00:02 HdkR: Yay for then switching to RISC-v and making their own lives easier 😛
00:02 HdkR: for them
00:03 fdobridge_: <g​fxstrand> And it's not like NVIDIA is special here. No hardware company is going to spend millions of dollars worth of engineering and legal effort on 10-year-old hardware just because some Linux users are complaining.
00:04 fdobridge_: <g​fxstrand> (Yes, I'm counting from the day they released the first GPU. At a hardware company, a product is over the moment the first unit ships.)
00:04 fdobridge_: <p​avlo_it_115> I had this in mind: has anyone tried to make a nouveau-fw type from the reclocking unlock firmware (which was extracted from the NVIDIA binary)?
00:04 fdobridge_: <p​avlo_it_115> https://cdn.discordapp.com/attachments/1034184951790305330/1200954539172450437/b88a28c2561d9324.png?ex=65c80f12&is=65b59a12&hm=882b8f2cb15d6a8906f339b1b892de3f026b4c36a13601eb16da0ae5f476ac3b&
00:04 fdobridge_: <k​arolherbst🐧🦀> though compared to the others NVidia has an impressive long support cycle for hardware
00:05 fdobridge_: <k​arolherbst🐧🦀> when did Kelper go EoL? last year?
00:05 fdobridge_: <k​arolherbst🐧🦀> that's like 10 years
00:05 HdkR: Impressive amount of time for tech hardware
00:06 fdobridge_: <k​arolherbst🐧🦀> extracting the firmware isn't the problem
00:06 fdobridge_: <r​edsheep> https://endoflife.date/nvidia-gpu "Ended 2 years and 4 months ago"
00:06 fdobridge_: <k​arolherbst🐧🦀> mh???
00:06 fdobridge_: <p​avlo_it_115> Problem adapting nouveau to this firmware?
00:06 fdobridge_: <r​edsheep> Going by previous pace Maxwell should already be dead
00:06 fdobridge_: <k​arolherbst🐧🦀> that's kinda bogus 😄
00:07 fdobridge_: <r​edsheep> How so?
00:07 HdkR: meanwhile AMD is cutting out Vega support from their software and they had hardware launch with it last year :(
00:07 fdobridge_: <r​edsheep> Yeah...
00:07 fdobridge_: <k​arolherbst🐧🦀> https://www.nvidia.com/download/driverResults.aspx/215840/en-us/
00:08 fdobridge_: <r​edsheep> Security support isn't over, but it's no longer considered active support
00:08 fdobridge_: <k​arolherbst🐧🦀> yeah well.. you get driver updates
00:08 fdobridge_: <k​arolherbst🐧🦀> and they do port to newer kernel versions afaik
00:09 fdobridge_: <k​arolherbst🐧🦀> but sure.. you won't get new features, but that's old hardware
00:10 fdobridge_: <k​arolherbst🐧🦀> from an end user pov only two things matter: 1. do I get updates and 2. does it run on the newest OS?
00:10 fdobridge_: <p​avlo_it_115> it seems there are only security updates
00:10 fdobridge_: <k​arolherbst🐧🦀> so? it still runs on the newest kernel
00:10 fdobridge_: <k​arolherbst🐧🦀> that's a big deal
00:11 fdobridge_: <r​edsheep> To be entirely clear, I do agree that having support for as long as they have provided it is commendable, and yes most new features just wouldn't work anyway.
00:11 fdobridge_: <k​arolherbst🐧🦀> they could have stopped at... whatever was released 2 years ago, but they didn't
00:12 fdobridge_: <k​arolherbst🐧🦀> if you get sec updates or perf opts for new games... meh
00:12 fdobridge_: <k​arolherbst🐧🦀> at this point the difference is pretty much not existing
00:13 fdobridge_: <k​arolherbst🐧🦀> yes
00:14 fdobridge_: <a​irlied> yeah I think the thing kepler users are hitting down is wayland integration, egl etc
00:16 fdobridge_: <a​irlied> yeah I think the thing kepler users are hitting now is wayland integration, egl etc (edited)
00:17 fdobridge_: <k​arolherbst🐧🦀> ahh yeah, fair
00:17 fdobridge_: <k​arolherbst🐧🦀> though not sure how usable that's on modern hw
00:17 fdobridge_: <k​arolherbst🐧🦀> I heard both ways
00:17 fdobridge_: <g​fxstrand> They tend to care a bit more about security support because the potential legal fallout of security bugs is a lot higher.
00:19 fdobridge_: <r​edsheep> For sure. I am also kind of playing devils advocate even bringing any of this up. In my day job I am annoyed if anybody is using something 6 years old, let alone 10.
00:19 fdobridge_: <r​edsheep> If it were only up to me everybody would buy new hardware every 4 years, but not everybody had the money and I don't want landfills full of cards that some people might still tolerate.
00:20 fdobridge_: <k​arolherbst🐧🦀> thing is.. some of their GPUs are used in DCs for eternity
00:20 fdobridge_: <k​arolherbst🐧🦀> and for most parts, it doesn't matter if your GPU is a DC one or a desktop one. Though they also have their quadro line
00:23 fdobridge_: <g​fxstrand> *Sandy Bridge has entered the chat.*
00:25 fdobridge_: <r​edsheep> Just in the last week I have worked with people using Nahalem and earlyish AMD FX. Those old drivers can't render anything right anymore.
00:25 fdobridge_: <k​arolherbst🐧🦀> who is still using Sandy in the business world? 😄
00:25 fdobridge_: <r​edsheep> Oh, it's a thing.
00:25 fdobridge_: <k​arolherbst🐧🦀> I mean.. sure
00:29 fdobridge_: <g​fxstrand> IDK about the business world but I can't count the number of people who asked me for a Sandy Bridge Vulkan driver. 😂
00:30 fdobridge_: <r​edsheep> Small businesses don't have the money to upgrade and don't have enough layers. The owner ends up directly making the decision to never upgrade, almost without fail.
00:30 fdobridge_: <r​edsheep> Something has to explode before they will authorize anything.
00:32 fdobridge_: <g​fxstrand> Yeah...
