ChanServ changed the topic of #lima to: Development channel for open source lima driver for ARM Mali4** GPUs - Kernel has landed in mainline, userspace driver is part of mesa - Logs at https://people.freedesktop.org/~cbrill/dri-log/index.php?channel=lima and https://freenode.irclog.whitequark.org/lima - Contact ARM for binary driver support!
jrmuizel has joined #lima
<anarsoul> enunes: rellla: I've just noticed that we're not using tiled textures
<anarsoul> it's just one line change and as result we get +10fps in q3a
<anarsoul> also looks like texture filtering is different for linear textures
<anarsoul> we need to RE it
jrmuizel has quit [Remote host closed the connection]
jrmuizel has joined #lima
megi has quit [Ping timeout: 245 seconds]
_whitelogger has joined #lima
jrmuizel has quit [Remote host closed the connection]
jrmuizel has joined #lima
jrmuizel has quit [Remote host closed the connection]
dddddd has quit [Remote host closed the connection]
mardestan has joined #lima
mardestan has quit [Ping timeout: 245 seconds]
Barada has joined #lima
mardestan has joined #lima
libv_ has joined #lima
libv has quit [Ping timeout: 240 seconds]
raimo has joined #lima
mardestan has quit [Read error: Connection reset by peer]
raimo has quit [Remote host closed the connection]
mardestan has joined #lima
<mardestan> CUDA and OPENCL are not so hard! Basically everything is opened up methodically by reading NVIDIA ptx docs and combining it with reading miaow and some stackoverflow comments about CUDA streams.
<mardestan> the maximum number of CUDA streams on the cards, which neccessarily do not run concurrently though, should also point to length of queue entries actually the way i read from in the middle of the lines.
<mardestan> I took my time to become an experts on this which i today definitely am, my eyes would not skip any of the more important stuff. Anger spikes is the bigger issue due to some circmustances in my private path, lots of people who know me they anticipate that once i take something on I will also succeed, I may not be of higher intelligence in anywhere, but those things i know
<mardestan> So you can freely trust me on those theories , I am confident today in what i say.
<mardestan> so here anarsoul talked about drm scheduler, which i have not looked into in code grounds.
<mardestan> but this type of cpu scheduling in the driver for the command queues has two missions, to make context switches and reset the gpu and ban the dirty contexts
<mardestan> in graphics you do not need to touch this, since context switches happen in frame granularity
<mardestan> but when mining some bitcoins i.e running the compute kernels, one may want to occasionally not to switch out contexts then either some priority needs to be used or indeed some patch done against, but in graphics we may find useful to switch contexts a little differently
<mardestan> but implementing this all should not be very complex
mardestan has quit [Quit: Leaving]
kaspter has quit [Quit: kaspter]
megi has joined #lima
libv_ is now known as libv
dddddd has joined #lima
<rellla> wtf? did i mess up my mesa setup?
<rellla> libEGL warning: MESA-LOADER: failed to open lima (search paths /opt/prefix/lib/aarch64-linux-gnu)
<rellla> piglit: error: waffle_display_connect failed due to WAFFLE_ERROR_UNKNOWN: eglInitialize failed with error EGL_NOT_INITIALIZED(0x3001)
<rellla> hm, have to look into it later...
megi has quit [Ping timeout: 246 seconds]
jrmuizel has joined #lima
megi has joined #lima
xdarklight has quit [Quit: ZNC - http://znc.in]
xdarklight has joined #lima
jrmuizel has quit [Remote host closed the connection]
Barada has quit [Quit: Barada]
jrmuizel has joined #lima
adjtm has quit [Ping timeout: 245 seconds]
<rellla> :(
megi has quit [Ping timeout: 265 seconds]
cp- has quit [Quit: Disappeared in a puff of smoke]
<anarsoul> do clean build?
<anarsoul> bisect if it still doesn't work
cp- has joined #lima
<rellla> anarsoul: shouldn't be enough to delete build dir?
<anarsoul> 'ninja -C build clean' would also work (and won't wipe your configuration)
adjtm has joined #lima
<rellla> doesn't work. i updated the kernel, but it worked since then. iirc i didn't reboot, just updated mesa ...
<rellla> i will clone again and do a clean build. checking a commit, that worked before...
<enunes> my automated run last night worked, so I think mesa is not broken
<rellla> enunes: thats what i think, too. though i wonder, because iirc i didn't change anything ...
<jernej> anarsoul: lima git master from yesterday doesn't even work with kodi, I get EGL_BAD_SURFACE from eglSwapBuffers
<jernej> but kmscube works fine
<jernej> and dmesg is full of "[drm:drm_internal_framebuffer_create] could not create framebuffer"
<jernej> that's with H3
<jernej> any idea?
<rellla> ftr... lima has to be card0 or card1?
