00:22 gfxstrand[d]: I really don't get all this index re-mapping. It seems like we should be able to just use the slots the client gives us
00:27 gfxstrand[d]: I'm tempted to get rid of all the remapping and see if tests pass
01:13 airlied[d]: That is the intent of the API, some fw interfaces act different
01:47 gfxstrand[d]: I think I'm starting to get what's going on, finally.
01:51 gfxstrand[d]: Okay, so we need some sort of hash-map-like thing that tracks the current decode state of all the bound images. But the actual slots don't need to be remapped, I don't think.
01:54 gfxstrand[d]: But it's 10:00 PM on a Friday. I need to be done for now.
10:09 azalearudi: I did the computation model for the community, but it's the combinatory and permutation lesson, was so long ago, that those formulas i no longer remember. But my hot take is, it's the decoder that is almighty, the multiplier curve goes very steady in the beginning until it starts to rise more sharply, the most of the heavylifting is committed by the decoder hence, this is like one to many
10:09 azalearudi: mapping, i do not understand how many combinations are for the start, are there 1024*1024 or only like 32 invariants in round robin so 1024, i started to think recently the last offer is correct, because it can result in any of the combinations during runtime when you post the index of the sum, than all the procedures count, and all of them are overcommitting rather than working on single
10:09 azalearudi: computer word they work on more lines, so it is possible that there is no mathematical function to the entropy also if i think deeper, rather there is procedural overcommit which leads to correct timing in the hardware VLSI. But it is fairly sure that the paradigm i offered is correct, and i graduated all the set of methods to get there, as well as someone from military headquarters already
10:09 azalearudi: approved it as electrical suppression method already known in the art. Normal method to get to all permutes of values in hardware is through one ISAs register and through half or full adders normally 8bit long internally, so this would help to spot the method, so the answer should be derivable from 1164 what the massive function would look like very accurately gotten from the VHDL/Verilog
10:09 azalearudi: library known as ieee1164 as told. And those adders are internal to the library synthesized from +, not such as ISA's add. So the wizdom i used to get there, comes from lifetime knowledge on computers in addition to schools education.
11:04 amanogali: Personally i think the answer is in how bits are built with high uncertainty, one has to wait if the higher bit is in, it will change the carry bits, and hence the itnernal register for the adder and result accumulator isa register hence it never knows how to output bits until the last one is known, in another occasion in my world, i do not wait for those, i overcommit the first 14bits later
11:04 amanogali: and already know where this answer is going to lead, by packing all invariance into the first 14 bits in case of 32bit representation. Both procedures are compatible but stitched adders have higher entropy, but also higher flexibility and security. I do not claim that the hardware state gets corrupted hence. I claim that functionally they are equal but one is cheaper as well as lower voltage,
11:04 amanogali: where another provides more security and flexibility. And what i claim despite some americans fraud on channels and working with fraud scamming auhorities from Estonia. Americans are our allies through nato, not nigerian anal stuffers and non existing gold sellers, now that swedish and finalnd joined, they are our allies too, but not those scammers. Northern Americans are very big power and
11:04 amanogali: successful people in the world, i communicate with them, they also pump the nigerian oil, cause nigerian companies can not do anything other than anal lessons and other fraud. So in other words, which i can provide with my allies to nigerian, italian, south-african, sweidsh, finnish , estonian and english scammers is bullets at most on our territory in exchange to their anal lessons and
11:04 amanogali: robbery of computers money they nicked from me, and gold they did not provide to indrek on a trade, and same goes to gloria terreurs vivid syndicate dreamed asshole, if i see this shitbreak with laura or any of their scammers again on our territory, i crack them up with military guided operations. We do not talk about Russians or latvians or such, who all provide metal when the deal has been
11:04 amanogali: made, not a anal exhibition through illborn crocidiles who mate with scammers, so there is pleanty of intelligent Russians way over the real fakes and scammers in world, so despite the war in ukraine which is more of a political agenda and indeed very nasty, i would had otherwise gotten Russians to be our allies with ukranians, they are often very low on scamming properties, powerful nations,
11:04 amanogali: powerful people in general and can be dealt with.