00:34 fdobridge_: <k​arolherbst🐧🦀> cursed rust stuff I'm working on: https://godbolt.org/z/Mv1droof9
00:34 fdobridge_: <k​arolherbst🐧🦀> yeah.. I mean.. Linux users be like that
00:34 fdobridge_: <k​arolherbst🐧🦀> I still fixed nv30 bugs in 2023 😄
00:34 fdobridge_: <g​fxstrand> Yup
00:35 fdobridge_: <k​arolherbst🐧🦀> or was that 2022?
00:35 fdobridge_: <k​arolherbst🐧🦀> I don't even remember
00:35 fdobridge_: <k​arolherbst🐧🦀> I need that for... things... but that cursed pattern matching is cursed
00:36 fdobridge_: <k​arolherbst🐧🦀> how does that even work... 😄
00:37 fdobridge_: <g​fxstrand> Uh oh... Do I want to know?
00:37 fdobridge_: <k​arolherbst🐧🦀> 😄
00:38 fdobridge_: <k​arolherbst🐧🦀> it's going to be awesome, but uhhh...
00:38 fdobridge_: <k​arolherbst🐧🦀> a lot of research on how I want things to be
00:38 fdobridge_: <k​arolherbst🐧🦀> apparently `ref base` matches on the actual field name...
00:38 fdobridge_: <p​avlo_it_115> I have a gt 630 kepler video card here now. And now I'm thinking about how to do automatic reclocking (just for the sake of the experiment). I'm thinking about how drivers determine when to increase the frequency. It's banal to think if something happens on the monitor. But I want to go the other way (because the option with tracking... I have no idea how it will work.) You can make a variable file type. And when starting the 3D program
00:39 fdobridge_: <k​arolherbst🐧🦀> arbitrary names: https://godbolt.org/z/eb77exqof
00:39 fdobridge_: <k​arolherbst🐧🦀> @gfxstrand I was wondering if you need something like that as well 😄 Or well.. needed 😄
00:41 fdobridge_: <g​fxstrand> Yeah, those matches are a little crazy
00:41 fdobridge_: <g​fxstrand> I've seen far worse
00:41 fdobridge_: <k​arolherbst🐧🦀> heh
00:43 fdobridge_: <k​arolherbst🐧🦀> anyway.. I kinda want to rework my `cl_mem` thing in rusticl, and I need to figure out how to do inheritance in a sane way I won't have to rewrite everything else as this also has to match with that pointer -> rust type thing going on... and pulling the base struct like that actually helps for stuff
00:43 fdobridge_: <g​fxstrand> Yeah...
00:45 fdobridge_: <k​arolherbst🐧🦀> and it's also going to be useful for CL pipes, as those are also just `cl_mem` objects :ferrisUpsideDown: honestly.. who designed this API
00:45 RSpliet: Lee Howes!
00:45 RSpliet: Well, among many others
00:50 fdobridge_: <g​fxstrand> Rust-based GL state tracker when? 😝
00:51 fdobridge_: <k​arolherbst🐧🦀> pain
01:03 fdobridge_: <g​fxstrand> I'm kidding, of course. No one should ever reimplement GL again.
01:03 fdobridge_: <k​arolherbst🐧🦀> tell that to google
01:04 fdobridge_: <g​fxstrand> I once had notions of cleaning up the Mesa state tracker but no... Best leave a working monstrosity alone... 😬
01:04 fdobridge_: <g​fxstrand> Google just likes reinventing things. At least Microsoft knew better.
01:17 fdobridge_: <g​fxstrand> Also, Google has only ever reimplemented GLES, to my knowledge. Even ES 3.1 + AEP is massively easier to get right than GL 4.6 compat profile
01:18 HdkR: On Android you can now just ship Zink and get GL :P
01:21 HdkR: Android even has the new-ish EGL extension to get frontbuffer rendering, which was a major concern five years ago or whatever
01:51 HdkR: Oh wait, I wonder if they have an equivalent Vulkan extension for that EGL extension. I was reaching that there would be one
02:01 fdobridge_: <g​fxstrand> There is a Vulkan extension for front buffer rendering. It was driven by a guy at Samsung so I assume it's shipped on some mobile somewhere.
02:12 HdkR: Nice. So at least some device has it, even if it is only like Exynos+RDNA which is super small percentage
02:15 HdkR: Or something more magical?
02:38 HdkR: `This allows the application to use the image while the presention engine is accessing it, in order to reduce the latency between rendering and presentation.` Sounds like it
04:16 TimurTabi: karolherbst: yes, I have an idea on how to clean it up.
05:04 fdobridge_: <g​fxstrand> Shared presentable image, I think. IDK it's on AMD or some Qualcomm device. It's been a while so it was probably happening in tandem with or before the AMD negotiations.
05:46 fdobridge_: <m​henning> @gfxstrand Oh, I just realized that your mmu fault could be related to a NAK bug I started fixing in September and then promptly forgot about. Maybe try this patch and see if it changes anything: https://gitlab.freedesktop.org/mhenning/mesa/-/commit/ee983530d51f2d4c383e10cb16224c4d2b1a6da9
05:46 fdobridge_: <m​henning> that's obviously not a complete fix and it isn't well-tested, but I could write up the rest of it pretty quickly if it helps
05:47 fdobridge_: <m​henning> (or if the shader in question doesn't use ldg, stg, or atomg, then it's something different)
05:55 fdobridge_: <g​fxstrand> Yeah, the shader doesn't touch memory. In fact, it crashes mid-MME, so... Nothing to do with shaders. 😅
05:55 fdobridge_: <g​fxstrand> Actually... I should look closer at `NVK_DEBUG=vm`... 🤔
05:56 fdobridge_: <g​fxstrand> It could be that the CTS is deleting the command buffer before it's done.
05:56 fdobridge_: <g​fxstrand> It wouldn't have been noticed on most drivers because they reference count somewhere. We just tear down the VA on the spot.
07:08 fdobridge_: <!​DodoNVK (she) 🇱🇹> Is it technically possible (just like Fermi or TeraScale Vulkan)? :cursedgears:
08:40 fdobridge_: <a​irlied> @gfxstrand that is possible problem, some CTS object being cleaned before idle, would match the pain
08:43 fdobridge_: <a​irlied> also https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27255
08:51 fdobridge_: <t​om3026> yay i feel useful beyond my endless nonsense in here, i actually helped find something!
