alyssa changed the topic of #panfrost to: Panfrost - FLOSS Mali Midgard & Bifrost - https://gitlab.freedesktop.org/panfrost - Logs https://freenode.irclog.whitequark.org/panfrost - <daniels> avoiding X is a huge feature
davidlt has quit [Ping timeout: 272 seconds]
raster has quit [Remote host closed the connection]
herbmillerjr has quit [Quit: Konversation terminated!]
herbmillerjr has joined #panfrost
vstehle has quit [Ping timeout: 272 seconds]
_whitelogger has joined #panfrost
tgall_foo has quit [Ping timeout: 245 seconds]
NeuroScr has quit [Quit: NeuroScr]
megi has quit [Ping timeout: 246 seconds]
herbmillerjr has quit [Quit: Konversation terminated!]
herbmillerjr has joined #panfrost
davidlt has joined #panfrost
davidlt_ has joined #panfrost
davidlt has quit [Ping timeout: 245 seconds]
davidlt__ has joined #panfrost
davidlt_ has quit [Ping timeout: 246 seconds]
vstehle has joined #panfrost
davidlt_ has joined #panfrost
davidlt__ has quit [Ping timeout: 248 seconds]
davidlt_ has quit [Remote host closed the connection]
megi has joined #panfrost
davidlt has joined #panfrost
davidlt has quit [Ping timeout: 246 seconds]
raster has joined #panfrost
pH5 has joined #panfrost
NeuroScr has joined #panfrost
NeuroScr has quit [Quit: NeuroScr]
davidlt has joined #panfrost
raster has quit [Remote host closed the connection]
raster has joined #panfrost
davidlt has quit [Ping timeout: 268 seconds]
raster has quit [Remote host closed the connection]
raster has joined #panfrost
raster has quit [Remote host closed the connection]
raster has joined #panfrost
afaerber has quit [Quit: Leaving]
afaerber has joined #panfrost
tgall_foo has joined #panfrost
megi has quit [Ping timeout: 258 seconds]
davidlt has joined #panfrost
jolan has quit [Quit: leaving]
jolan has joined #panfrost
megi has joined #panfrost
adjtm_ has quit [Ping timeout: 268 seconds]
<alyssa> So I have the identical shader in both deqp and freedreno tests (working in the latter only), identical compiled shader code
<alyssa> So what forth could be the difference
<alyssa> And I now have identical cmdstream too um
davidlt has quit [Ping timeout: 244 seconds]
<alyssa> Got it.. missing a flush
<alyssa> Now we're passing our first few dEQP compute shader tests
<raster> alyssa: the glmark offscreen tests are the same res ass the default windowed one right?
<alyssa> raster: should be
<raster> i thought so :) faster to ask here than double check src
<alyssa> raster: --size 0x0 for explicit setting
<alyssa> 800x600 should be def for both
<raster> just getting ddk number s- booted into an older kernel - gaev up with getting kbase module on current one
<raster> funtime .weston doesnt work on ddk
<raster> e does work
<raster> :)
<raster> but glmark in a window in wayland is vsync limited :(
<alyssa> Conformant apps not working on proprietary drivers is a feature, not a bug :^)
<raster> ok
<raster> 397 is what i get on ddk
<raster> thats the measuring stick in the sand for me right noe
<raster> now
<raster> now its back to panfrost land
<alyssa> Brrgh
<alyssa> Something something LTS released called "Permafrost"
<alyssa> Something new IP something global warming
<daniels> raster: 397 vs. what on panfrost?
<raster> well 222 onscreen
<raster> but that's not fair
pH5 has quit [Quit: bye]
<daniels> yeah, offscreen comparison would be good :)
<alyssa> raster: Also, if you're doing off-screen, also try PAN_MESA_DEBUG=afbc just for fun
<alyssa> AFBC is disabled by default (hidden behind a flag) since it's unfinished and not yet conformant, but enough is implemented for glmark to work
<alyssa> So take the numbers with a grain of salt but also realize the blob is using AFBC everywhere so
<raster> i can only compare under e as that's the only thing that is working on both
<alyssa> raster: Compare what-to-what?
<raster> so for offscreen (where wayland is just a connection) i get 303fps
<alyssa> on blab or on panfrost?