<enunes> jernej: did kodi ever work better than that? if so bisect would be helpful
<jernej> of course, at one point it was almost fine
<jernej> but currently I don't see anything on screen
<enunes> yeah so bisect is the most interesting information
<jernej> with current setup it will take ages
<enunes> I wonder if it is the same issue that affects glmark2
<enunes> (which is a regression in mesa, not in lima)
<rellla> ok. so an ooold one works: fbd5d9eb
<rellla> current master is broken.
<rellla> bisecting now ...
<jernej> ah, good to know :)
<rellla> good to have a cross compile environment here btw :)
<jernej> I have it, but 7+ years old laptop with slow update system doesn't help
<anarsoul> jernej: I'm still using x230 :)
<anarsoul> technically it's 7 yo laptop
<jernej> yes, but you didn't do any LibreELEC development, right?
<anarsoul> nope, just kernel and lima
<jernej> every time you make a image, it creates squashfs image from scratch, which takes about a minute
<jernej> and then you have to upload update to board and reboot
<jernej> where update procedure extracts update, check checksum and copy over squashfs file, kernel and dtb
<jernej> so it takes time to do one round
<anarsoul> ouch
<jernej> I guess I could do some hacks to speed up, but still it wouldn't be quick
<anarsoul> yeah, I guess rellla can do it faster
<rellla> but has to restart, because i mistyped good/bad :)
<jernej> I can play with other drivers in the mean time :)
<anarsoul> can't be too many commits
<anarsoul> rellla: btw 32ea4c2c5ea461659a5b779dfe6d6b1334747ee6 still works for me
<jernej> does that have most lima improvements?
<rellla> could that be sth llvm related ?
<anarsoul> rellla: nope
<anarsoul> jernej: yes
<rellla> ... give me 5 mins and i'll give you the commit :)
<rellla> ok, it's that one:
<anarsoul> report it on #dri-devel?
<rellla> ...manual check
<rellla> yes, it's the first bad one.
<anarsoul> try reverting it?
<anarsoul> and don't forget to report it on #dri-devl
<anarsoul> *devel
<jernej> anarsoul: your commit still doesn't work for me
<anarsoul> :\
<jernej> does this configuration looks sane to you? https://pastebin.com/raw/2dd3Leu4
<anarsoul> can you test the same commit on your laptop with I presume i965?
<anarsoul> seems to be OK
<jernej> I'm not sure if it's ok to have no dri drivers
<jernej> or is that only for x11?
<rellla> jernej: that's mine... https://pastebin.com/7pzxT3iV
<jernej> ok, then I guess it's fine
<rellla> voila, git revert 2c1983f7 and mesa is back again :)
<rellla> time to make a break :p
<jernej> anyway, thanks for help, but my focus is currently elsewhere, this was just drive-by attempt to see current state
<jernej> so I'll try again at later time
<enunes> it seems that I'm not able to reproduce that, mesa master works for me with xorg and glmark2, glmark2-es2-drm, kmscube
<jernej> just a thought, does anyone test this on 32-bit SoC?
<enunes> yeah my automated nightly piglit test runs on 32-bit and passed too
<rellla> anyone with aarch64?
<enunes> I just tested on A64 and it's working for me
<rellla> it's mali450 on H5 btw
<rellla> :( so it's me
<jernej> although this message: "[drm:drm_gem_fb_create_with_funcs] Failed to lookup GEM object"
<jernej> might suggest something in the kernel
<enunes> jernej: which kernel is this?
<jernej> 5.3-rc8
<jernej> glmark2-es2-drm doesn't work
<jernej> eglCreateWindowSurface failed with error: 0x3009
<enunes> this message with glmark2-es2-drm mesa master and unmodified 5.3-rc8?
<jernej> define unmodified :)
<jernej> it has patches, but nothing regarding lima or display driver
<enunes> ok
<enunes> which glmark2-es2-drm scene?
<jernej> just running without any parameters
<anarsoul> jernej: it needs "--visual-config stencil=1:alpha=0" argument
<anarsoul> it's not related to lima and broke somewhere in mesa for panfrost as well not so long ago
<jernej> that help
<jernej> *helps
<anarsoul> I'm not sure why kmscube works
<enunes> kernel message is still strange though
<anarsoul> (my guess that alpha=0 is only relevant thing)
<jernej> shading=phong -> seg fault
<enunes> that one is known
<anarsoul> enunes: btw, do you remember why we disabled tiling for textures?
