what that is yet
02:15 gfxstrand[d]: mhenning[d]: Well, as long as it's not in a loop...
02:16 mhenning[d]: oh, right. ugh
02:17 gfxstrand[d]: Well, maximal reconvergance might actually save our bacon here
02:19 mhenning[d]: mhh yeah, maybe since we're reconverging continues it works out?
02:20 mhenning[d]: need to think about it more
03:14 airlied[d]: Blackwell also adds uniform texture offsets
03:14 airlied[d]: So add another one to the list, though I'm not using them
03:30 gfxstrand[d]: Yeah, let's at least make sure we know how to encode all that, even if we're not using it.
03:32 gfxstrand[d]: The thing I really hate is that we need to make these decisions when lowering from NIR to NAK. It may be that we can safely shuffle things around in NAK but I need to have a good hard look at the texture ops and their data layout to determine if that's practical. I really don't want to get ourselves in another codegen situation where we have arbitrary vectors of data which we re-arrange 3
03:32 gfxstrand[d]: different times in 3 different lowering passes and no one quite knows what a texture instruction does at any given moment.
07:14 frederikraj: I am sick for life , my neck injury can not be fixed, and it was series of killoff attempts to me arranged by people on IRC, we are going to charge you for all of those in next years, you are disgusting mindill people. I died actually probably clinically in one of those assaults and was planted a chip behind the neck, it was a series starting from 2011 bomb detonation in Russia, then the
07:14 frederikraj: bacterial poisoning and beaming and finally very impatient raw attack, where it's temptation of such looneys got me finally down , i have several issues now , i do not know if the neuralink was getting me back to life, or just they consume again all the substances, such as amniotic fluid on my back, on the other eye another depot is being felt, where as on left eye the depot got destroyed
07:14 frederikraj: or consumed, it might be both, medicine is gangsterous result fixing anyhow. It's really horrible how ILLBORN you are.
23:41 mangodev[d]: https://cdn.discordapp.com/attachments/1034184951790305330/1386129049168117891/image.png?ex=685894c6&is=68574346&hm=db14210953d6bd2e4ac4a31b15d918d3f924d442a8aa40ab8491048fb0d3f3ea&
23:41 mangodev[d]: hmm…