<raster> 397 is blob with older kernel+kbase (4.4)
<raster> thats offscreen
<raster> onscreen is 57.. its vsync limited so useless as a comparison
<raster> onscreen with panfrost is 222, offscreen is 303
<raster> so far afbc on is looking far worse than off :)
<raster> now i'm feeling confident of what my numbers are and where they are from.
<raster> oh wait
<raster> thsi is oscreen
<raster> let me roll back
<raster> why are my numbers so wildly variant
<alyssa> I dunno
<alyssa> I just did `glmark2-es2-wayland --off-screen` and got 717fps
<raster> different board/gpu/clock so ... :)
<raster> ok 225 onscreen
<raster> for sure
<raster> let me run an offscreen run... then afbc after that
* alyssa should probably finish up the AFBC implementation and push it
<raster> for onscreen weston vs e are very different i think mostly due to e not doing explicit fence/sync
<alyssa> Er, not push it, enable it
<raster> i could easily imagine that as being a huge difference there
<alyssa> Particularly for off-screen, yeah
<raster> is there some way to turn explicit fence/sync off in weston?
<alyssa> The AFBC implementation we have now is correct but incomplete, since CPU access is stubbed out
<raster> oh for offscreen i doubt it should matter...
<raster> oh wait not talking afbc - fences/sync protocol client/server for wl
<raster> afbc is another kettle of fish
<alyssa> (For linear, we can do CPU access arbitrarily ofc. For Utgard-style tiled textures, we have CPU routines for software tile/untile. For AFBC, we have nothing rn)
<alyssa> It wouldn't be hard to wire up something with u_blitter to do AFBC<->linear in hardware
<alyssa> Justhaven't gotten around to it
<raster> ok
<raster> offscreen 511
pH5 has joined #panfrost
<raster> offscreen+afbc 633
<alyssa> Nice
<alyssa> Vs 397 for the blob?
<anarsoul> alyssa: well, enable at least utgard-style tiling, it should improve performance a bit
<raster> yes
<alyssa> anarsoul: It is enabled, yeah
<raster> but this is an ancient blob
<alyssa> raster: All the blobs are..
<raster> its whatever binaries were put togethr ages ago by rockchip in userspace and ayfan's kernel tree
<raster> but its a line in the sand
<alyssa> See, the neat thing about proprietary stacks, is that they're *always* ancient by the time they're on your board and working.
<raster> true
<raster> now... thats it for this week
hlmjr has joined #panfrost
herbmillerjr has quit [Ping timeout: 268 seconds]
jrb has quit [Remote host closed the connection]
sphalerit has quit [Write error: Connection reset by peer]
thefloweringash has quit [Write error: Connection reset by peer]
EmilKarlson has quit [Write error: Connection reset by peer]
TheCycoTWO has quit [Write error: Connection reset by peer]
_whitelogger has joined #panfrost
thefloweringash has joined #panfrost
sphalerit has joined #panfrost
EmilKarlson has joined #panfrost
TheCycoTWO has joined #panfrost
jrb has joined #panfrost
raster has joined #panfrost
<alyssa> umulExtended looks annoying
<alyssa> Although that's more because I don't have 64-bit support wired in yet
<alyssa> And anyway I kinda feel like this series is way too long already
<alyssa> No need to throw 64-bits/atomics into the mix
<alyssa> Wonder if I can cherrypick all but the hacks and get something okay rebased out
raster has quit [Remote host closed the connection]
<alyssa> daniels: Oh, re the sleep theory
<alyssa> I had a dream last night that included a snippet of a disassembled Midgard shader, including two new/unknown ALU ops
<alyssa> Do what you will with that information.
davidlt has joined #panfrost
<alyssa> Anyway, I got the SSBO series more or less cleaned up.
<alyssa> There's one patch I'm leaving out (a two line hack that would regress stuff)
<alyssa> So SSBOs mostly won't be functional on master... but keeping a 1 insert, 1 delete patch out of tree is much preferable to uh
<alyssa> 14 files changed, 417 insertions(+), 78 deletions(-)
<alyssa> out of tree (plus the two line patch)
<alyssa> { This branch, plus the 2 line patch, make compute shaders work :) }
raster has joined #panfrost
NeuroScr has joined #panfrost
stikonas has joined #panfrost
<stikonas> looks like I can switch from sway to KDE Plasma with panfrost now...
<stikonas> very impressive job you are all doing here
<alyssa> stikonas: Thank you! :)
<alyssa> Let me know how Plasma is. I don't think it's as well-tested as other environments but it should work.