14:41 keithbecker: Derivative as rate of change, like times two, *3 and *4 that my algorithm used is two centuries older concept than markov chains. And the limes aka limit that extracts the values through the calculus, the same thing can be done as said in integral computation also just derivatives come also integrating those flows in form of circles sin cos tan, same can be done in calculus and linear
14:41 keithbecker: algebra combined. the circles rate modifiers just flip the sign and are probably one to one match according to my beliefs. You forget to understand maths and you become pain in the ass, more importantly those crews who tried to kill me in cambodia i remember those all. It looked like i have weaker hand, but they were all recruited by this medical fraud of milking my proteins to the same
14:41 keithbecker: area, they were unsuccessful and we kill all of this crew. Good luck with your markov chain and sniffing daves the heroic fart discuis and aids heros fart again. You have no rights to assicoate me with those idiots nor from cambodia nor from here. Estonians get killed in masses in upcoming war, they themselves digged a hole to them.
14:50 snowycoder[d]: karolherbst[d]: That works for chip GK208B, I'm not seeing a warning anymore
14:52 snowycoder[d]: I can also test for GK104, but I don't have GK110 and GK110B (though I guess it's the same).
14:52 karolherbst[d]: yeah.. just need the same change in their specific files
14:52 snowycoder[d]: How did you know that that was the right bit to flip?
14:53 karolherbst[d]: I assumed they wouldn't choose a different bit and it was also the same reg that works on maxwell2
14:53 karolherbst[d]: and they haven't moved the other reg between kepler and maxwell2
14:53 karolherbst[d]: sooo...
14:53 karolherbst[d]: just a guess
14:53 snowycoder[d]: There are docs for Maxwell2?
14:54 karolherbst[d]: nope
14:54 karolherbst[d]: but there is `nvk_push_draw_state_init`
14:54 karolherbst[d]: the `get_sms_hww_warp_esp_report_mask_reg` part
14:56 karolherbst[d]: nvidia generally keeps things more or less backwards compatible between generations, and if they change things, they often also just move things around
14:57 karolherbst[d]: I'm sure most of the changes there are because now there are more SMs per arch possible and they needed to move things around 🙃
15:07 snowycoder[d]: This is both really interesting and really complex, but it makes sense.
15:16 snowycoder[d]: I think I encountered a kernel panic when I tried arch's fork of 6.16, when modprobing `nouveau` it found my Blackwell GPU and froze.
15:16 snowycoder[d]: Is it a known bug? I worked-around it by returning to 6.15 (I only needed to use Kepler)
16:07 gfxstrand[d]: There was something in rc2 or 3, I think
22:39 dwfreed: having tested nouveau on 6.12 with a kepler (GT750M), I must unfortunately report that I went back to nvidia 470 drivers; I do blame nvidia, not you guys (and I'm sure apple should get some blame as well)
22:52 snowycoder[d]: dwfreed: What went wrong?
22:55 dwfreed: first problem: rather annoying cursor flicker
22:55 dwfreed: second problem: waking up displays, the HDMI attached one would come up with nothing on it as if it had been disconnected and reconnected
22:57 dwfreed: third problem: this laptop has a mux chip, and the backlight controls for the laptop screen are part of the mux chip; putting displays to sleep, the laptop screen backlight stays on
22:59 dwfreed: and the effort of getting hardware accelerated video decoding working didn't seem worth it (i'm on debian, so no existing packaging) with all the other problems
23:06 steel01[d]: dwfreed: Huh. I get a corrupted cursor on android with tegra using gallium. Interesting to hear that's not just a tegra or android issue.
23:16 snowycoder[d]: dwfreed: Thanks for the list of bugs!