13:20 fdobridge_: <!​DodoNVK (she) 🇱🇹> https://mstdn.games/@ReverseModule/111833704210505779
14:12 fdobridge_: <g​fxstrand> No. No compute shaders.
14:52 fdobridge_: <g​fxstrand> I thought about that but then I remembered I'm running with `NVK_DEBUG=push_sync` so it can't be unbinding while it's running but then I remembered that threaded submit exists but then I remembered that I'm able to successfully dump so my CPU maps must still exist and those get torn down before the GPU unmap. I also just ran w/ valgrind and didn't get any use-after-free so all my CPU data structures still exist all the way through.
14:52 fdobridge_: <g​fxstrand> At this point, I really think we're looking at either a GSP bug or a VM_BIND bug.
14:58 fdobridge_: <!​DodoNVK (she) 🇱🇹> GSP firmware or GSP implementation? :nouveau:
15:00 fdobridge_: <k​arolherbst🐧🦀> could be both
15:38 fdobridge_: <g​fxstrand> https://tenor.com/view/why-not-both-why-not-take-both-gif-11478682
15:43 fdobridge_: <g​fxstrand> That's possible but we've already had bugs with not properly sizing some fixed size array somewhere. This feels a lot like that.
15:52 fdobridge_: <k​arolherbst🐧🦀> ~~just wait until we have our new driver then all those bugs will be fixed~~
15:52 fdobridge_: <k​arolherbst🐧🦀> which in all seriousness it probably indeed the quickest way to get rid of those bugs :ferrisUpsideDown:
15:57 fdobridge_: <S​id> ~~nova?~~
15:59 fdobridge_: <!​DodoNVK (she) 🇱🇹> I doubt that will be upstreamed though
16:07 fdobridge_: <s​amantas5855> I think tesla 1 and 2 has those
16:07 fdobridge_: <s​amantas5855> What if, TesVK
16:09 fdobridge_: <g​fxstrand> If the bug is that we did a calculation wrong, a new driver won't fix that if it uses the same calculation.
16:09 fdobridge_: <!​DodoNVK (she) 🇱🇹> I might have to get a 9500 GT again if that becomes a thing 🔺🔺🔺
16:11 fdobridge_: <s​amantas5855> I have a tesla
16:11 fdobridge_: <s​amantas5855> I think 9600
16:12 fdobridge_: <!​DodoNVK (she) 🇱🇹> Is it a Model GT or Model GTX? 🚀
16:15 fdobridge_: <s​amantas5855> https://cdn.discordapp.com/attachments/1034184951790305330/1201198773750218802/signal-2023-01-04-15-24-10-756.jpg?ex=65c8f288&is=65b67d88&hm=750e7377278baee668f940236142c07a5f0cd815559d929591ebac74fd27188c&
16:15 fdobridge_: <s​amantas5855> here
16:15 fdobridge_: <s​amantas5855> tesla support would unironically be useful since gt210 and gt220 gpus are still sold as display adapters
16:15 fdobridge_: <s​amantas5855> at least here
16:18 fdobridge_: <!​DodoNVK (she) 🇱🇹> This doesn't seem to be a 9600 GT according to images on Google
16:19 fdobridge_: <g​fxstrand> Tesla probably could do very basic Vulkan but that's going to be a different driver.
16:19 fdobridge_: <g​fxstrand> Yeah, Tesla is the new GeForce 2.
16:20 fdobridge_: <g​fxstrand> Well, it would require a very creative kernel driver....
16:20 fdobridge_: <s​amantas5855> if it is enough to run a compositor it would be useful
16:21 fdobridge_: <s​amantas5855> what about fermi?
16:21 fdobridge_: <s​amantas5855> https://cdn.discordapp.com/attachments/1034184951790305330/1201200316155187390/Screenshot_20240128-181835.png?ex=65c8f3f8&is=65b67ef8&hm=25421c048cb1aece87f6d0ae3273e99773d16092f288dadafb96153b4ba6a221&
16:21 fdobridge_: <s​amantas5855> if I go to gpus and sort by price tesla comes first
16:21 fdobridge_: <g​fxstrand> Fermi is possible. It's all about how much work you want to do.
16:21 fdobridge_: <s​amantas5855> is fermi that much different from kepler?
16:22 fdobridge_: <g​fxstrand> Like, as long as you have compute, vertex, and geometry shaders, it's probably possible.
16:22 fdobridge_: <g​fxstrand> It's additional compiler work and the copy engine is different.
16:23 fdobridge_: <g​fxstrand> Worst case, we have to implement copies with shaders like other hardware.
16:23 fdobridge_: <!​DodoNVK (she) 🇱🇹> Fermi is probably more useful as a radiator thanks to its high power consumption /s
16:23 fdobridge_: <g​fxstrand> Note that I didn't say "I". 😝
16:24 fdobridge_: <s​amantas5855> https://tenor.com/view/thermi-nvidia-gpu-580-480-gif-19880982
16:25 fdobridge_: <k​arolherbst🐧🦀> Ben said that we _could_ implement the copy engine via firmware (which we'd have to write ourselves)....
16:25 fdobridge_: <k​arolherbst🐧🦀> or something
16:25 fdobridge_: <m​henning> The gl driver should already be enough to run a compositor on tesla, but it doesn't get much attention any more and might be a little buggy
16:25 fdobridge_: <k​arolherbst🐧🦀> but apparently nvidia has firmware for that as well...
16:25 fdobridge_: <k​arolherbst🐧🦀> we could ask for that (if we don't forget)
16:25 fdobridge_: <k​arolherbst🐧🦀> but...
16:25 fdobridge_: <k​arolherbst🐧🦀> but...
16:25 fdobridge_: <s​amantas5855> but they won't give it?
16:25 fdobridge_: <k​arolherbst🐧🦀> nvk on fermi is pointless
16:26 fdobridge_: <k​arolherbst🐧🦀> nvidia decide that fermi is the gen where they want to boot with the lowest clocks possible, and without reclocking you have GPUs being slower than CPU rendering
16:26 fdobridge_: <k​arolherbst🐧🦀> literally
16:27 fdobridge_: <s​amantas5855> I thought we could reclock fermi
16:27 fdobridge_: <k​arolherbst🐧🦀> well.. maybe not if you use a CPU from the same time :ferrisUpsideDown:
16:27 fdobridge_: <k​arolherbst🐧🦀> nah..
