06:24airlied[d]: oh now I noticed coopmat has a feature bit to say whether you can be bothered doing robust or not
06:24airlied[d]: so if I don't advertise that bit, I can probably bail on a bunch of stuff
10:26karolherbst[d]: does robust matrix accesses account for elements or the entire matrix?
12:07snowycoder[d]: If somebody has free time to review, I hope https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/37108 is ready!
12:07snowycoder[d]: (pinging mhenning[d] since she previously reviewed)
12:13jennymaiden: I actually responded to anything about the formula, normally you could solve fractional equation yet more precisely so it has maximum capacity, through division by reducing the equation, we call such things not quadratic aka ruutvõrrand which takes square root over two possible solutions which where one is two equals and on unique, but we call this murdvõrrand, where you simplify and reduce
12:13jennymaiden: the formula. and then you get very accurate result precalculated , i did not give fully accurate, but i got to run again.
14:54gfxstrand[d]: mhenning[d]: I think I know what's going on with those failing interpolation tests. They're setting a sample location that isn't (0.5, 0.5) for the non-multisampled case and then using a `sample` qualifier in the shader and expecting things to get interpolated there.
14:54gfxstrand[d]: I honestly have no idea what the hardware is doing there
14:55gfxstrand[d]: Unfortunately, the tests don't log the actual coords so we're going to have to hack up the tests
15:26karolherbst[d]: ~~use printf~~
15:26snowycoder[d]: hold on
15:26snowycoder[d]: you can with my hacks!
15:26snowycoder[d]: (in nir_lowering)
15:55mhenning[d]: gfxstrand[d]: Right, I figured that much out
15:55mhenning[d]: what's not clear to me is what we could change about our setup code to make that work
15:56gfxstrand[d]: Yeah, not sure
15:56gfxstrand[d]: I didn't see anything obviously special-casing MSAA count 1
15:57mhenning[d]: Yeah, that's what I was looking for too
15:58gfxstrand: We used to but I got rid of most of that for shader object
16:04gfxstrand: I wonder if it's CENTROID_PER_FRAGMENT that's messing us up
16:09gfxstrand[d]: GL had 1x MSAA so there's got to be a way to make it work
16:10gfxstrand[d]: :frog_clown:
16:10HdkR: Everyone loves 1x MSAA.
16:10gfxstrand[d]: It's the best kind of anti-aliasing!
16:11zmike[d]: I still don't support it :stressheadache:
16:11gfxstrand[d]: You can have any number of samples, as long as the number is 1!
16:11gfxstrand[d]: The most evil implementation of multisample render to single sample: We support it but only with samples = 1
16:12mhenning[d]: I hacked it to always use SET_HYBRID_ANTI_ALIAS_CONTROL_CENTROID_PER_PASS and that doesn't fix the test
16:12gfxstrand[d]: 😭
16:17gfxstrand[d]: We could always change `load_barycentric_sample` to `load_barycentric_at_sample(sample_id)`
16:17gfxstrand[d]: makes sample interpolation a bit more expensive but puts us in control
16:23gfxstrand[d]: Okay, I've got patches. I hate them.
16:25HdkR: Anyone that turns on AA knows they are going to have pain anyway right? :)
16:26mohamexiety[d]: MSAA is so old school. give me all the temporal effects
16:30HdkR: If I'm not seeing six frames in to the past then what am I even doing with my life.
16:30mhenning[d]: gfxstrand[d]: okay, I might continue hacking at it and try to see what proprietary does
16:31mohamexiety[d]: I am kind of confused though. shouldn't MSAA 1x be basically MSAA off? or is actually enabling MSAA and only doing 1 sample the correct behavior?
16:31HdkR: Sadly it's not the same as off.
16:31mohamexiety[d]: 🐸
16:31mohamexiety[d]: but... why
16:32gfxstrand[d]: Thanks! I hate it. https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/38008
16:33HdkR: :D
16:33mhenning[d]: mohamexiety[d]: We're currently having issues with VK_EXT_sample_locations + 1 sample, so that extension has a little more context https://docs.vulkan.org/refpages/latest/refpages/source/VK_EXT_sample_locations.html
16:33gfxstrand[d]: Most drivers optimize `interpolateAtSampleId(gl_SampleId)` to `interpolateAtSample()`. On NVIDIA, we "optimize" the other way!
