07:26tomeu: kherbst: ok, will take that into account
07:27tomeu: kherbst: btw, I see that PIPE_COMPUTE_CAP_MAX_GLOBAL_SIZE in clover_support_nir_target is a constant, instead of using vram_size
07:27tomeu: guess there's a branch somewhere with more compute stuff?
07:28kherbst: tomeu: https://gitlab.freedesktop.org/karolherbst/mesa/commit/8339e2ed1f986344e04dc9cacfad1e082d7ff8f1
07:29kherbst: but there isn't that much more which is usefull anymore
07:29kherbst: most of the commits would have to be rewritten
10:22mrsinisalu: very nice US&A!
10:28mrsinisalu: MrPooper is very obsessively sharp guy, say like cryptomwk instead of playing with the brain neurons and doing sane stuff, he at AMD logs my activity, like usually i behave this and sometimes this, have you allready built something that predicts how i think VOLKOV BRAINPAIN over markov sanechain?
10:33mrsinisalu: So anyways back to the point! I introduced some of the compression of the command stream in a known primitive fashion
10:35mrsinisalu: The point is you'll have to do some resource computation and delay computation, highly easy arithmetic calculations.
10:36mrsinisalu: to be able to plug such things in, the code should not be very difficult.
10:36mrsinisalu: for the starters i also said, that actually VLIW is slightly nicer on this method than that of SIMD.
10:41mrsinisalu: most important is -- how many queue entries chip has aka num_maxTG*num_maxTG and also maybe how many ALUs it has to able to run decompressor in parallel for various opcodes
10:44mrsinisalu: and for VLIW chips for instance cypress it goes like this: there are 32waves on 16wide VLIW's SIMD, queues are hence 4*32*32 so it is calculated num_maxTG*num_maxTG*bundle_width on VLIW instead
10:45mrsinisalu: so it is 2048 Per one SIMD
10:45mrsinisalu: now alus per SIMD can be taken all in parallel on cypress chip so those are obviously 16*4 hence
10:46mrsinisalu: 64 can be done in parallel on SINGLE CU/SIMD VLIW chips
10:46mrsinisalu: those specs are lower for pure SIMD aka hw scoreboarding chips quite abit
10:52mrsinisalu: next thing what is also preferable to be aware about or to know is how many VGPR or SGPR registers based oc occupancy a single TG -- Thread group would have
10:53mrsinisalu: this calculation is again a simple one, so if the cluster of 16wide/lane VLIW SIMD processor has 256000bytes aka 256kb register file for instance
10:53mrsinisalu: and maximum TG count is 32
10:54mrsinisalu: you divide 256000 with 4 then 64 and then 32
10:55mrsinisalu: result should be 32VGPRs of float or 16 as double
10:56mrsinisalu: if you have fewer obviously that will hinder parallelism but you will have more then 32 VGPRs actually 8*4 vector registers
10:57mrsinisalu: now the compression and scheduling i talked about on #dri-devel this has more variations on how to do this, many ways suite
11:00mrsinisalu: before any of the coding is being started those 5mentioned things should be something that a programmer is aware about.
11:02mrsinisalu: that is why hw journalists provide some of the cool info on the web, so a developer would have some less nervy start to the stuff, however scheduling and compression bits were not given on those journals or articles
11:03mrsinisalu: as we know allready i did sniff out those bits intelligently from miaow hunks and multi2sim and attilagpu and such, mostly miaow in case -- as i was simulating some of the flow and debugged it a bit
11:11mrsinisalu: and yep, for compressing and redirecting the queue entries to be executed in the corresponding functional units aka ALUs on chip
11:12mrsinisalu: i suggested the fast pipe of subtract on two's complement system, this is equivalent of bitwise AND and OR basically two gate delays
11:13mrsinisalu: the correct responce or answe comes with just ingoring the carry on twos complement subtract
11:14mrsinisalu: i think all hw does it this way
11:16mrsinisalu: I also referenced a turkish uni. page for this, i would had thought all were aware of this, however this is not seeming to be a case.
11:17mrsinisalu: http://eem.eskisehir.edu.tr/egermen/EEM%20232/icerik/Week%206%20Arithmetic%20Functions.pdf just read page32 and forward from this pretty nice tutorial or paper or article or slides, whatever
11:18mrsinisalu: Yes i understant that paging should be something that makes things securer while doing it slower by major factors
11:19mrsinisalu: I myself am not very big in RAM buffering supporter or security fan
11:19mrsinisalu: not during runtime of an application anyways, since it slows down things to crawl imo
11:20mrsinisalu: for an example
11:20mrsinisalu: even though ddr2 or pc1600 standard ram should have 1600MB/s
11:21mrsinisalu: which is slightly more then 32bit transfer per nanosecond for instance
11:21mrsinisalu: then reality is more like 26MB/S transfers with all the paging in-kernel faciilities, so it all that delay even on pretty fast DDR is outstanding
11:22mrsinisalu: it is just going to hurt perfromance too bad
11:31mrsinisalu: what i basically think is: that it would be time to offer saner and performing code for low-end hardware and to some poorer people, in an event not to consume so much resources etc.
