06:46pabs: (Tegra related)
11:20omeringen: @imirkin @karolherbst
18:46Wally: For envyas, what is the ISA appropriate for TU106?
19:19imirkin: Wally: pretty sure that volta/turing support wasn't added to envyas
19:26Wally: Is there another way assemblies could be assembled then?
19:27Wally: or n/a
19:29imirkin: not that i know of
19:29imirkin: but ... maybe?
19:31Wally: Is there an ISA I can refrence off of then(to add assembler support
19:32imirkin: you can look at the emitter in codegen
19:33imirkin: Wally: https://cgit.freedesktop.org/mesa/mesa/tree/src/gallium/drivers/nouveau/codegen/nv50_ir_emit_gv100.cpp
20:01omeringen: @imirkin hi, is my new kernel 5.17 log useful ?
20:02imirkin: omeringen: haven't looked at it yet
20:44Wally: How are the tables generated eg. https://github.com/YusufKhan-gamedev/envytools/blob/master/envydis/gm107.c#L483
20:45mwk: they're not generated, they're human-made
20:45Wally: How are they made
20:46mwk: mostly by feeding a lot of stuff to nvidia compiler, running disassembler on output, figuring out the unknown instructions
20:46mwk: that + cross-checking with nvidia's own disassembler
20:47Wally: nvidia has a assembler?
20:47mwk: they sure have a disassembler
20:47mwk: didn't have an assembler last I checked (but that was a while ago)
20:49Wally: what are those tables...tabeling then?
20:50mwk: various parts of the instruction encoding
20:51Wally: In what formate
20:55Wally: ie. How are 0x0000000100000000ull, 0x0000000380000000ull, and N("covered") connected to each other
20:55Wally: And how would I make those tables
22:59karolherbst: mwk: well ptx is kind of an assembler :D
23:01karolherbst: Wally: I don't think it makes much sense to add support for a new arch to envydis/envyas now that we have something inside mesa to do this kind of stuf
23:02karolherbst: read up on src/compiler/isaspec/ inside mesa
23:04Wally: Is it ~nvidia ptx?
23:04karolherbst: no, so the idea of isaspec is, that you probably don't want to write assember/disassembler tooling _and_ something for your compiler inside mesa
23:04karolherbst: so you have an XML describing everything making use of a C struct
23:05karolherbst: or something
23:05karolherbst: I never used it, but normally that's what people would use in the future
23:05karolherbst:should add it as a GSoC project idea
23:06Wally: Why isnt xf86-video-nouveau using these shaders?
23:07Wally: initial motivation for envytools extention
23:08karolherbst: because it doesn't make sense to care much about the nouveau ddx besides sentimentality
23:09karolherbst: time is better spend on more useful things tbh
23:12Wally: so isaspec is XML files which are ~ low level assemblies?
23:12karolherbst: isaspec kind of lives inside mesa, because that's the only place where you really need this kind of stuff
23:12Wally: So I cant use it externally
23:13karolherbst: there is no point to use it externally
23:13karolherbst: the nouveau ddx is already a worse user experience, due to many reasons (mainly tearing and performance). And I can't stress enough how much of a waste of time it is to care about it besides supporting very old hardware
23:14karolherbst: just work on more interesting project
23:15karolherbst: regardless of that, using isaspec is probably a good idea anyway
23:15karolherbst: or well.. porting our compiler stack to it
23:15Wally: Then remove the nouveau ddx ;)
23:16karolherbst: we don't have solid GL support on all hardware
23:17karolherbst: old hardware is really the only good reason to use it
23:17Wally: So can I remove the ddx from my system?
23:17Wally: Itll be faster?
23:17karolherbst: depends on the GPU you currently have
23:18karolherbst: then it wouldn't change anything as it's not used there at all
23:18karolherbst: or shouldn't be
23:21karolherbst: Wally: what's the problem you want to solve anyway?
23:21Wally: Shit I forgot
23:21Wally: uh...exa was the initial motivation then there was something else...
23:22karolherbst: exa is broken
23:22karolherbst: the thing is.. the nouveau ddx is not even used for wayland, so it's all kind of a waste of time to care much about any of this
23:22karolherbst: if acceleration appears to be slow, it's probably a better idea to figure out what's slow inside mesa
23:22Wally: wayalnd doesnt work on nouveau devices though
23:23karolherbst: and help speeding up like everything besides just 2D accel in some legacy windowing system nobody cares about anymore
23:23karolherbst: Wally: it does
23:23karolherbst: why shouldn't it?
23:23Wally: How could Mesa be too slow
23:23Wally: wl didnt work for me
23:23karolherbst: many reasons.. maybe we stall somewhere? maybe we block too often? maybe our GPU pipeline is too slow?
23:24karolherbst: Wally: yeah well.. that could be wls fault
23:24Wally: "maybe we stall somewhere" - yes most likely
23:24karolherbst: yeah dunno
23:24karolherbst: it wasn't really slow on any hardware I have here, except you do like dual 4K displays, but then you are screwed by low GPU clocks one way or the other
23:25karolherbst: yeah.. can't do that yet on turing, maybe we will in the future oh well
23:25karolherbst: if wl is crashing or not working with nouveau that's something we should figure out, but I know that gnome just works fine
23:26Wally: gnome is awful though
23:26karolherbst: yeah well.. it does work though
23:26karolherbst: so either wl is buggy or wl triggers a path not triggered by mutter
23:26Wally: mirror of that
23:28karolherbst: Wally: yeah well.. that doesn't really help figuring out what's wrong :/
23:29karolherbst: maybe I take a look tomorrow and see what happens
23:29karolherbst: but from my past experience plasmas wayland support isn't great
23:30Wally: From my experience plasmas X support isnt great either
23:30Wally: prolly worse than wayland now
23:31karolherbst: I ditched it for gnome and things got better :D
23:31karolherbst: gnome has at least solid wayland support
23:31Wally: I ditched gnome for plasma and now I have spent less time debugging kernel panics ;D
23:32karolherbst: anyway, coredumpctl has a gdb command
23:32karolherbst: with that you could print the trace
23:32karolherbst: but it looks like the GPU context crashes or something
23:32karolherbst: we probably do something stupid, maybe it got fixed by a newer mesa
23:36Wally: ---how would I find whats slow in mesa---
23:39karolherbst: yeah... no clue :) So there was a project I was starting to figure at least out if the GPU is busy doing stuff, but....