16:27 fdobridge_: <k​arolherbst🐧🦀> there are patches, but it's never been really tested
16:27 fdobridge_: <k​arolherbst🐧🦀> and probably quite a bit of work
16:27 fdobridge_: <s​amantas5855> so kepler and maxwell can be reclocked but fermi not?
16:28 fdobridge_: <k​arolherbst🐧🦀> correct
16:28 fdobridge_: <s​amantas5855> that includes maxwell 1 and 2 right?
16:29 fdobridge_: <k​arolherbst🐧🦀> yes and no
16:29 fdobridge_: <k​arolherbst🐧🦀> maxwell 2 is uhm.. needs signed firmware for fan controls
16:29 fdobridge_: <!​DodoNVK (she) 🇱🇹> At least work is being freed up from the CPU
16:29 fdobridge_: <k​arolherbst🐧🦀> but if the GPU fans are controlled by something else, it's fine
16:29 fdobridge_: <k​arolherbst🐧🦀> we don't support it out of the box tho
16:30 fdobridge_: <s​amantas5855> Any progress on pascal and volta?
16:30 fdobridge_: <k​arolherbst🐧🦀> yeah.. for rendering a desktop it's good enough, well.. if you use a FHD display
16:30 fdobridge_: <k​arolherbst🐧🦀> will never happen
16:30 fdobridge_: <k​arolherbst🐧🦀> Kepler and maxwell1 will probably be the only gens we will reclock before GSP (Turing+)
16:31 fdobridge_: <k​arolherbst🐧🦀> NVK does kinda support it, but it's just equally pointless
16:31 fdobridge_: <!​DodoNVK (she) 🇱🇹> Can't Tesla be reclocked too?
16:31 fdobridge_: <k​arolherbst🐧🦀> yeah.. to some degree
16:31 fdobridge_: <k​arolherbst🐧🦀> some can, some can't
16:31 fdobridge_: <k​arolherbst🐧🦀> it's just that if you support Kepler and Turing+, supporting all the boards in the middle is more or less for free 😄
16:32 fdobridge_: <k​arolherbst🐧🦀> took me a day to fix all volta issues and never has ran nvk on volta before :ferrisUpsideDown:
16:32 fdobridge_: <k​arolherbst🐧🦀> took me a day to fix all volta issues and nobody has ran nvk on volta before :ferrisUpsideDown: (edited)
16:32 fdobridge_: <m​arysaka> I have two of those cuties around still
16:32 fdobridge_: <m​arysaka> https://cdn.discordapp.com/attachments/1034184951790305330/1201203165517860904/IMG_2111.jpg?ex=65c8f69f&is=65b6819f&hm=e2b24a0bd50472991ecccacdf85694fd408c1da5b0792ae9b0a0e321ad63e030&
16:32 fdobridge_: <k​arolherbst🐧🦀> I have the same
16:32 fdobridge_: <!​DodoNVK (she) 🇱🇹> Volta is much more niche than Pascal
16:33 fdobridge_: <k​arolherbst🐧🦀> sure, but also cursed
16:33 fdobridge_: <k​arolherbst🐧🦀> 😄
16:33 fdobridge_: <k​arolherbst🐧🦀> it's like pascal with a new ISA
16:34 fdobridge_: <s​amantas5855> 9500gt?
16:35 fdobridge_: <m​arysaka> 6600LE
16:35 fdobridge_: <s​amantas5855> also damn is that svideo
16:35 fdobridge_: <m​arysaka> I still have the motherboard and SLI module too
16:36 fdobridge_: <s​amantas5855> I have never used sli
16:36 fdobridge_: <s​amantas5855> or crossfire
16:36 fdobridge_: <s​amantas5855> but I remember using these cards on agp 8x
16:44 fdobridge_: <m​arysaka> one of mine is not working anymore but I have a feeling that changing one of those capacitors could help, might be dry considering how old this thing is
16:45 fdobridge_: <m​arysaka> (and looking at the image again, one on that card might be close to dead too)
16:48 fdobridge_: <k​arolherbst🐧🦀> yeah.. same
16:48 fdobridge_: <k​arolherbst🐧🦀> though my Teslas are mostly fine
16:49 fdobridge_: <k​arolherbst🐧🦀> what worries me more are the pre Tesla ones 😄
16:49 fdobridge_: <k​arolherbst🐧🦀> I think half of them are dead
16:49 fdobridge_: <k​arolherbst🐧🦀> maybe it's nouveau being broken
16:49 fdobridge_: <k​arolherbst🐧🦀> but....
16:49 fdobridge_: <k​arolherbst🐧🦀> it works flawlessly one some and random rendering corruption and other funky things on others, where they hit the same path in the driver
16:49 fdobridge_: <k​arolherbst🐧🦀> it's weird 😄
16:49 fdobridge_: <m​arysaka> I think I archived some drivers for Windows for those on the webarchive :P
16:50 fdobridge_: <m​arysaka> (I should really revive my archiver again)
16:50 fdobridge_: <k​arolherbst🐧🦀> uhhh
16:50 fdobridge_: <k​arolherbst🐧🦀> what would I need? Windows 2000?
16:50 fdobridge_: <k​arolherbst🐧🦀> though I guess XP might work 😄
16:50 fdobridge_: <m​arysaka> https://archive.org/details/nvidia-drivers
16:51 fdobridge_: <m​arysaka> Should have everything down to 18x I think
16:51 fdobridge_: <m​arysaka> (and until 2021 as I haven't started the scanner again)
16:51 fdobridge_: <s​amantas5855> I have some REALLY old nvidia drivers on some CDs
16:51 fdobridge_: <s​amantas5855> they are for windows tho
16:51 fdobridge_: <k​arolherbst🐧🦀> `Linux x64 (AMD64/EM64T) Display Driver NVIDIA Certified 304.137` on linux 😛
16:51 fdobridge_: <k​arolherbst🐧🦀> that's not _that_ old
16:51 fdobridge_: <m​arysaka> that's "recent"
16:52 fdobridge_: <k​arolherbst🐧🦀> though mhh
16:52 fdobridge_: <k​arolherbst🐧🦀> maybe that's a tesla one?