11:32mrsinisalu: and best way to do it is to form some dedicated channel to approach towards it, some little team to carry that out for CPUs and GPUs of various kind, where the requirements are that you understanding the issues too.
11:35mrsinisalu: no one is demanding open source world to stay in the interest field of hw vendors, you can also say fuck off just like that, we are not going to follow your spec of overconsuming world resources.
11:35mrsinisalu: it is not like they would have something to tell about this, or do they really send some men in black?
11:43mrsinisalu: I do not recognize NVIDIA , INTEL or AMD or ARM to be gods saying how should i use their hardware and to tell me what i can i do and what can't i do, i should decide my own and so should you, what is right and what is wrong, i.e the point of stanley milgram famous test.
11:44mrsinisalu: when tutor says that you have to make an electric shock because the attendent in the test did not answer a random question correctly, you say i am not willing to give this shock to one, I have independent thinking and i do not obey such order, cause the person will have pain involved due that shock
11:55mrsinisalu: and about myself, perhaps i have been also destroyed in most part cause i have independent thinking which is rare enough those days. I am not going to carry out any changes despite that or regardless of that still :(
12:12mrsinisalu: i've been yeah punished bad by my fellow countrymen, but they have evil people anywhere, i've been punished every time i did not defend myself enough, even though the trouble is about me, i did not cause this, and i still have many native good friends locally here, and it isn't like you can not bring ott tänak down either if you really wanted, samewise it isn't looking like you compete with him in always winning style either.
12:14mrsinisalu: in/as perhaps, so yeah i need to go, i was brought down conceptually very early in childhood allready, i am not very capable talent anymore.
12:18mrsinisalu: you are thinking that i could do those jobs to get rich, but an as injured cripple what would i acheive with this money? really nothing anymore, my health was conctantly under attack ever since i was 10, i have not strong views anymore, just every day life with some fewer violations present then when i was more promising.
12:20mrsinisalu: drive the lamborghini to tell people i am not the one i was suppose to be cause i faced too many violations
12:30mrsinisalu: I will never see any reason to blame or attack innocent people cause of that, even though i understand from many nutters behavior why terror takes place in ugly forms.
14:09juri_: wow. that was.. something.
15:50FLHerne: Sometimes I'd swear someone trained one of these OpenAI text-generation bots on the graphics-dev channel logs
15:51FLHerne: It's got the same characteristic: the words are right, the grammar is right, the sentences are structured coherently -- and yet there's no meaning...
18:27mrsinisalu_: juri_: I never got any meaning from one of those here either. I also suspect something cruel as once there was some human being that trained bunch of parrots long time ago, they randomly yell some words of english like markov chain and Ai about which trained parrots know nothing about since there have not been a so called deep dive to the internals apart from those words, i suspect there is speech recognizer synthesis kit involved somehow which
18:27mrsinisalu_: routes those speech synthesized soundwaves straight to IRC.
18:28mrsinisalu_: Afaik FLHerne is plausible one of the many parrots identified.
18:34mrsinisalu_: juri_: this was just a joke, i never have thought about such things (as it's really retarded and paranoid path of thinking), parroting is one of many psychological tricks used on me, it is so called info starvation and part of cracking a nice person down that way.
18:36mrsinisalu_: it has been used on me for ages in groups , they switch roles spreading the words and parrot all the time exactly the same things, very annoying to tolerate but it works well on cracking someone up, old trick but had success on me too.
18:49mrsinisalu_: it is a form of conspiration or sabbotage against me, also mostly attended by male opponents in history, it has taken such heights that maybe it is confusing someone.
18:51mrsinisalu_: science vs. evilness is always in competition with each other.
19:05mrsinisalu_: another path is predictive programming which normally does not carry so much affect if faught against minorly, you see a human being on this planet makes oneself to belive what one wants to belive itself, and fools onself to belive it.
19:07mrsinisalu_: it really does not mean that female will belive what they say about me most the time in mass psychosis effect, it just means that men cockblockers are stronger to terrorise female.
19:11mrsinisalu_: in other words, such what we see is an attempt to cause braindamage to me, i have lots of experience with raw psychology tricks and tacticks, cause i have suffered under such for decades and thought about them all through
19:15frobos: I'm using kernel 4.14.118 and decided to test nouveau for a while. It seems that the driver stability improved since my last usage. I'm running the system for more than a day and no system freeze yet
19:15frobos: tested on several games and reclocked to max, all seems fine
19:17frobos: for this Kepler NVE4 it seems to be good