ChanServ changed the topic of #panfrost to: Panfrost - FLOSS Mali Midgard & Bifrost - Logs https://oftc.irclog.whitequark.org/panfrost - <macc24> i have been here before it was popular
<icecream95>
tfw your opengl app sorta does what you want it to but only because you are doing it wrong
<alyssa>
icecream95: you know what they said, two wrongs make a right!
<alyssa>
("No they don't." "They do when one wrong is accidentally rendering upside down and the other wrong is OpenGL.")
<icecream95>
Good point. Maybe the problem I had *was* that everything was upside down...
<icecream95>
Yep, that fixes it
pendingchaos_ is now known as pendingchaos
<alyssa>
🤣
<icecream95>
It was actually fixed by setting the handy screen_fb_y_flip bool I made to *false*. But I still need it for another rendering path
<icecream95>
uh oh, I forgot that school changed its timetable this term, I'm late for the next class..
<macc24>
yes, they are slow but it doesn't mean they are slow
<macc24>
*blames gnome and runs away*
rasterman has joined #panfrost
Daanct12 has joined #panfrost
Danct12 has quit [Ping timeout: 480 seconds]
<icecream95>
macc24: You should add the Krane "Internal Mic" UCM section to Cadmium
<icecream95>
You might as well add all of the other sections while you're at it, but I'm not sure if I want "SCO" anywhere near my system...
<macc24>
icecream95: ack
<macc24>
SCO?
<icecream95>
macc24: From grepping the kernel it seems to be related to bluetooth audio, but that wasn't the SCO I was referring to...
alpernebbi has joined #panfrost
rando25902 is now known as rando25892
robmur01_ is now known as robmur01
Daanct12 is now known as Danct12
camus1 has joined #panfrost
camus has quit [Ping timeout: 480 seconds]
karolherbst_ has joined #panfrost
karolherbst has quit [Remote host closed the connection]
camus has joined #panfrost
camus1 has quit [Read error: Connection reset by peer]
alpernebbi has quit [Quit: alpernebbi]
nlhowell has joined #panfrost
<alyssa>
macc24: Heh
nlhowell is now known as Guest5158
nlhowell has joined #panfrost
Guest5158 has quit [Ping timeout: 480 seconds]
karolherbst_ is now known as karolherbst
camus1 has joined #panfrost
camus has quit [Ping timeout: 480 seconds]
nlhowell has quit [Ping timeout: 480 seconds]
nlhowell has joined #panfrost
nlhowell is now known as Guest5167
nlhowell has joined #panfrost
Guest5167 has quit [Ping timeout: 480 seconds]
rando25902 has joined #panfrost
rando25892 has quit [Ping timeout: 480 seconds]
nlhowell is now known as Guest5168
nlhowell has joined #panfrost
Guest5168 has quit [Ping timeout: 480 seconds]
warpme_ has joined #panfrost
nlhowell has quit [Ping timeout: 480 seconds]
camus1 has quit [Remote host closed the connection]
camus has joined #panfrost
camus has quit [Remote host closed the connection]
camus has joined #panfrost
gcl has joined #panfrost
megi1 has joined #panfrost
megi has quit [Ping timeout: 480 seconds]
<alyssa>
Nothing more "exciting" than debugging a Midgard fault on a Monday morning ...
<macc24>
on topic of midgard, i'm considering dropping 32bit support in cadmium
<macc24>
buuuuuuut people still using those machines would be unhappy ,_,
<alyssa>
alas
<macc24>
dropping 32bit platform would make kernel development easier... and would make one-for-all image a real possibility
<alyssa>
oh, this isn't a regression. um, still broken though.
nlhowell has joined #panfrost
<alyssa>
me, out loud a minute ago
<alyssa>
Aaaaaaaaaaaaaaa blit shaders
<macc24>
...blit?
<alyssa>
Fail (Verification of renderbuffer 0 failed.)
<alyssa>
grrr. well, it's not faulting now.
camus has quit [Read error: Connection reset by peer]
camus has joined #panfrost
nlhowell is now known as Guest5187
nlhowell has joined #panfrost
Guest5187 has quit [Ping timeout: 480 seconds]
camus has quit [Remote host closed the connection]
camus has joined #panfrost
nlhowell has quit [Read error: Connection reset by peer]
<macc24>
i'm not envious of m1's performance or anything
<alyssa>
bbrezillon: If you ever get around to MRT on panvk, you'll need the workaround in 1c4dd26888a ("panfrost: Port v5 blend shader issue to blitter")
<alyssa>
Or I can do it because it's like two lines of code now. Um
<alyssa>
oh I guess panvk isn't doing blend shaders on midgard at all yet right ok
<alyssa>
well, if you get around to it-- here's a preemptive bug report 😉
camus1 has joined #panfrost
camus has quit [Ping timeout: 480 seconds]
camus has joined #panfrost
adjtm has joined #panfrost
rasterman has quit [Quit: Gettin' stinky!]
<icecream95>
I wonder if we should make the copy_resource path in panfrost_ptr_map use a compute shader for copying, as otherwise the WC reads kill performance
<icecream95>
For example, glmark2-es2 -b buffer:update-dispersion=0.9:update-fraction=0.5:update-method=subdata currently spends *way* too much time copying resources
<icecream95>
If the resource is only being used by the vertex shader, we could also add a compute job that only uploads changes which runs inbetween the vertex jobs
<alyssa>
Hum
<alyssa>
I'm hesitant to merge things that bump up driver complexity substantially based on potential perf increases, given that has bit us so many times before..
<icecream95>
Yes, but how many times has that happened when it was my code?
<icecream95>
("What about all those CRC bugs, including the rather serious one that still isn't fixed?")
<robclark>
icecream95: without having looked so much at the panfrost code.. how do you end up with so much WC read?
<robclark>
other than like deqp/piglit which do readpix
<icecream95>
memcpy of one resource to another
<icecream95>
Because otherwise we need to flush the frame to be able to update the resource
<robclark>
oh, ok so resource shadowing
<robclark>
I did play around a bit with doing that on the blitter (for adreno gen's that have a blitter).. it seemed like we needed to be memcpy'ing a few pages or more to justify it..
<alyssa>
might be worth gating the resource shadowing path on bo->size < some constant
<robclark>
looks like I never landed the patch that moved it to the blitter, so I guess it didn't translate into enough of a win in any cases I could find
nlhowell has joined #panfrost
<icecream95>
QuakeSpasm (with noafbc) spends about 30% of CPU time doing resource copying
<robclark>
oh, I guess you are talking about more than just PIPE_BUFFER then? Ie. shadowing uploaded textures?
<alyssa>
the panfrost logic doesn't discriminate...
<robclark>
non-buffer we always try to do on blitter (on gens that have one) or else u_blitter
<alyssa>
equal opportunity memcpy
<robclark>
sure.. textures tend to be larger.. I wasn't thinking of that case since I assumed you already did it on the hw ;-)
<icecream95>
(Except for AFBC. All formats are equal, but some are more equal than others)
<robclark>
but if you already to batch reordering, just toss it to u_blitter for the !buffer case
<alyssa>
Two lines good, four lines better!
<robclark>
gl drivers tend to accumulate lines of code to deal with perf edge cases over time
nlhowell has quit [Read error: Connection reset by peer]