anarsoul has quit [Remote host closed the connection]
anarsoul has joined #panfrost
raster has joined #panfrost
raster has quit [Remote host closed the connection]
raster has joined #panfrost
<ente>
I read that as "wayland is already dying" and got really confused for a minute
WoC has quit [Remote host closed the connection]
stikonas has quit [Remote host closed the connection]
karolherbst has quit [Ping timeout: 265 seconds]
atler is now known as Guest83679
Guest83679 has quit [Killed (weber.freenode.net (Nickname regained by services))]
atler has joined #panfrost
vstehle has quit [Ping timeout: 260 seconds]
* alyssa
playing around with Panfrost on a 4k display
<alyssa>
Noticeably sluggish :|
<alyssa>
Admittedly T860 is no speed demon
<icecream95>
alyssa: What DE? Xorg or Wayland?
<alyssa>
icecream95: GNOME, Wayland
<alyssa>
GNOME Xorg hit too many display drive issues
<macc24>
has anyone tried ubports on panfrost device?
<macc24>
it hangs my kevin's display
rcf1 has joined #panfrost
<alyssa>
like, the phone?
rcf has quit [Ping timeout: 246 seconds]
<macc24>
the phone what?
<alyssa>
UBports?
<macc24>
yep
<macc24>
i want it because c o n v e r g e n c e
rcf1 is now known as rcf
<ente>
I've been wanting to write a wayland compositor, but I haven't been wanting it bad enough to do it in C
<ente>
(there's no good reason for it to be in C other than readily available libraries, but that's a fairly big selling point. then again, in theory, it's something that's relatively self-contained. it's just.... someone already started doing ocaml bindings for wlroots and they're not finished)
<jekstrand>
I tried to write one in Java many moons ago
<alyssa>
jekstrand: Fancy seeing you here :p
raster has quit [Quit: Gettin' stinky!]
<jekstrand>
alyssa: I thought I might join the fun just to see what's happening. :)
<jekstrand>
y'all stopped chatting on #dri-devel. I have to spy on the project somehow. :)
<alyssa>
jekstrand: Let's see... I infodump and/or complain about Mali, macc24 and kinkinkijkin ramble about Duet, icecream95 yells at me for not doing the thing I said I was going to 6 months ago, and bbrezillon is off hiding in a corner somewhere while managing to get actual work done.
<alyssa>
Did I miss anyone?
<ente>
is #panfrost-offtopic still a thing?
<icecream95>
Hey, I ramble about duet too!
<kinkinkijkin>
I stopped rambling about duet
<ente>
I haven't heard of duet before
<ente>
but I haven't been on irc much lately
<alyssa>
ente: Enchantee 🕺
<ente>
santé?
<HdkR>
and I'm always busy doing random things :)
<ente>
yeah lol I remember when I was still seriously trying to buy a pinebook pro to help at least test panfrost
<ente>
or one of these unobtanium chromebooks
rcf has quit [Quit: WeeChat 3.0]
<ente>
I didn't find a single chromebook in Shenzhen
rcf has joined #panfrost
<jekstrand>
My only mali GPUs are a couple of 400s 🙃
<macc24>
alyssa: i don't ramble about duet here
<macc24>
i ramble about duet on ##panfrost-offtopic
<ente>
I still have a pine64, meant to test-drive lima too but I'm not as interested as I would be in panfrost
<ente>
ah, two ##
<macc24>
oh and i pretend that i'm doing anything in cadmium while not sleeping and procrastinating literally everything
<ente>
but trying to take the wayland compositor to -offtopic doesn't work if the only one who admitted to writing one isn't there ^^
<jekstrand>
ente: Then take it to #wayland. :-P
* icecream95
gives up trying to get gdm to work again
rcf has quit [Quit: WeeChat 3.1-dev]
rcf has joined #panfrost
* alyssa
looks at options to avoid EXT_shader_framebuffer_fetch needing a shader key
<alyssa>
On Bifrost, it's easy, since the conversion descriptor is sourced from a register, with the only condition of warp-invariance
<alyssa>
So we can just pass a sysval for the framebuffer format and everything is fine.
<alyssa>
Midgard is where things get tricky.
<macc24>
alyssa: how about valhall?
<alyssa>
🤷
<alyssa>
On midgard, ld_tilebuffer doesn't handle format conversion for non-blendable formats
<alyssa>
worse, the format descriptor used by DECODE is an immediate, not a register -- so we can't pass as a sysval
urjaman has quit [Read error: Connection reset by peer]
urjaman has joined #panfrost
<alyssa>
DECODE does support a special "tile buffer format" mode but I've only ever seen the blob use it for blend shaders
<HdkR>
macc24: There's no reason to ask about Valhall at this point in time. Wait a minimum of six months
<macc24>
HdkR: D:
<macc24>
there should be more people working on panfrost
<icecream95>
macc24: Thanks for volunteering!