16:52 fdobridge_: <k​arolherbst🐧🦀> ahh no
16:52 fdobridge_: <k​arolherbst🐧🦀> FX is even older
16:52 fdobridge_: <k​arolherbst🐧🦀> that's nv30, right?
16:52 fdobridge_: <k​arolherbst🐧🦀> 5 FX I mean
16:52 fdobridge_: <m​arysaka> I *think*
16:53 fdobridge_: <k​arolherbst🐧🦀> "177.67" on "2008.8.19 " 😄
16:53 fdobridge_: <m​arysaka> the oldest driver I have around https://archive.org/details/174-16-geforce-winxp-32bit-english-whql
16:53 fdobridge_: <k​arolherbst🐧🦀> I don't even want to know...
16:53 fdobridge_: <k​arolherbst🐧🦀> but yeah..
16:53 fdobridge_: <m​arysaka> pre 2008 doesn't follow the proper CDN format that NVIDIA uses
16:53 fdobridge_: <s​amantas5855> I have older ones
16:53 fdobridge_: <k​arolherbst🐧🦀> latest driver for nv40 is "304.137 September 19, 2017", which is impressive
16:53 fdobridge_: <s​amantas5855> windows 98, me and stuff
16:53 fdobridge_: <k​arolherbst🐧🦀> that's on linux
16:54 fdobridge_: <m​arysaka> nice
16:54 fdobridge_: <k​arolherbst🐧🦀> 4.x something I guess?
16:54 fdobridge_: <k​arolherbst🐧🦀> yeah.. 4.13 or so
16:54 fdobridge_: <k​arolherbst🐧🦀> maybe 4.12
16:54 fdobridge_: <s​amantas5855> I should archive them
16:54 fdobridge_: <k​arolherbst🐧🦀> depending on luck
16:55 fdobridge_: <k​arolherbst🐧🦀> though I guess 4.4 would be the better choice here because longterm
16:55 fdobridge_: <k​arolherbst🐧🦀> wait...
16:55 fdobridge_: <k​arolherbst🐧🦀> that's still supported :ferrisUpsideDown:
16:55 fdobridge_: <k​arolherbst🐧🦀> ehh wait
16:55 fdobridge_: <k​arolherbst🐧🦀> that got dropped
16:56 fdobridge_: <!​DodoNVK (she) 🇱🇹> I have at least 2 phones with older stock kernel versions than that
16:56 fdobridge_: <k​arolherbst🐧🦀> last 4.4 release is in 2022...
16:57 fdobridge_: <m​arysaka> If I ever find the energy to rewrite my scanner, I will try to catch those too I guess
16:57 fdobridge_: <s​amantas5855> https://cdn.discordapp.com/attachments/1034184951790305330/1201209375956680785/signal-2021-11-24-02-30-12-718.png?ex=65c8fc68&is=65b68768&hm=79dfc94ab47b1e235a67ec2c36ccf262ef7f5abb1159c936a7785afac7db726d&
16:57 fdobridge_: <s​amantas5855> https://cdn.discordapp.com/attachments/1034184951790305330/1201209430000279633/signal-2021-11-24-00-33-01-360.png?ex=65c8fc75&is=65b68775&hm=50608993c53d45da08bf35465806d34a4e3c205d23824286934a3219b78aaeb9&
16:57 fdobridge_: <s​amantas5855> matrox man
16:57 fdobridge_: <s​amantas5855> https://cdn.discordapp.com/attachments/1034184951790305330/1201209549848326144/signal-2021-11-24-20-10-53-847.png?ex=65c8fc92&is=65b68792&hm=54efec39f5d58043f249c91fc6623b1495e052c25efb6a018949d70cadf5a5f1&
16:58 fdobridge_: <!​DodoNVK (she) 🇱🇹> I wonder if it will end up like Bibliogram
16:58 fdobridge_: <k​arolherbst🐧🦀> the time where all GUI was horrible
16:58 fdobridge_: <s​amantas5855> yeah
16:58 fdobridge_: <k​arolherbst🐧🦀> without exceptions. I don't take any questions
16:59 fdobridge_: <k​arolherbst🐧🦀> I mean.. XP was also horrible, but in a way the 2000s deserved
17:00 fdobridge_: <s​amantas5855> xp was color pukw
17:00 fdobridge_: <s​amantas5855> xp was color puke (edited)
17:00 fdobridge_: <m​arysaka> I just hope I will not crash the archiver pool again... Last time I may have uploaded 2k installers and that was representing 50% of the driver collection on the archive.org, so I understand that it might be a bit too much for tapes to catch up 😅
17:00 fdobridge_: <s​amantas5855> have you donated to archive too.
17:00 fdobridge_: <!​DodoNVK (she) 🇱🇹> And also nostalgic
17:00 fdobridge_: <s​amantas5855> have you donated to archive too? (edited)
17:00 fdobridge_: <k​arolherbst🐧🦀> I mean.. the style was unique in its own
17:00 fdobridge_: <k​arolherbst🐧🦀> so it's fine
17:00 fdobridge_: <k​arolherbst🐧🦀> but GUI from the 90s is all terrible
17:00 fdobridge_: <k​arolherbst🐧🦀> 😄
17:01 fdobridge_: <s​amantas5855> I never used windows so I'm thankful I'm not nostalgic for xp
17:01 fdobridge_: <s​amantas5855> we did have 95 at school tho
17:01 fdobridge_: <k​arolherbst🐧🦀> I think what I'm trying to say is, I understand people liking the XP style
17:01 fdobridge_: <s​amantas5855> I'm nostalgic for ubuntu with gnome 2
17:01 fdobridge_: <k​arolherbst🐧🦀> mhhh
17:02 fdobridge_: <k​arolherbst🐧🦀> I mean.. unity the X11 compositor was kinda cool
17:02 KitsuWhooa: I still use unity :p
17:02 KitsuWhooa: (also technically compiz is the compositor and unity is a plugin for it)
17:04 fdobridge_: <s​amantas5855> not really
17:04 fdobridge_: <s​amantas5855> everyone hated it back then
17:04 fdobridge_: <s​amantas5855> also perf was bad
17:04 fdobridge_: <s​amantas5855> it even had amazon ads
17:05 fdobridge_: <s​amantas5855> When unity came out I think I had a pentium 2 or 3
17:12 waterisnice: so the main routine 524 624 724=1872 624+624+512=1760 so is 724+724+512=1960 and 12+12=24 that 288+88+24 so what another buffer expects is 224+88+488 in other words 724+724-1024=424, so if you generate 288 instead of 224 you get 512 to eliminate, if 224 its sums as 448, so three times that value 672-448 comes back as 224, 224+500 is 724, for an example 724 400 300 would use the same logic, after eliminating the stuff 400 and 300 that combined
17:12 waterisnice: to 1024, 724+724=1448/2 is 724. Such routines are for access and are generated by compiler, queue jumping is with indexes which i talk about next time after evading the ban again.
