08:09airlied[d]: uggh seeing some diagonal crap on a vkrunner rectangle draw
08:19airlied[d]: is this one of the magic register things?
08:41karolherbst[d]: yep
08:41karolherbst[d]: could be the helper invocation thing
08:42karolherbst[d]: the reg/bit might have changed
08:42karolherbst[d]: nvidia also uses an mme macro to set it, so you should see it in traces
09:32airlied[d]: I'll try and dig into it tomorrow then
11:47performingdarts: pavlo_kozlenko[d]: well if wasn't clear for you , I am not even sure why I am helping your mad cunts here.361+328+324+336+361+322+−512−144−72−144 you see, two almost equal sides of the hash and 336-324=12 in the middle, the function is hence 12+12+4 28 is the value added and queried from the hash, and procedures are like above 361+328+324+336+361+322+−512−144−72−144=1160-1024-36-144=28
11:47performingdarts: and 361+328+324+336+361+322+−512−144−72−144=1160+361+328+324+336+361+322-1024-36-144=2060 hence 361+328+324+336+361+322-2060=-28 you use same polarity of the things but on request you cache the minus value to get things eliminated, the real format has gaps of 4, i.e starting as ideal distribution/sequence of uniform kind to represent powers of twos as invariant non-colliding sums from 5
11:47performingdarts: 9 13 17 21....etc. and so forth. That is what the function is, twice the value+delta of the halves. it can be 0+0+1=0+1 as bare minimum then 0+0+2 0+0+3 etc ,but also 2+2+1*2-1 as 9 1+1*2+1=5 for second one etc. so forth and the easiest to remember as to how to store this format is 328-324is 4 while 328-322 is six, so it tends to know that to use 4 as delta since it is smaller, but you
11:47performingdarts: can use also use 1 to get the difference down to minimal as told.
11:55performingdarts: it wasn't me who engaged into sexual interactions with idiots like you, it was laura tornado doing extorting based of such achievements too, and the trash is no longer allowed to be near customers, both the abuse cranks and the abuse sluts, or firearms are used to wipe you out from out territory.
11:55performingdarts: I just perform works that adhere to logics also, in the world of programming, and have done my last two years little parasports too.
12:06performingdarts: estonian programming related higher education people, requested that info from me, that can you describe the function of that format, well of course, cause such number system that would not allow to place such hacks onto sw stack would be rejected by mathematics community around the world, logically enough, cause mathemaatics is way older than personal computing.
12:19performingdarts: it's both europes and american policy together even more so europes one that outsiders abusers in the world are hired to ban logics and real people around the world, and park their asses to annoy eastern people until the war breaks out, absolutely intolerable violators and annoyers, and entirely idiotic people.
12:28performingdarts: Yes china manufactures a lot of atoms/arm/nvidia chips but they can be placed sanctions to those chips which would not be effective measure, since they would manufacture their own chips very soon, they are pretty powerful, though my labs are way above chinese intelliegence, but they requested this info to manufacture drones, and yes when europe gets their feet off the stomach back to
12:28performingdarts: the militarization projects quickly that escalates into massive war that probably nato in the end wins.
12:29performingdarts: its the technology of super calculators that plays the curcial part and throughput of the manufacturing, which is higher with also quality for europe and usa even than for russia as the way i know.
13:23gradientofraster: well dad compared to my sisters, has or had put maximum abuse at me, just like many others, but difference is there, he has paid me microsums of money for this, while others have extorted in addition to persistent abuse daily. For those people who scammed the world, They got approval and hopes from my dad and allowance/permission to carry out such shit also. I gotta be honest, nothing
13:23gradientofraster: is the same ever since where such abuse retards got privileged over people like me. But they likely anyways get all killed for the shit they did. My world wide friends and supporters network is extraordinary and only lines of real powers that can go against Russians are theoretically my lines. It's that i fail to see huge violations or tactical amoral mistakes that russians ever did.
