hanetzer1 has joined #panfrost
hanetzer has quit [Ping timeout: 480 seconds]
hexdump01 has joined #panfrost
hexdump0815 has quit [Ping timeout: 480 seconds]
alpernebbi has quit [Ping timeout: 480 seconds]
alpernebbi has joined #panfrost
Daanct12 has joined #panfrost
alarumbe has quit []
rasterman has joined #panfrost
warpme has joined #panfrost
tchebb has joined #panfrost
acryo has quit [Ping timeout: 480 seconds]
Moe_Icenowy has joined #panfrost
MoeIcenowy has quit [Read error: Connection reset by peer]
acryo has joined #panfrost
warpme has joined #panfrost
rasterman has quit [Quit: Gettin' stinky!]
averne_ has quit [Ping timeout: 480 seconds]
rasterman has joined #panfrost
warpme has joined #panfrost
averne has joined #panfrost
alarumbe has joined #panfrost
alarumbe has quit [Remote host closed the connection]
alarumbe has joined #panfrost
Daanct12 has quit [Quit: WeeChat 4.6.3]
rasterman has quit [Quit: Gettin' stinky!]
dsimic is now known as Guest18684
dsimic has joined #panfrost
Guest18684 has quit [Ping timeout: 480 seconds]
hanetzer1 has quit [Read error: Connection reset by peer]
hanetzer has joined #panfrost
warpme has joined #panfrost
paulk has quit [Ping timeout: 480 seconds]
rasterman has joined #panfrost
paulk has joined #panfrost
<
alyssa>
Mary: wow that looks so much nicer than the nir_builder we had in 2020 !
<
Mary>
yeah it's quite nice right, not too happy with the amount of repacking going on tho...
<
Mary>
the indirect offset could use some atomics too to remove the barriers around but at the very least this pass CTS nicely
<
alyssa>
probably definitely want to parallelize it too
<
alyssa>
at least the minmax search part
<
Mary>
oh yeah totally but this could probably do for now (will have to document everything around)
<
Mary>
could probably make variants of the minmax search depending on the index size too
<
alyssa>
definitely do that!
<
pac85>
Oh indirect on bifrost that is cool
sally has quit [Ping timeout: 480 seconds]
<
Mary>
pac85: oh so that would remove the need for variants too I should probably try both ways and see thanks!
<
alyssa>
Mary: what's wrong with the variants?
<
alyssa>
also, this is Mali, you can just use a formatted load and do the conversion in hardware
<
Mary>
alyssa: probably nothing... just never tested them so far actually
<
Mary>
right we could also just use that
<
alyssa>
i'm jealous, it'd be really uesful on AGX
<
alyssa>
especially for geom/tess stuff
<
alyssa>
i guess on mali your other option is to setup an attribute descriptor for it
<
alyssa>
but ld_cvt feels a lot less heavyhanded and more bindless
<
alyssa>
(honestly if i designed mali i'd be inclined to delete most of the attribute unit and just use ld_cvt all over)
<
pac85>
I'd imagine letting the hw convert would be faster yeah. Also less iterations
rasterman has quit [Quit: Gettin' stinky!]
sally has joined #panfrost
sally has joined #panfrost
sally has quit [Remote host closed the connection]
sally has joined #panfrost
sally has quit [Remote host closed the connection]
sally has joined #panfrost
sally has quit [Remote host closed the connection]
sally has joined #panfrost
sally_ has joined #panfrost
sally has quit [Ping timeout: 480 seconds]
sally has joined #panfrost