17:21 fdobridge_: <g​fxstrand> Ah... Compiz...
17:30 fdobridge_: <!​DodoNVK (she) 🇱🇹> That Estonian is at least nice in this message (I wonder when that person started doing weird stuff)
17:38 fdobridge_: <p​avlo_it_115> Hope dies last)
17:45 fdobridge_: <p​avlo_it_115> Wtf
17:45 fdobridge_: <p​avlo_it_115> https://cdn.discordapp.com/attachments/1034184951790305330/1201221635181641858/IMG_20240128_194408.jpg?ex=65c907d3&is=65b692d3&hm=a33525c93b78048f1d369cf182a7236ba60060c489603e1f7bf868e964bcb5a5&
17:47 fdobridge_: <p​avlo_it_115> Wtf) (edited)
17:47 fdobridge_: <p​avlo_it_115> https://cdn.discordapp.com/attachments/1034184951790305330/1201221635181641858/IMG_20240128_194408.jpg?ex=65c907d3&is=65b692d3&hm=a33525c93b78048f1d369cf182a7236ba60060c489603e1f7bf868e964bcb5a5&
17:47 fdobridge_: <p​avlo_it_115> Maybe xf86-video-modesettings crash this
17:49 soreau: wayfire is probably the best bet for 'compiz on wayland'
17:50 fdobridge_: <p​avlo_it_115> Maybe xf86-video-modesettings crash this
17:50 fdobridge_: <p​avlo_it_115> Yes! (edited)
17:50 fdobridge_: <k​arolherbst🐧🦀> somebody needs to do this, and it won't be any of the current nouveau developers 😛
17:51 fdobridge_: <p​avlo_it_115> Were there any attempts to bypass the firmware?
17:52 fdobridge_: <p​avlo_it_115> I don't remember very well, but someone here did that. and added "nvidiaboost = 0-2" or something like that to the kernel parameters
17:57 fdobridge_: <k​arolherbst🐧🦀> that has nothing to do with the firmware
17:57 fdobridge_: <k​arolherbst🐧🦀> that's just how power management works
17:58 fdobridge_: <k​arolherbst🐧🦀> 0 (the default) just means safe clocks you probably will never hit thermal limits with
17:58 fdobridge_: <k​arolherbst🐧🦀> 1 is generally something GPUs are marketed with as their block and 2 is "until thermal limits"
18:28 fdobridge_: <t​om3026> If thats the same person as "marrt" or similiar nick it was going on in 2015 or so
18:31 fdobridge_: <g​fxstrand> Yeah, he's been around a long time
18:32 fdobridge_: <p​avlo_it_115> Do you have a list of VK_EXT/VK_KHR extensions that are in the works/already done for specific chip generations?
18:35 fdobridge_: <p​avlo_it_115> Do you have a list of VK_EXT/VK_KHR extensions that are in the works/already done for specific chip generations? If there are, look for them on gitlab? (edited)
18:41 fdobridge_: <g​fxstrand> nvk_physical_device.c is pretty easy to read. Otherwise, there's the GitLab issue tracker.
18:47 fdobridge_: <g​fxstrand> And of course features.txt and mesamatrix if you're looking for a quick 1000 mile view.
18:48 liquidben: so first let's elaborate how those numbers are generated, easy -- 224 comes as 624-400 424 724-300 24 as 524-500 288 comes as 512-224 etc, so all is self contained, but the alpha state compiler is still months away since i rarely work on it, and having some other issues still, but i do work on it whenever i get time, hope to make a prototype this year. The talks about order vs queue jumping conditionals etc and index generation are
18:48 liquidben: the next things up tomorrow, still testing. but same note goes for avoiding the division on 724 line, and same logic means 724-512 is 212 times to from times three gets single without division, i had success on queue jumping too but i test a bit more how to describe this on another day, cheers. that too and off to bed now.
18:52 fdobridge_: <g​fxstrand> In terms of plans, the closest thing we have is the issue tracker. I think we've got most of the NVK and Zink features ticked off but there's still some test failures with Zink that need to be triaged. Then VKD3D is next.
18:58 fdobridge_: <p​avlo_it_115> thx
19:02 fdobridge_: <p​homes_> wrt plans I am thinking about trying VK_KHR_performance_query unless that conflicts with anyone else's plans?
19:23 fdobridge_: <a​irlied> Zcull is out there if someone wants to make a real difference 🙂
21:01 fdobridge_: <k​arolherbst🐧🦀> I wonder if we want allow yuzu to send NVK the plain GPU binaries 😄 though that's going to be pain in regards to binding all the resources properly
21:23 liquidben: Short rant! People tend to get insulted that some courageous bloke comes to talk who never lift any technological weights , so i comment i actually did not intend to fire anyone who is not assaulting me, and is the system bad, absolutely no, it's incredibly correct, cause essentially the same back frame or so to speak exactly the all invented methods work in flow but they tick inside compressed, so i can not be able to fire anyone either even
21:23 liquidben: if i wanted to, cause essentially i am using their code, but i stash it with a trick, i mean really the science has been amazing in computer tech. The compiler and hw is all very correct. So linux is a wonder too or unixes, so that is the short thing, IT's exactly the same code that runs in just compressed format. I AM nobody to tell you what to do, by law i have human and etiquette honoring rights, for an example having gotten assaulted behind
21:23 liquidben: the back and all the abuse is not something i can like, but systems is another thing, this i approve and state is very correctly done, just modern ways none wants to share as of yet. So this was short rant, back to sleep. world is perhaps at war but this is not a war to me, i do not disrespect the computer science at all, i approve all of that being above my capabilities that has been achieved. back to sleep.
