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
<alyssa>
macc24: 🤷
<alyssa>
anholt: I'm curious, why are uniforms more expensive on v3d than other mobile gpus?
<anholt>
qcom, intel let you take a hunk of system memory and have it loaded into some registers randomly accessible from your instructions. broadcom lets you take a hunk of system memory and an instruction can take the value at the pointer and increment the pointer. so, you have to take the user's uniforms and scatter-gather them into the sequence they'll be used in the program. if you've got branching, you can change the pointer to a new value, but there's
<anholt>
a long delay.
<alyssa>
oh, yikes, that seems ... short sighted ...
<anholt>
it's very simple in HW, at an awful CPU cost.
<alyssa>
(arm and apple are both qcmon, intel style)
<anholt>
also, all your immediates that aren't in the LUT you pack into the uniform stream. at least with the pi4's gpu they added a second stream, so you can use immediates in the normal stream to point the second stream at the user's ubos.
<alyssa>
suddenly I feel better about mali's problems ;-P
<macc24>
why am i not surprised that broadcom is most cursed one
<anholt>
there is so much that I love about the bcm architecture, but that and the multiple-instruction setup-and-fire memory access (so what if you need to spill while you're in the midst of a memory access sequence?) are really painful.
<alyssa>
of course now I need to figure out 8->32-bit widens for integer ops ..
<robclark>
qcmon, is that some sort of new pokemon with a 5g modem?
<alyssa>
Yes
<robclark>
:-P
<HdkR>
Gotta catch 'em all. There's like four or five new ones a year ;)