04:30steel01[d]: steel01[d]: Okay, I get it now. This section is handling if the size is not a multiple of 4 bytes. Then the first thing it does is copy a u32 at the end of the array. Which of course is... past the end of the array. :lesigh:
04:32steel01[d]: The gk20a gpccs inst fw is 8171 bytes. Basing off this, it seems that none of the firmware used by gm20b and gp10b is non-multiple-of-4, cause I don't hit this on my 64-bit targets.
07:49seatarnosko: So the dependency chain! You see it needs to know the maximum sum of packed presentation that of dependent instruction, but in the sequence of answers we do not know to where it's going to hit. It's now all about padding which would be one of the simplest paradigm to handle that, so that you call all indexes but the ones that does not combine to a dependent instruction eliminates itself, so
07:49seatarnosko: if every one instruction has dependency it will go all the way per operand to the last dependent inst. say we have 1024 instructions to get executed in the hash, that means operand 1 is the answer of inst1 which is negative when going forward, and all irrelevant pc's would be eliminating themselves and positive for debug values so negative would give . so negative hash arithmetic per pc
07:49seatarnosko: would go straight to the end, positive would give something to reindex. such as 1+8+11 being three dep pc's for operand 1 operation results used in inst1 to inst8 toinst11 we pack to every that index a maximum sum aka offset or whatever , so the count is three , so pc1 has 1, then pc2 has 8 then pc3 has 11 and rest of the cells in dep bank are padded. metadata such as variable names are all
07:49seatarnosko: indexed to be asked hence and everything would work in both ways variable at pcX or pcX having which variables etc. So debug presentation being more expensive to compile together, but still fast, we ask if IO can have dependencies? So in case we are at IO the value returned is in negative, but the indexes used must be upto certain point in the stream, best to do it, is to have IO hotspot
07:49seatarnosko: hash so always what gets done is comparison with the IO fifo. just like are we passed the last IO. IO would bottleneck the calculations indeed but not kill the performance altogether i'd suppose. So if 5th instruction is IO, one needs to decode the intermediate result and write it in the handler then receive the async value out of IO and continue. It's not a fifo actually it's that you ask
07:49seatarnosko: what is the first io, it says pc indexes 1-5 from data banks, then you pull those in. you receive a result you write something and reload the value gotten back to pc'X, so IO can be delayed if it has no alu dependencies. i.e if it's constant based writing.
09:40luc64627490: is nouveau capable of reclocking Pascal GPU (except GP10B) now?
10:05asdqueerfromeu[d]: luc64627490: It still doesn't (the closest thing is experimental and partially broken Maxwell reclocking support in some branch)
12:11naminingh: So the compiler takes 1year to be nearly or fully done and tested, and needs some testing framework , but i ain't gonna carry this out in public, if it was programmed for general public than we would need a bugtracker for it etc. So technically i am now done with my side of the story, it looks like Ryan wanted me to suck his dick, but he is going to end up with his own dick and genitals deep
12:11naminingh: in his throat like the estonian assaulters butcher tyrans and jack. We have no problem to carry this out.
12:43naminingh: The thing what happened in cambodia was by far a joke, it was entire deadly story, 5 people died in the gangster coward meeting, and the slut you talked about was from takeover events it played the nastiest sexual humiliation part, they always play the same, smuggle them into others territory then one thing leads to another they start to fabricate charges, call out the motogangster anal
12:43naminingh: buffolos who claim to buy things out, and start to terror people this way as scrub fuckers paradoxically enough. The slut swapped such bums several times and humiliated me daily every single day that i am a freak wanker and scared all our customers away, so it was mature time for one to get handled and after my assault it got handled, all the front drawer was emotied out from teeth and the
12:43naminingh: sluts bones were all broken, nastiest human trashes you had ever seen. But juri and martin akva got killed in the gangsters battles there, who were one of nicest people ever possible, so always in the conflicts best people suffer first, then a force against tyrans is launched as retaliation.
13:01naminingh: Have you ever thought about what is terrorist? Terrorist very often takes responsibility of an attack done, however gangster coward does never that, they start to issue lies as to how Mart Martin killed the people they actually killed etc. And i advise not to listen to any word that comes out of those lumpleg gangsters illborn people, because there just isn't any truth in there, people such
13:01naminingh: as indrek raud, alex enrico sif , laura keskinen, gloria terreur and their other actors in the so called syndicate, they do not release liable information at all.
