11:35 <karolherbst> I've heard you have the same issue as we have with tegra that userspace not using modifiers is quite busted, is that correct? In any case, I opened a feature request against mutter to enable their kms modifier support on some devcices where it is actual required and things won't work without it: https://gitlab.gnome.org/GNOME/mutter/-/issues/1408

11:36 <karolherbst> I was wondering if that's the actual case with etnaviv as well or not?

11:36 <mntmn> karolherbst: OOC, what kind of visible problems does that cause?

11:36 <karolherbst> mntmn: the screen is completly black when starting mutter :)

11:36 <mntmn> ah lol!

11:37 <karolherbst> yeah..

11:37 <karolherbst> the issue is we can't render into a linear depth buffer

11:37 <mntmn> i've seen that only when using Xorg

11:37 <karolherbst> sooo... things just break if we get one

11:37 <mntmn> like, xorg+modesetting+etnaviv

11:37 <karolherbst> mntmn: yeah.. it's an issue with 1.20

11:37 <karolherbst> it should work on 1.21

11:37 <karolherbst> I know that weston just works as it uses modifier

11:37 <karolherbst> but mutter does not

11:38 <mntmn> yes, all wayland compositors work for me on etnaviv, incl. mutter when started in wayland mode

11:38 <karolherbst> ohh, interesting

11:38 <karolherbst> maybe etnaviv does something inside mesa so the issue is not visible.. mhh

11:38 <karolherbst> are you doing some magic detiling stuff or something?

11:38 <karolherbst> I know that I have to enable kms-modifiers in mutter first

11:39 <mntmn> mutter --wayland --display-server works for me

11:39 <mntmn> i think etnaviv does automatic detiling for scanout, but i'm not an expert in that area

11:39 <karolherbst> mhhhh

11:39 <karolherbst> yeah.. we were wondering to add this stuff to the tegra driver as well

11:40 <karolherbst> but using kms modifiers would give better perf anyway

11:40 <austriancoder> karolherbst: If needed we do an extra blit to detile

11:40 <karolherbst> austriancoder: ohh, I see.

11:41 <karolherbst> austriancoder: when are you doing it? We have an MR for tegra do support this but maybe there are some pitfalls we are not aware of

11:41 <karolherbst> https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6426/diffs

11:42 <karolherbst> so essentially we do it at flush_resource

11:57 <lynxeye> karolherbst: you really want some age counters on the resources. flush_resource might be called multiple times without actual rendering happening to the resource

11:57 <lynxeye> so you want to optimize this into only doing the blit when there were changes to your internal shadow

12:05 <karolherbst> yeah.. I think that sounds like a reasonable thing

12:07 <karolherbst> but I was more interested on the high level overview on how the resource are created and when you do the detiling... I mean, I know that the MR seems to work well enough, just wondering if you do it differently or not

12:08 <mntmn> gnome 3.36 (wayland) works perfectly now on etnaviv/gc7000l btw. needs those 2 little fixes to glamor for xwayland apps, otherwise stable and surprisingly fast

12:12 <karolherbst> mntmn: what two fixes?

12:18 <mntmn> karolherbst: i had to remove one glClear from glamor to fix texture/fbo corruption of xwayland windows. and put in one glFinish to fix sync issues that make everything flicker

12:19 <karolherbst> mhh, let's see if I hit the same thing

12:19 <lynxeye> karolherbst: In general etnaviv was in the nice place that we did never support winsys buffers in tiled layout without modifiers. So if we get to import a buffer or create one without a specified modifier we just assume linear and then have internal shadows for actual rednering

12:20 <lynxeye> Not sure how well this translates to tegra

12:20 <karolherbst> well.. the issue with tegra is that tegra is essentially kmsro, and it uses nouveau for the actual operations

12:21 <mntmn> karolherbst: this and the following patch https://source.mntmn.com/MNT/reform-system-image/src/branch/reform2-redo/reform2-imx8mq/mkuserland2.sh#L150

12:23 <mntmn> the glFinish() is a workaround that causes some buffer flushes to hard sync. because apps that draw directly via legacy X apis don't have a sync mechanism with etnaviv

12:23 <karolherbst> right..

12:23 <mntmn> removing the glClear() is a workaround for unknown corruption bug in etnaviv/gc7000l

12:23 <karolherbst> I think I remember sawing a similiar issue,not sure if it was X or xwayland though

12:24 <mntmn> i also found that Xwayland works much better when used in rooted mode. it is tricker when wayland also acts as the window manager

12:26 <austriancoder> karolherbst we are using the renderonly library for scanout buffers (renderonly_create_kms_dumb_buffer_for_resource(..) ) - see etna_resource_alloc(..) and the flush should happen in etna_flush_resource(..)

12:44 <karolherbst> mntmn: mhh.. can I launch mutter from ssh?

13:14 <mntmn> karolherbst: not sure, maybe with openvt?

13:19 <daniels> mntmn: yeah, rooted mode means that you get a real surface with eglSwapBuffers; rootless mode means that the user buffers get directly passed through for GL clients, but for software-rendered clients, you just periodically blit into a shared surface, which is ... a bit of a disaster tbh

13:19 <daniels> but it's the same when running X11 with any other composition manager, not specific to Xwayland

13:21 <mntmn> daniels: btw, remember our little experiment with glFinish? do you think it's realistic to get a patch specific to etnaviv into xorg/glamor?

13:22 <mntmn> i guess the glClear thing has to be fixed on etnaviv side somehow, but for the sync issue i don't know any other solutions

13:23 <daniels> mntmn: I do remember it well (found the branch the other day in fact), but I don't really know how it is we'd detect that we need to be using it

14:38 <marex> lynxeye: did you make any progress on the GPC/PGC ? I'll slowly get back to it todayish

14:39 <marex> I also got no reply from NXP, of course