22:26 fdobridge_: <g​fxstrand> You're welcome to try! That would actually help a lot of we can get some perf counters.
22:29 fdobridge_: <g​fxstrand> Meh. Most desktop GPUs are going to be so much more powerful than a switch that a little more ALU won't matter much. I have thought about it for the sake of swapping out NAK shaders for blob shaders
22:29 fdobridge_: <g​fxstrand> Yeah, I should give that a crack once I've got 1.3 conformance
22:30 fdobridge_: <k​arolherbst🐧🦀> yeah.. not saying it's useful 😄
22:30 fdobridge_: <k​arolherbst🐧🦀> but it can reduce the CPU overhead a lot
22:30 fdobridge_: <k​arolherbst🐧🦀> and might allow to run it on tegra based devices...
22:31 fdobridge_: <k​arolherbst🐧🦀> nintendo is goint to ~~kill~~ sue us
22:37 fdobridge_: <g​fxstrand> Yeah. I'm curious to know how close we can get to the original shaders with just NIR optimizing things. Same thought with consuming PTX.
22:38 fdobridge_: <k​arolherbst🐧🦀> I'm more concerned about CPU overhead here
22:39 fdobridge_: <k​arolherbst🐧🦀> like... the dolphin folks were able to do a massive workaround to get rid of mid-game stuttering, because they had a full hardware emulation shader they could run in place.. the yuzu folks probably don't have that luxury ever...
22:40 fdobridge_: <k​arolherbst🐧🦀> so instead of parsing/translating/compiling the entire shader, you just run it
22:44 fdobridge_: <g​fxstrand> Yeah...
22:45 fdobridge_: <a​irlied> Won't they need a recompile on most NVIDIA hw though?
22:45 fdobridge_: <k​arolherbst🐧🦀> sure
22:45 fdobridge_: <k​arolherbst🐧🦀> but not on one generation 😄
22:45 fdobridge_: <k​arolherbst🐧🦀> I'm not saying it's very useful
22:45 fdobridge_: <k​arolherbst🐧🦀> however, it has its niche use case 😄
22:45 fdobridge_: <g​fxstrand> If we could disassemble to NAK and reassemble, it might be possible...
22:46 fdobridge_: <a​irlied> Like emulating a switch on a switch is hardly interesting, well maybe at little,
22:46 fdobridge_: <k​arolherbst🐧🦀> could buy a jetson
22:46 fdobridge_: <a​irlied> And if it targets Maxwell our reclocking sucks
22:46 fdobridge_: <k​arolherbst🐧🦀> ~~people could sell knock-off switches for half the price and get sued into oblivion~~
22:46 fdobridge_: <k​arolherbst🐧🦀> ahh, so here is the thing
22:47 fdobridge_: <k​arolherbst🐧🦀> on tegra that kinda works
22:47 fdobridge_: <k​arolherbst🐧🦀> well..
22:47 fdobridge_: <k​arolherbst🐧🦀> not dynamically
22:47 fdobridge_: <k​arolherbst🐧🦀> however
22:47 fdobridge_: <k​arolherbst🐧🦀> one could fix the stuff
22:47 fdobridge_: <a​irlied> Will anyone though? It's been fixable for years
22:47 fdobridge_: <k​arolherbst🐧🦀> in my fantasies somebody will
22:47 linkmauve: I haven’t followed much, what is the state of NVK on the current 6.7 on the Switch?
22:48 fdobridge_: <a​irlied> When they fix it the can write the Vulkan extensik
22:48 linkmauve: Running Switch games without rebooting and without wasting CPU cycles to disassemble/recompile would be useful I think.
22:48 fdobridge_: <a​irlied> When they fix it the can write the Vulkan extension (edited)
22:48 fdobridge_: <k​arolherbst🐧🦀> look.. if somebody wants to sell knock-off switches, they might even bother to fix it 😄
22:49 fdobridge_: <d​adschoorse> why would you use maxwell tegra for that though
22:49 fdobridge_: <a​irlied> I don't think there is much value in us fixing it without a clear use case, just because we could, writing Vulkan exts it pretty easy
22:53 fdobridge_: <g​fxstrand> As an interesting compiler project, I think you can translate most of Maxwell to Turing without redoing RA. There's a couple instructions that might be tricky.
22:53 fdobridge_: <k​arolherbst🐧🦀> mhhhh
22:53 fdobridge_: <k​arolherbst🐧🦀> bitfield stuff is different
22:53 fdobridge_: <k​arolherbst🐧🦀> like that insert/extract stuff
22:53 fdobridge_: <g​fxstrand> Yeah. That's the tricky bit
22:54 fdobridge_: <k​arolherbst🐧🦀> and that texgrad lowering is going to be a pain...
22:54 fdobridge_: <k​arolherbst🐧🦀> and apparently some games hit that
22:54 fdobridge_: <g​fxstrand> Rearranging texture sources would be tricky
22:55 fdobridge_: <k​arolherbst🐧🦀> nah, that's the trivial part :ferrisUpsideDown:
22:55 fdobridge_: <g​fxstrand> Well yeah.
22:55 fdobridge_: <k​arolherbst🐧🦀> texgrad lowering is that funky "put shader into quad more and to quad ops" thingy
22:55 fdobridge_: <k​arolherbst🐧🦀> *mode
22:55 fdobridge_: <k​arolherbst🐧🦀> and it uses the barrier stuff on volta+
22:56 fdobridge_: <k​arolherbst🐧🦀> mhh.. though might be possible
22:56 fdobridge_: <k​arolherbst🐧🦀> dunno
22:56 fdobridge_: <p​ac85> This recompiling stuff is something Nintendo themselves will have to do to run switch games on switch2 isn't it?
22:57 fdobridge_: <g​fxstrand> Yes if they let you plug in the same cartridge.
22:57 fdobridge_: <k​arolherbst🐧🦀> they could also just ship two SoCs 😄
22:57 fdobridge_: <g​fxstrand> But if they are in the loop, they can recompile or transpile offline.
22:57 fdobridge_: <k​arolherbst🐧🦀> yeah...
22:57 fdobridge_: <g​fxstrand> That's been the way in the past...