16:43gfxstrand[d]: Whoops! Did I just plug in a GK106? :bim_think:
16:46mangodev[d]: what happens in this channel in the mornings ðŸ˜
16:57aligatornrone: now essentially the highest latency or most latent operation goes to encoder , it technically is extremely fast, since memory is byte addressable, but we know that this operation isn't supported often enough on pre glsl1.2 chips like bitshift by say 16bits to right and you have troubles often with dma fw or hw to do rmw, but it has floating point mul which is the same, it also has
16:57aligatornrone: saturated overflow from claimed microsoft fragment program arb reliance of the spec, but no sign interrupts etc. so what happens is with the procedures i posted all you ever want to do is build a scaffold for offsets, since decoder is now also extremely performant, but FPGA-s was the thing i wanted to talk about, they are the biggest performers that are ever possible, and hw multipliers
16:57aligatornrone: like dsp blocks on fpga-s are not needed even for decoder, cause obviously it would decode the number in hardware. however all dma and things alike are expected to be stable on those too. FPGA-s are fun fucture projects on 3d and general compute. But in general what is fun also is mjt/mlc flash fpga's which is not basing on euv etc. first of them has no wareout problems. What happened in
16:57aligatornrone: the morning was your fucking abuse case is over , we arrest your shitbags soon by force also. That does not mean i have problems when you are not around, it means just like russians had enough of you, i said several time i had enough of your shit, if your fuckers ever harass my territory in life, you get owned and absolutely finished off with torture and violance.
17:01x512[m]: Looks fine from Matrix side.
17:04HdkR: x512[m]: Like a fine wine. 🎉
17:19snowycoder[d]: gfxstrand[d]: You want to add support for keplerA?
17:46mohamexiety[d]: mohamexiety[d]: hm, so this is how it looks like on Ada, without shared memory:
17:46mohamexiety[d]: mthd 3bc4 NVC7C0_CALL_MME_DATA(120)
17:46mohamexiety[d]: .VALUE = 0x34040000
17:46mohamexiety[d]: mthd 3bc4 NVC7C0_CALL_MME_DATA(120)
17:46mohamexiety[d]: .VALUE = 0x24030
17:46mohamexiety[d]: mthd 3bc4 NVC7C0_CALL_MME_DATA(120)
17:46mohamexiety[d]: .VALUE = 0x10001
17:46mohamexiety[d]: mthd 3bc4 NVC7C0_CALL_MME_DATA(120)
17:46mohamexiety[d]: .VALUE = 0x210ff
17:46mohamexiety[d]: and this is with:
17:46mohamexiety[d]: mthd 3bc4 NVC7C0_CALL_MME_DATA(120)
17:46mohamexiety[d]: .VALUE = 0x340c0100
17:46mohamexiety[d]: mthd 3bc4 NVC7C0_CALL_MME_DATA(120)
17:46mohamexiety[d]: .VALUE = 0x24030
17:46mohamexiety[d]: mthd 3bc4 NVC7C0_CALL_MME_DATA(120)
17:46mohamexiety[d]: .VALUE = 0x10001
17:46mohamexiety[d]: mthd 3bc4 NVC7C0_CALL_MME_DATA(120)
17:46mohamexiety[d]: .VALUE = 0x610ff
17:46mohamexiety[d]: of note is that compared to Blackwell, it's two entries, and things don't change as much as they do on Blackwell.
17:46mohamexiety[d]: is ada QMD v03 or v02?
18:40freitagatspace: i know every single one from osdev, their addresses names like bcos, kleine, john hodges, travis fecalbrecht, kazinsal all that trash who bullied me, i need no logs for that, from every channel debian or llvm not to mention also from discoord also from matrix rooms, i have complete absolute memory, also from device tree overlay rooms every last one i punish if you do not fuck off from my
18:40freitagatspace: sights and regions in real life.
18:45mhenning[d]: mohamexiety[d]: ada is clc9*. clc9c0qmd.h shows both 02_04 and 03_00 support. I'd guess that they're more likely to use the recent one which is 03_00
18:45f_: karolherbst dwfreed ^^
18:46mhenning[d]: I'd guess that you can check that by looking at the QMD_VERSION or QMD_MAJOR_VERSION fields
20:17TimurTabi: karolherbst: I have an update to the Nouveau wiki for you to consider. https://gitlab.freedesktop.org/nouveau/wiki/-/merge_requests/54
20:17karolherbst[d]: yeah, already saw it
20:17karolherbst[d]: I'm wondering if the tone is a bit too much, but it's probably fine...