<stikonas> well, so far I tested it for like 5 seconds
<alyssa> 5 seconds longer than me :)
* alyssa is not a KDE user
<stikonas> but plasmashell is without visible defects
<alyssa> Krita is my love though :)
<stikonas> last week it was still quite distorted
<alyssa> stikonas: 16edd56fccad7ba0e1b938a60f0f5e927a104585 probably fixed things then! :)
<stikonas> I was thinking maybe I should recompile kwin with gles instead of opengl, but it looks like opengl is now good too
<alyssa> which landed early this week
* stikonas going to git log
<stikonas> alyssa: I think kernel driver crashed
<stikonas> but I'm still on 5.2 kernel, so maybe that's why...
<alyssa> stikonas: Hmm
<alyssa> robher: ^^
raster has quit [Remote host closed the connection]
somy has joined #panfrost
adjtm_ has joined #panfrost
<robher> stikonas: I'm not seeing anywhere we could have a null. I need to know what's at 'drm_gem_shmem_create+0x48' and/or offset 0x11c of drm_gem_object. It has to be done with your gentoo kernel (and config).
davidlt has quit [Ping timeout: 258 seconds]
<stikonas> robher: how can I get those offsets?
<stikonas> do I need to compile kernel with drm symbols?
<stikonas> s/drm/debug/
<robher> stikonas: start with the code offset as the 0x11c offset may be something else besides drm_gem_object.
<robher> stikonas: objdump on the vmlinux file will give you the disassembly.
<stikonas> hmm, but to get line numebrs I guess I need to recompile...
<robher> stikonas: If you give me the disassembly of drm_gem_shmem_create that's maybe enough.
<stikonas> ok, let me try that first
<robher> note, that's the aarch64 version of objdump
<stikonas> robher: yes, I have ovjdump locally, but I'm having the problem with EFI header...
<stikonas> objdump does not recognize it
<stikonas> I've already ungziipped my kernel image...
<stikonas> but I guess i need to do something more
afaerber has quit [Quit: Leaving]
<robher> The EFI header would be in Image file. You need the elf file vmlinux.
<stikonas> oh, I would need to compile kernel again
<stikonas> I guess I'll just enable debug symbols then...
<stikonas> I only have image file at the moment
<stikonas> my build tree was already cleaned
<robher> stikonas: then you need to retrigger the crash. debug symbols may change the code generation.
<stikonas> yeah, I understand
<stikonas> well, I hope it's easy to retrigger
<stikonas> I wasn't doing anything
<stikonas> logged in, went away
<stikonas> came back and it already locked up
<alyssa> Ok, I've done my due dilligence on XFB/SSBO
<alyssa> I guess I can come back to gfx now
<alyssa> Er lemme deal with outstanding email first
<stikonas> I'll try to report back my findings in an hour... If you are away, I guess tomorrow...
<alyssa> stikonas: Tomorrow is saturday ;)
<stikonas> oh, robher is away... Well, for me it's vice versa... I can only check this during my free time...
<robher> alyssa: BTW, what's protecting the BO cache list?
<alyssa> robher: Protecting it? Like, threadly?
<robher> alyssa: yes.
<alyssa> The BO cache is the least of your worries for multithreaded Panfrost things.
<alyssa> next question.
* robher goes back to the kernel
<robher> lalalala
raster has joined #panfrost
<alyssa> Okay caught up on things... MRT I guess..
<alyssa> The only thing I can think of is maybe a hw limitation that different RTs have to live in the same BO
<alyssa> But that doens't make sense.
<alyssa> Or there's some hidden state I don't see
mearon has quit [Ping timeout: 258 seconds]
<alyssa> Very bizarre
<alyssa> So weird.
mearon has joined #panfrost
raster has quit [Remote host closed the connection]
<alyssa> Oh my I wonder
<alyssa> All zero. Ok.
sravn has quit [Ping timeout: 272 seconds]
sravn has joined #panfrost
pH5 has quit [Quit: -_-]
<stikonas> still can't retrigger crash for now...
<stikonas> will try some other day
stikonas has quit [Remote host closed the connection]
stikonas has joined #panfrost
<stikonas> ok, triggered...
<stikonas> but I'll investigate tomorrow
<stikonas> I think it happens when it goes to sleep
stikonas has quit [Remote host closed the connection]