<enunes> anarsoul: I didn't follow this discussion
<enunes> we had it enabled and disabled it?
drod has joined #lima
<anarsoul> well, prior to going mainline
<anarsoul> I tried enabling it and it works
<anarsoul> gives 10fps boost in q3a
<anarsoul> (20-30fps without tiling 30-40fps with tiling)
<enunes> if no piglit regressions... seems good
<anarsoul> haven't tested it
<anarsoul> busy with work this week
<enunes> yeah same, this weekend I might have some time to work on something
<anarsoul> enunes: also looks like texture filtering differs for linear and tiled textures
<anarsoul> and our code works only for tiled
<anarsoul> e.g. mipmap filtering doesn't work (mipmaps work fine though) - easy to notice in q3a
<anarsoul> I guess I need to ask yuq how we made blob to use linear textures
<jernej> anarsoul: This commit message seems interesting: https://github.com/Miouyouyou/glmark2/commit/b7cce86e830ce39f3642a48dd520744af4205391
drod has quit [Read error: Connection reset by peer]
<jernej> so issue also seems to be with binary drivers
<anarsoul> yeah, looks like mesa now has the same issue
<anarsoul> jernej: I wonder if glmark2 tries to use vi plane that doesn't have alpha
drod has joined #lima
<jernej> that could be also the reason, but that should be easy to check, right?
<jernej> anarsoul: just for test, you can enable formats with alpha channel in kernel for VI plane
<jernej> it will work, HW just ignores alpha
<anarsoul> jernej: I'm at work now
<jernej> ok, I guess I can try quickly
<jernej> anarsoul: you're right, this patch allows running glmark2 without any parameters: http://ix.io/1V0C
<jernej> e.g. by adding formats with alpha channel to VI plane
<anarsoul> jernej: well, then I guess glmark2 should be patched?
<jernej> so glmark2 plane selection should be improved
<jernej> yes
<anarsoul> does it also fix kodi?
<jernej> no, Kodi has good plane selection method
<anarsoul> then I wonder why it doesn't work :(
<jernej> drmModeAddFB2 is failing
<jernej> so maybe some corruption of GEMs?
<jernej> [drm:drm_gem_fb_create_with_funcs] Failed to lookup GEM object
<anarsoul> are you creating dumb bo with render node or with display node?
<anarsoul> IIRC sun4i-drm needs contiguous BO
<anarsoul> but lima doesn't need it (it has its own MMU) and thus BOs allocated by lima aren't contiguous
<jernej> card0, so by display node
<anarsoul> jernej: I don't think that lima is at play at this time
<jernej> could be and I tried a git hash which is known to work with amlogic
<anarsoul> oh, panfrost just got job dependencies and it doesn't flush on fb switch anymore
<jernej> anyway it's just strange that it works with binary driver well
<anarsoul> rellla: I can merge your undef MR if you add Erico's r-b
<anarsoul> jernej: make sure that card0 is display node and not lima
<anarsoul> it can be lima if both drivers compiled as modules
<jernej> I did, I checked sysfs 5min ago
<jernej> and only lima is compiled as a module
<anarsoul> OK
adjtm has quit [Ping timeout: 276 seconds]
<rellla> anarsoul: i see if i can motivate myself to do it later...
<anarsoul> :)
<rellla> jernej, anarsoul: is my issue related to jernej's? haven't backread the whole log...
<anarsoul> probably not
megi has joined #lima
<rellla> ok. then i will first do some more checks (also on H3/A20) and different kernels, to see how it's reproducible. .... before i pop up in #dri-devel again ;)
<rellla> anarsoul: pushed it again.
nerdboy has quit [Ping timeout: 245 seconds]
nerdboy has joined #lima
<enunes> I clicked to merge it after the pipeline
<rellla> thanks.
megi has quit [Quit: WeeChat 2.6]
megi has joined #lima
adjtm has joined #lima
nerdboy has quit [Excess Flood]
romainmahoux[m] has quit [Remote host closed the connection]
bshah|matrix has quit [Write error: Connection reset by peer]
Danct12 has quit [Write error: Connection reset by peer]
nerdboy has joined #lima
z3ntu has quit [Write error: Connection reset by peer]
nerdboy has quit [Changing host]
nerdboy has joined #lima
dllud has quit [Ping timeout: 240 seconds]
<anarsoul> jernej: I think I can try installing kodi-gbm if it's in archlinux arm repo
<jernej> ah, arm repo
<anarsoul> it's x86_64 :)
bshah|matrix has joined #lima
<jernej> isn't as simple as using same script to build it for arm with makepkg?
dllud has joined #lima
<anarsoul> jernej: I'm using pine64, it may take ages :)
<jernej> distcc?
<anarsoul> maybe one day
<anarsoul> I've had some issues with it last time so I resorted to using native compilation
<jernej> it seems that alarm doesn't have gbm package
<anarsoul> yeah :(
<anarsoul> also they don't support a lot of allwinner-based boards officially
Danct12 has joined #lima
z3ntu has joined #lima
romainmahoux[m] has joined #lima
dllud has quit [Ping timeout: 250 seconds]
dllud has joined #lima
jrmuizel has quit [Remote host closed the connection]
<anarsoul> enunes: rellla: if you haven't started working on indirect uniform and varyings load for ppir - I'll start working on it this weekend
<anarsoul> also we should probably ping danvet for merging this: https://patchwork.kernel.org/patch/11009715/
<anarsoul> but he's offline currently