20:26airlied[d]: ah looks like the regs moved
20:33airlied[d]: okay fixed in my branch
20:59TimurTabi: Yeah, the tone might be a bit much, but I REALLY don't want anyone thinking that as soon as Nouveau fails, that they should follow these steps and then starting emailing tarballs.
20:59TimurTabi: karolherbst: the openrm git repo is full of people not following the rules and filing bug reports because they don't want to go through official channels.
21:00karolherbst[d]: what's the harm done there though?
21:01TimurTabi: the harm is that we have people already in place who can handle bug reports, but instead they want the RM developers to drop what they're doing and deal with the reports.
21:02TimurTabi: And when those developers are too busy to look at whatever obscure issue is being reported, it makes us look bad.
21:12airlied[d]: gfxstrand[d]: https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/34513 just want to get rid of some of the basic patches
21:44gfxstrand[d]: Well, Kepler A is now faulting. The good news is that the disassembler seems to be okay with my shader. Just need to figure out what's busted.
21:59mohamexiety[d]: mhenning[d]: thanks! yeah I'll try next with bigger shared variables then as right now it's too clear. :thonk:
21:59mohamexiety[d]: from clc9c0qmd.h, we do have these which look like they'd be impacted by choice of shared mem:
21:59mohamexiety[d]: #define NVC9C0_QMDV03_00_SHARED_MEMORY_SIZE MW(561:544)
21:59mohamexiety[d]: #define NVC9C0_QMDV03_00_MIN_SM_CONFIG_SHARED_MEM_SIZE MW(567:562)
21:59mohamexiety[d]: #define NVC9C0_QMDV03_00_MAX_SM_CONFIG_SHARED_MEM_SIZE MW(574:569)
22:00mohamexiety[d]: #define NVC9C0_QMDV03_00_QMD_VERSION MW(579:576)
22:00mohamexiety[d]: #define NVC9C0_QMDV03_00_QMD_MAJOR_VERSION MW(583:580)
22:00mohamexiety[d]: #define NVC9C0_QMDV03_00_OCCUPANCY_MAX_SHARED_MEM MW(591:584)
22:00mohamexiety[d]: it's not too clear*
22:15gfxstrand[d]: gfxstrand[d]: And... `hw_tests::test_op_prmt` is now passing on GK106.
22:15gfxstrand[d]: snowycoder[d]: I'm mostly curious. Also maybe a little bored and don't want to conflict with what you're hacking on.
22:16gfxstrand[d]: Also, the codegen emit code for KeplerA/Fermi is utterly unreadable. :blobcatnotlikethis:
22:25mhenning[d]: gfxstrand[d]: lol have you tried reading some of the optimization passes
22:26gfxstrand[d]: Oh, I have.
22:26gfxstrand[d]: I didn't say that was the *only* unreadable code in codegen. 🤪
22:26gfxstrand[d]: The texture lowering is impossible
22:29gfxstrand[d]: Also, I really hope our spilling code is amazing because 63 regs is gonna suuuuuuuck
23:29snowycoder[d]: gfxstrand[d]: more than keplerb? that's an achievement
23:29snowycoder[d]: gfxstrand[d]: don't worry, right now I'm still figuring out sucalc ops, `suclamp` is entirely folded though!
23:33gfxstrand[d]: snowycoder[d]: It helps that it has 4-byte encodings for most of the ops but it's all dead code and never gets used. 🤪
23:37mhenning[d]: yeah, beware of dead code in codegen
23:37mhenning[d]: I delete it sometimes when I come across it but there's still plenty lingering in the corners
23:44imprisonment: i am not much worried neither of the nigerian gang, khmer gang, mob of thailand nor anything related to the slut. They will be with rest of the mentioned already europeans aussies south-africans and imprisoned in the jail, that looks like this for similar elsalvador gangs https://www.youtube.com/watch?v=H42zWaD4A4s&pp=ygUdY2Vjb3QgcHJpc29uIGZ1bGwgZG9jdW1lbnRhcnk%3D they and the slut are yet
23:44imprisonment: to decide whether they want that life, or will fuck off out of my territory..