16:34themaister[d]: mhenning[d]: isn't that just called vkCmdSetViewports with fractional offsets? 🐸
16:34gfxstrand[d]: I mean... We could...
16:36mohamexiety[d]: mhenning[d]: ahhh I can see what that would break. thanks!
16:48gfxstrand[d]: mhenning[d]: I seem to recall the prop driver leaning on their sample positions in a built-in cbuf pretty hard and NGL blindly copying it.
16:48gfxstrand[d]: But I don't remember what they did in this case
17:29mhenning[d]: The proprietary driver also fails that test
17:35gfxstrand[d]: Maybe I should send the NVIDIA team my MR. 😉
17:36gfxstrand[d]: Ugh.. Looks like it breaks centroids
17:37gfxstrand[d]: There's also these unbound VB tests
17:38mhenning[d]: How important is it that we support this case? Can we just turn off VK_SAMPLE_COUNT_1_BIT in VkPhysicalDeviceSampleLocationsPropertiesEXT.sampleLocationSampleCounts ?
17:38gfxstrand[d]: I guess maybe we could
19:27djdeath3483[d]: mhenning[d]: you can report no support for 1x msaa
19:27djdeath3483[d]: anv does because there is no hw concept of 1x msaa
19:28karolherbst[d]: it always confused me what 1x msaa is supposed to mean anyway
19:29djdeath3483[d]: troubles is what it means
19:29mhenning[d]: djdeath3483[d]: Right, I suggested we could, more narrowly, disable it in sampleLocationSampleCounts above
19:31karolherbst[d]: djdeath3483[d]: I figured as much
19:36gfxstrand[d]: Either of those would be okay, IMO. We could also merge my draft MR but I'm really not thrilled with all the extra UBO and math just for a really dumb case.
19:42karolherbst[d]: would be funny if there would be native X1 support..
20:14mhenning[d]: gfxstrand[d]: yeah, I'd rather not pessimize the common case for this
20:21gfxstrand[d]: If just disabling sample locations on 1x passes CTS, RB me
20:22gfxstrand[d]: Let me try quick
20:24gfxstrand[d]: CTSing now
20:58mohamexiety[d]: phomes_[d]: is the kernel warning with the WARN_ON in the compression kernel reproducible?
21:02phomes_[d]: kind of. It does not always happen but it does if I e.g. open a new tab in the text editor or firefox. So it requires some kind of interaction there and then I see it. But the stack trace still says gnome-shell
21:03mohamexiety[d]: Hmm that’s going to be rough
21:04mohamexiety[d]: This is without installing the compression nvk, right? So old nvk + the new kernel is what triggers this
21:05phomes_[d]: yes. old nvk + new kernel
21:05phomes_[d]: I can try to find a nicer way to reproduce
21:06mohamexiety[d]: I’ll keep trying as well, thanks!
21:11marysaka[d]: phomes_[d]: Pretty certain it's this case:
21:11marysaka[d]: - You allocate 16KiB
21:11marysaka[d]: - You proceed to bind at an address not 16KiB aligned
21:13marysaka[d]: This was valid with previous kernels as it always used 4KiB granularity but now that we unlocked bigger page sizes it means that depending of the page size selected, we now cannot bind the BO without downgrading to something that align properly to the address
21:13marysaka[d]: So we really need that downgrade codepath for compatibility with old NVKs
21:14marysaka[d]: Otherwise we will break userspace...
21:43handrynash: i could not provide _that_ accurate and also fast formula all maximums and their capacity on calculator, because of difficulty, but I applied some iterations and when all are maximums 230k words can be packed to the hash for 32bit on 32bit machine word. (230303)×(2+230303)÷16 then (2^32−3314995775.9375)/4224=232k approx, that's gonna waste a bit but really , whatever. I expect this may be
21:43handrynash: more correct now. let's write a program that iterates from there, and uses the maximum , but it is worthless calculations data is entirely random,and it needs to check the overflow at store time anyways. Those are placebo or no-op stuff.
22:34operationsenc: Don't tell me about my family they belong to black dolphin all 4 , i have evidence too as to what they did do. They hate me and are afraid to get handled in a war now, more than 20killoff attempts plus torture and fraud arranged to me, from as very little child, why would otherwise a person like me be so cracked? I am getting a win finally from all the fraud, due to war, it's histories
22:35operationsenc: nastiest concpiracy. Entirely idiodic murderers. They do not operate their kill off attempts with their own earned money.