13:02naminingh: it's not even about drugs, those actors are illborn, their dreams are entirely crooked.
13:07naminingh: junkies they undoudedly are, but drugs only do not make a devil out of a human.
13:17naminingh: I tried to have a vacation there for 2.5 years, and in reality could not really ever get it going due to the social environment aka. those cunt wank spammers, and male crank gangster scrubs who just humiliated me every day and fabricated charges of different kind ontop, i really do not know what to say about this stuff, it's a disaster that such people exist in the world.
13:23naminingh: in other words, yes there were many nice people around, but the actors turned half of the residential scrubs all against me, to humiliate me in such way that it ended up being untolereable , so i quit the country before my death and was cured in mental insitution for the first time to level 1 to 1 , they saved my life this time, where as their previous blooper to kill me did not carry out
13:23naminingh: earlier, i found some peace of mind after admitting the fact that the score is one one, and they really did their best to heal me up this time.
14:45misyltoad[d]: chatgpt off
14:45dwfreed: sadly that is not chatgpt
15:14mangodev[d]: misyltoad[d]: i'm sorry but i don't think chatgpt could say cuss words that much
15:14mangodev[d]: if it was, this chat would just be
15:14mangodev[d]: > Sorry, but as an AI model, it is against OpenAI policies for me to talk about the terrorists in Estonia.
15:14mangodev[d]: > Sorry, but as an AI model, it is against OpenAI policies for me to talk about sexual topics.
15:14mangodev[d]: > Sorry, but as an AI model, it is against OpenAI policies for me to talk about bombing horses.
15:14mangodev[d]: > Sorry, but as an AI model, it is against OpenAI policies for me to make defamatory claims about NVIDIA employees.
15:18mangodev[d]: what if the word "estonian" blacklisted you from irc >:)
15:19cwabbott: the person behind this has been doing this since way, way before ChatGPT existed
15:19mangodev[d]: i'm aware
15:19mangodev[d]: and i'm aware that it's not an algorithm
15:19mangodev[d]: unless the human mind is an algorithm now
15:19karolherbst: that's just the sad part of reality
15:20mangodev[d]: not markov chains, not chatgpt, just an unfortunate situation
15:21cwabbott: oh no, imagine 100 years from now ChatMARTM keeps doing this
15:21mangodev[d]: ChatMART sounds like a store
15:22mangodev[d]: 10% off all tokens! summer sale!
15:22cwabbott: hahaha, but people around when they started call them martm because that's the handle they started with
15:23cwabbott: before the escalating bans/blocks/etc that forced them to get creative
23:13steel01[d]: steel01[d]: ```
23:13steel01[d]: diff --git a/drivers/gpu/drm/nouveau/nvkm/falcon/v1.c b/drivers/gpu/drm/nouveau/nvkm/falcon/v1.c
23:13steel01[d]: index dd2ddc54ac60..313fb9d5709a 100644
23:13steel01[d]: --- a/drivers/gpu/drm/nouveau/nvkm/falcon/v1.c
23:13steel01[d]: +++ b/drivers/gpu/drm/nouveau/nvkm/falcon/v1.c
23:13steel01[d]: @@ -49,13 +49,13 @@ nvkm_falcon_v1_load_imem(struct nvkm_falcon *falcon, void *data, u32 start,
23:13steel01[d]: * does not get written
23:13steel01[d]: */
23:13steel01[d]: if (rem) {
23:13steel01[d]: - u32 extra = ((u32 *)data)[i];
23:13steel01[d]: + u32 extra = 0;
23:13steel01[d]: + memcpy(&extra, ((u8 *)data) + (i * 4), rem);
23:13steel01[d]: /* write new tag every 256B */
23:13steel01[d]: if ((i & 0x3f) == 0)
23:13steel01[d]: nvkm_falcon_wr32(falcon, 0x188 + (port * 16), tag++);
23:13steel01[d]: - nvkm_falcon_wr32(falcon, 0x184 + (port * 16),
23:13steel01[d]: - extra & (BIT(rem * 8) - 1));
23:13steel01[d]: + nvkm_falcon_wr32(falcon, 0x184 + (port * 16), extra);
23:13steel01[d]: ++i;
23:13steel01[d]: }
23:13steel01[d]: Does this seem like a reasonable workaround for the issue? There's no more kasan traces after this. Though, it still doesn't entirely fire up stating ctx timeouts. I'm presuming that's an unrelated problem.