02:05fdobridge: <gfxstrand> Here's a codegen run:
02:05fdobridge: <gfxstrand> `Pass: 312031, Fail: 1296, Crash: 137, Skip: 1673845, Flake: 66, Duration: 28:48`
04:09fdobridge: <gfxstrand> NAK: `Pass: 311397, Fail: 920, Crash: 1098, Skip: 1673427, Flake: 33, Duration: 1:00:1`
07:31fdobridge: <gfxstrand> NAK: `Pass: 311523, Fail: 917, Crash: 1042, Skip: 1673845, Flake: 48, Duration: 59:58`
16:43fdobridge: <gfxstrand> NAK: `Pass: 311510, Fail: 912, Crash: 1042, Skip: 1673845, Flake: 66, Duration: 59:51`
16:43fdobridge: <gfxstrand> Image atomics were totally busted and fixing them fixed... 5 tests?!?!?
16:50fdobridge: <gfxstrand> Aparently, all the image atomic tests require SPIR-V 1.3 🙃
17:05fdobridge: <gfxstrand> VK_KHR_spirv_1_4 here we go!
17:16fdobridge: <![NVK Whacker] Echo (she) 🇱🇹> Have you tested vulkanMamoryModel yet?
17:16fdobridge: <![NVK Whacker] Echo (she) 🇱🇹> Have you tested vulkanMemoryModel yet? (edited)
17:21fdobridge: <gfxstrand> No but it should mostly pass with NAK. It's mostly about not optimizing too much which means codegen will probably fail.
17:26fdobridge: <gfxstrand> Codegen with VK_KHR_spirv_1_4:
17:26fdobridge: <gfxstrand> `Pass: 313672, Fail: 1355, Crash: 161, Skip: 1672133, Flake: 54, Duration: 29:33`
17:26fdobridge: <gfxstrand> Apparently image atomics are broken in codegen. 🙄
17:29fdobridge: <gfxstrand> It's probably getting signed vs. unsigned wrong. You have to look at the atomic op itself, not the image type for that.
18:28fdobridge: <mhenning> oop, yep we do that wrong
18:39fdobridge: <mhenning> @gfxstrand Untested but https://gitlab.freedesktop.org/mhenning/mesa/-/commit/99810829b3b60765d7f4202d66cfb39551892e1c might fix codegen
19:06fdobridge: <karolherbst🐧🦀> ahh yeah... do we have some nir functions to query that information these days?
19:10fdobridge: <mhenning> I don't see signedness in nir_intrinsic_info
19:17fdobridge: <mhenning> I also don't immediately see anything in the gl cts that would test that patch
19:17fdobridge: <karolherbst🐧🦀> isn't there a function to retrieve the type directly or something now?
19:21fdobridge: <mhenning> looks like Faith has a pr https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22724
19:21fdobridge: <mhenning> but not yet
20:21fdobridge: <gfxstrand> Looks like it should work. I'm running the CTS on it now. If it works, I'll push it to the nvk/main branch and let you know.
23:12fdobridge: <gfxstrand> Fixes the bug. Feel free to slap a `Reviewed-by: Faith Ekstrand <firstname.lastname@example.org>` on it and merge upstream.
23:13fdobridge: <gfxstrand> Or, yaknow, maybe make an MR and give others a chance to review too if they want. 😅 Go ahead and put my tag on it, though. I'm gonna push it to nvk/main right now.
23:46fdobridge: <Samantas5855> is codegen like the default compiler for all drivers on mesa?
23:47fdobridge: <![NVK Whacker] Echo (she) 🇱🇹> (it's not)
23:48fdobridge: <Samantas5855> what is it then
23:48fdobridge: <![NVK Whacker] Echo (she) 🇱🇹> I think it's src/compiler 🐸
23:51fdobridge: <Ziemas> codegen is the old nouevau compiler
23:51fdobridge: <Ziemas> codegen is the old nouveau compiler (edited)
23:52fdobridge: <Samantas5855> ohhh
23:53fdobridge: <Samantas5855> how's perf btw
23:53fdobridge: <gfxstrand> F1 2015 runs fine. I didn't have an overlay up so I don't know how many FPS but it was smooth.
23:54fdobridge: <Samantas5855> hype
23:54fdobridge: <Samantas5855> so your 980m laptop is maxwell2
23:54fdobridge: <Samantas5855> but my 940m is maxwell 1?
23:55fdobridge: <Mohamexiety> NVIDIA does not have the best naming scheme for mobile to say the least. there's even a cursed 9x0m SKU that's based on _fermi_
23:55fdobridge: <Mohamexiety> (the 920m iirc?)
23:56fdobridge: <karolherbst🐧🦀> don't think too much about it
23:57fdobridge: <karolherbst🐧🦀> they got a lot better tho
23:57fdobridge: <Mohamexiety> yeah. these days are better
23:58fdobridge: <Mohamexiety> there's still a weird SKU in this lineup, but it's ultra low end so meh..
23:58fdobridge: <Mohamexiety> (RTX 2050... it's actually Ampere, not Turing)
23:58fdobridge: <Samantas5855> there's a 940m with nvenc
23:58fdobridge: <Samantas5855> and there's my 940m without nvenc
23:58fdobridge: <Samantas5855> terrible naming
23:58fdobridge: <karolherbst🐧🦀> there are like cursed models like the GT 640 (GF116, GK107 or GK208) or the GT 630 (GF108, GK107, GK208
23:59fdobridge: <Mohamexiety> yeah....
23:59fdobridge: <Samantas5855> so karol technically can we reclock pascal
23:59fdobridge: <Samantas5855> without PMU firmware?
23:59fdobridge: <karolherbst🐧🦀> no
23:59fdobridge: <karolherbst🐧🦀> I mean...
23:59fdobridge: <karolherbst🐧🦀> technically yes, but we can't change the voltage