<macc24>
icecream95: if i had valhall device, i'd put minimum effort into panfrost on it until Someone Else picks up valhall
<HdkR>
limiting factor is hardware availability
<chrisf>
"all zero hackable valhall boards"
* macc24
points at asurada
<macc24>
err, mt8192 chromebooks
<HdkR>
Those haven't launched
<macc24>
yet
<HdkR>
Which is why I'm saying minimum of six months
<HdkR>
Need the hardware in hand to do work on it
<macc24>
asurada launch isn't that far
<HdkR>
Plus time for me to RE it :)
<macc24>
i sure hope that someone gets mainline linux on it
WoC has joined #panfrost
vstehle has joined #panfrost
Ntemis has joined #panfrost
davidlt has joined #panfrost
jschwart has quit [Ping timeout: 260 seconds]
kaspter has joined #panfrost
davidlt_ has joined #panfrost
davidlt has quit [Ping timeout: 265 seconds]
jschwart has joined #panfrost
kaspter has quit [Ping timeout: 264 seconds]
<icecream95>
Firefox 85 crashes when run with PAN_MESA_DEBUG=sync: "pandecode_inject_free: Assertion `mem->gpu_va == gpu_va' failed"
<icecream95>
Maybe I should keep running Firefox with PAN_MESA_DEBUG=sync, even though I don't have to anymore, just to catch all the pandecode bugs
warpme_ has joined #panfrost
alpernebbi has joined #panfrost
Net147 has quit [Ping timeout: 240 seconds]
Net147 has joined #panfrost
Ntemis has quit [Read error: Connection reset by peer]
davidlt_ has quit [Ping timeout: 264 seconds]
<amonakov>
bbrezillon, ping, I'm looking at implementing split_div in 32-bit arithmetic, and I need to understand which divisors it needs to handle
<bbrezillon>
any 32B divisor
<amonakov>
in pan_attributes.c, panfrost_compute_magic_divisors is called only for results of panfrost_padded_vertex_count
<amonakov>
is it really different from how it is used on host side?
<amonakov>
the other call site invokes it for hw_divisor which is a multiple of panfrost_padded_vertex_count
<bbrezillon>
right, in practice is seems to be a multiple of padded vertex_count
<amonakov>
it is important, since e.g. if it may be converted to a float without round-off, it admits a nice implementation
<amonakov>
and results of panfrost_padded_vertex_count do have that property
<bbrezillon>
indeed
stikonas has joined #panfrost
<amonakov>
okay, so "difficult" cases are when per-instance vertex count is tiny, but attrib divisor is huge (above 2**24)
<amonakov>
would the caller of split_div be able to handle that by decomposing the draw to (at most 64) separate draws?
raster has joined #panfrost
<amonakov>
bbrezillon: so what do you think, would that be an appealing compromise?
<bbrezillon>
amonakov: you mean calling split_div() iteratively?
<bbrezillon>
not sure I understand what you're proposing
<amonakov>
bbrezillon: I meant splitting the instanced indirect draw into a sequence of draws for each attrib
<amonakov>
the problem arises when there's a huge attrib divisor, making hw_divisor not fit in a float without round-off
<amonakov>
but if divisor is huge, there can't be many elements to draw, because the product of instances*vertices still fits in an uint
<amonakov>
so the proposal is then handle the draw as a sequence of individual draws for 0'th attribs, 1st attribs, and so on
<amonakov>
"there can't be many elements to draw" <-- there can't be many distinct attribs
<bbrezillon>
duh, no, I don't want to have to split draws like that, the cmdstream patching in NIR is painful enough already, and I feel (assuming that's even possible) we'll end up with something even more complicated than the idiv64 lowering
<bbrezillon>
thanks for looking at it BTW
<amonakov>
bbrezillon: it's good exercise? can I ask you a couple questions (long-ish form, so over email probably)?
warpme_ has quit [Quit: Connection closed for inactivity]
karolherbst has quit [Remote host closed the connection]
karolherbst has joined #panfrost
raster has quit [Quit: Gettin' stinky!]
alpernebbi has quit [Quit: alpernebbi]
<amonakov>
bbrezillon: how about split_div that handles easy cases with 6 or 8 FMAs, plus needs a simple fixup loop with 9 iterations for "difficult" cases? :)