<icecream95>
tomeu: That sounds better than on chromebooks, where you have to manually change the partition flags on the recovery image, replace the install script with one that just runs flashrom or crossystem, then plug the drive in and hope it works...
<narmstrong>
tomeu: nice !
<tomeu>
icecream95: I think with newer models one can do some stuff from the usb port, with CCD
<narmstrong>
tomeu: did you boot linux try panfrost yet ?
<tomeu>
I'm copying the fs to emmc right now
<tomeu>
so I'm almost there
<icecream95>
tomeu: Don't you need a USB-C cable with some of the wires switched round for that?
<tomeu>
icecream95: something like that, yeah
<tomeu>
smaller than a servo :p
andrey-konovalov has joined #panfrost
stikonas has joined #panfrost
icecream95 has quit [Ping timeout: 265 seconds]
raster has joined #panfrost
yann has joined #panfrost
stikonas has quit [Remote host closed the connection]
<urjaman>
alyssa: did someone fix the perf regression you mentioned like a bit over a week ago?
<urjaman>
that is, should i update mesa :P (or should I run a bisect? :P)
<robmur01>
there's not really much to diagnose without the contents of memory to hand
<robmur01>
and the timeouts are probably caused by the page faults anyway ;)
<tomeu>
Xalius: works here :/
<Xalius>
on what distro are you running?
<tomeu>
debian
<Xalius>
I have swapped around most components from arch-arm/manjaro-arm without success
<tomeu>
kernel 5.3
<Xalius>
xorg stable or xorg master also produced the same
<urjaman>
hmm i do still have the not-yet-upstreamed part of bbrezillon's patch set in my kernel - that dealt with some of the page faults i was seeing before
<Xalius>
I haven't tried kernel 5.3
<urjaman>
(some still happen occasionally but it's not that easy to track down)
<urjaman>
i decided that i needed some sort of a comparison on this mesa update so i did x11perf -shmputxy100 before and after
<urjaman>
from about 70/sec to 27/sec
<urjaman>
i'ma bisect
davidlt has joined #panfrost
buzzmarshall has joined #panfrost
robher has quit [Remote host closed the connection]
daniels has quit [Remote host closed the connection]
robher has joined #panfrost
daniels has joined #panfrost
Xalius has quit []
davidlt has quit [Remote host closed the connection]
davidlt has joined #panfrost
Lyude has quit [Read error: Connection reset by peer]
Lyude has joined #panfrost
daniels has quit [Remote host closed the connection]
robher has quit [Remote host closed the connection]
robher has joined #panfrost
daniels has joined #panfrost
<urjaman>
alyssa: "32851ff71521d203c08e26b3198d96d0f6467bb5 is the first bad commit"
<urjaman>
that is "panfrost: Cleanup tiling selection logic"
<raster>
aaaaaah "cleanup" commits breaking things... always happens over the decades. you learn to beware of "cleanup" commits :)
<urjaman>
the commit does claim to have a non-zero effect (fixing dEQP case(s)), so it's not really a pure "cleanup" commit in that sense anyways
<raster>
the classic sneaking real changes into a "shouldn't change anything" cleanup :)
pH5 has quit [Quit: bye]
karolherbst has quit [Ping timeout: 268 seconds]
yann has quit [Ping timeout: 265 seconds]
<alyssa>
urjaman: Ohhh, interesting, okay.
<alyssa>
urjaman: Thank you so much for bisecting that :)
enunes has quit [Ping timeout: 248 seconds]
<alyssa>
Mistake #1 in that commit is that the checksumming logic changed for no good reason
<alyssa>
Though checksumming hasn't made a big difference in perf when I was poking at it
<alyssa>
Really though we just need the tiling routine to be 1000x faster, grr
<anarsoul>
:)
<anarsoul>
that'd be nice
<anarsoul>
I guess some arm assembly should do the trick?
<anarsoul>
maybe not 1000x
<anarsoul>
but 2x should be possible
<alyssa>
It's ... a lot harder than you might hope, unfortunately.
<alyssa>
I might take another stab at it once the GLES3 drowning slows down
<alyssa>
Let's uh investigate to see what the issue is these days
karolherbst has joined #panfrost
<alyssa>
Problem #1 is that it's annoyingly difficult to instrument this code sanely since there's a subtle dependency on Gallium
<alyssa>
Just need pipe_box, probably better to eliminate that entirely..
<anarsoul>
alyssa: maybe it would make sense to utilize GPU for tiling/untiling?
<alyssa>
anarsoul: That's a last restort imho
<anarsoul>
why?
<anarsoul>
it should be able to do that faster
<alyssa>
I mean
<alyssa>
hrm
<anarsoul>
*faster than CPU
<alyssa>
I suppose it's still an option worth looking into
<robmur01>
cf. the "higher throughput" vs. "significantly larger setup overhead" from crypto and pretty much every other hardware offload argument ever :P
<alyssa>
robmur01: A bug in libgcrypt's implementation of AES-GCM via armv8 crypto prevented me from chatting with a friend on XMPP
<alyssa>
Software makes me happier sometime s:V
<alyssa>
anarsoul: We agreed to drop RGB (24bpp) , yeah?
<alyssa>
Although wait 6bpp formats are a thing in GL
raster has quit [Quit: Gettin' stinky!]
<alyssa>
and 12bpp
<anarsoul>
alyssa: yeah
<anarsoul>
I dropped 24bpp in lima
<alyssa>
*nods*
Ke has joined #panfrost
<alyssa>
Really we need better templating etc
<alyssa>
like etna has
<alyssa>
Er no there are, agh
<alyssa>
This code is at the mix of complex and slow alyssa is not happy
<anarsoul>
you mean code in etnaviv?
<alyssa>
Yeah
<alyssa>
Okay, so, job#1 is to cleanup this code since there's a lot of half-baked opts here and it's just making things hard to work with
<alyssa>
Supporting different bpp is annoying..
megi has quit [Ping timeout: 240 seconds]
yann has joined #panfrost
<anarsoul>
alyssa: sure it is
<anarsoul>
but we need that
<anarsoul>
see the positive side - you don't have to deal with separate channels
* robmur01
recalls struggling to comprehend weird planar PC video modes (EGA, maybe?) as a youngster
* anarsoul
started with zx spectrum so is used to weird framebuffer layouts