22:57 fdobridge_: <k​arolherbst🐧🦀> I know
22:58 fdobridge_: <k​arolherbst🐧🦀> but the switch actually has working game updates in place
22:58 fdobridge_: <k​arolherbst🐧🦀> sooo....
22:58 fdobridge_: <k​arolherbst🐧🦀> but I guess nvidia could also provide a transpiler.. I at least don't think games actually contain command buffers and it all goes through some API, but not quite sure
22:58 fdobridge_: <g​fxstrand> Yup
22:59 fdobridge_: <k​arolherbst🐧🦀> not sure if yuzu needs to emulate that low-level...
22:59 fdobridge_: <k​arolherbst🐧🦀> but I think they have to...
23:00 fdobridge_: <p​ac85> Mmm, but it's one thing to put a tiny gameboy arm CPU in the Ds, it's another thing to have to tegras
23:00 fdobridge_: <p​ac85> Mmm, but it's one thing to put a tiny gameboy arm CPU in the Ds, it's another thing to have two tegras (edited)
23:00 fdobridge_: <k​arolherbst🐧🦀> you can turn the other off
23:00 fdobridge_: <k​arolherbst🐧🦀> it's not really a space issue, just a cooling one
23:00 fdobridge_: <p​ac85> The user space driver is statically linked with the game I think
23:00 fdobridge_: <p​ac85> I'd guess yuzu does interpret command buffers
23:00 fdobridge_: <k​arolherbst🐧🦀> yeah...
23:00 fdobridge_: <k​arolherbst🐧🦀> probably
23:01 fdobridge_: <p​ac85> That's what all emulators do, emulating at an api level has never been a thing AFAIK
23:01 fdobridge_: <k​arolherbst🐧🦀> might make more sense to just target the UAPI directly 😄
23:01 fdobridge_: <p​ac85> Yeah exactly
23:02 fdobridge_: <p​ac85> You emulate at the kernel level and just run all of the userspace stuff
23:24 fdobridge_: <g​fxstrand> Not really. If you take the 20 nm X1 design and fab it at 5nm, it'll sip power.
23:25 fdobridge_: <k​arolherbst🐧🦀> tho they already shrinked it to 16nm
23:25 fdobridge_: <g​fxstrand> And you can just power it off when playing switch 2 games and run the bigger one low power for switch 1 games.
23:25 fdobridge_: <p​ac85> I suppose you really only need the gpu
23:26 fdobridge_: <p​ac85> I don't think games have low level access to anything else
23:26 fdobridge_: <g​fxstrand> Or they could potentially make a GPU that supports both instruction sets.
23:26 fdobridge_: <k​arolherbst🐧🦀> cursed
23:26 fdobridge_: <k​arolherbst🐧🦀> but...
23:26 fdobridge_: <k​arolherbst🐧🦀> the sub-channels have the API for that :ferrisUpsideDown:
23:26 fdobridge_: <p​ac85> Yeah cursed
23:27 fdobridge_: <k​arolherbst🐧🦀> I wonder if they added it because of stuff like that
23:27 fdobridge_: <k​arolherbst🐧🦀> though I have _no_ idea if it works at all
23:27 fdobridge_: <g​fxstrand> Or put it in the shader header or something
23:27 fdobridge_: <k​arolherbst🐧🦀> yeah...
23:27 fdobridge_: <g​fxstrand> It depends on how much custom chip design they're willing to pay NVIDIA for
23:27 fdobridge_: <p​ac85> Have two decoders in the shaders?
23:28 fdobridge_: <p​ac85> Sounds messy
23:28 fdobridge_: <k​arolherbst🐧🦀> they'll probably just put two SoCs on that thing...
23:28 fdobridge_: <g​fxstrand> Probably
23:28 fdobridge_: <p​ac85> I think the most likely thing is they use sw emulation
23:29 fdobridge_: <k​arolherbst🐧🦀> you just mux the peripherals and boot it up on demand or something...
23:29 fdobridge_: <g​fxstrand> Or that
23:29 fdobridge_: <k​arolherbst🐧🦀> yeah...
23:29 fdobridge_: <k​arolherbst🐧🦀> tho that's kinda expensive
23:29 fdobridge_: <k​arolherbst🐧🦀> I mean...
23:29 HdkR: What's some dark silicon hanging off the AXI between friends? :P
23:29 fdobridge_: <k​arolherbst🐧🦀> they do ship emulators 😄
23:29 fdobridge_: <g​fxstrand> Now that we're running real GPUs in consoles, emulation is a lot more practical.
23:29 fdobridge_: <k​arolherbst🐧🦀> not sure they have a GC emulator already on the switch?
23:30 fdobridge_: <p​ac85> Yes
23:30 fdobridge_: <k​arolherbst🐧🦀> I know there is the N64 one
23:30 HdkR: NGC emulator was used for the Mario 3D collection
23:30 fdobridge_: <k​arolherbst🐧🦀> ahhh
23:30 fdobridge_: <p​ac85> The super Mario galaxy ports run in sort of emulation
23:30 fdobridge_: <k​arolherbst🐧🦀> but that's like...
23:30 fdobridge_: <k​arolherbst🐧🦀> not pure emulation, is it
23:30 fdobridge_: <p​ac85> Not sure about the details
23:31 HdkR: For Mario Sunshine it's theoretically pure emulation, Mario Galaxy has been reported to just be a GPU emulator with the rest recompiled or something
23:31 fdobridge_: <p​ac85> I think they had an n64 emulator on the Wii btw
23:31 fdobridge_: <k​arolherbst🐧🦀> I even own the 3D collection :ferrisUpsideDown:
23:31 fdobridge_: <p​ac85> Actually I know they did
23:31 fdobridge_: <k​arolherbst🐧🦀> yeah.. recompiling is way more sane
23:32 fdobridge_: <p​ac85> I guess the flipper is peculiar enough that porting it made little sense
23:32 HdkR: Emulating a 485Mhz PowerPC CPU isn't too hard :P
23:32 fdobridge_: <p​ac85> I guess the flipper is peculiar enough that porting thay part made little sense (edited)
23:32 fdobridge_: <k​arolherbst🐧🦀> well..
23:33 HdkR: Especially with games that don't push the CPU and GPU to the maximum, which is nicer. Dolphin has had "idle skip" forever
23:34 fdobridge_: <k​arolherbst🐧🦀> right...