<anarsoul>
alyssa: apparently it's hexchat issue, I see similar behavior under X11
rcf has joined #panfrost
chewitt has quit [Quit: Adios!]
stikonas has quit [Remote host closed the connection]
vstehle has quit [Ping timeout: 255 seconds]
icecream95 has joined #panfrost
buzzmarshall has quit [Quit: Leaving]
<icecream95>
While scrolling in hexchat, `sudo perf top --sort dso -p $(pidof Xwayland)` shows [kernel] at over 50%
davidlt has joined #panfrost
<anarsoul>
I wonder if lima has the same issue
vstehle has joined #panfrost
rcf has quit [Quit: WeeChat 2.7]
rcf has joined #panfrost
<icecream95>
Possibly related: When text is drawn in STK there can be ~3000 jobs per frame with only two triangles
<icecream95>
Uniforms are used for the character position, which I don't think was a very good idea...
<anarsoul>
icecream95: on utgard it's just a matter of changing uniform pointer (rest of job can be reused except render state), it's not the same on midgard?
pH5 has quit [Quit: bye]
pH5 has joined #panfrost
<tomeu>
thecycoone: actually, apparently we don't really fiddle with the voltage in linux because of a bug, so we change sometimes to clock frequencies at which the GPU don't get enough power
<tomeu>
the OPP in mainline may be right, we don't know
<tomeu>
but that patch limits the frequencies we run the GPU at, so the voltage as left by depthcharge may be enough to run things stably
yann has quit [Ping timeout: 255 seconds]
<icecream95>
anarsoul: Still, that's 3000 extra jobs that need to be processed every frame, when there only need to be a couple.
<icecream95>
It gets worse when you find out how they do bold text...
<alyssa>
anarsoul: Midgard ought to be similar but we have some drawtime overhead atm, bbrezillon is working on fixing that :)
<bbrezillon>
alyssa: hm, we'd still have 1+ jobs per draw (assuming we're talking about vertex/tiler/write_value/fragment jobs, not batches)
<bbrezillon>
or am I missing something.
<bbrezillon>
?
<alyssa>
bbrezillon: What's that in reply to?
<bbrezillon>
alyssa: your reply to anarsoul saying that I'm working on optimizing the draw path, except it won't reduce the number of emitted jobs
<alyssa>
bbrezillon: Right, but it should lower CPU time overhead at least per-job
<alyssa>
GPU overhead, there's not much we can do there
<tomeu>
anarsoul: don't know how it is in lima, but we submit more than one job in one go
jolan has joined #panfrost
tgall_foo has quit [Ping timeout: 272 seconds]
tgall_foo has joined #panfrost
davidlt has joined #panfrost
<bbrezillon>
alyssa: would it make sense to allocate a separate BO (not from the transient pool) for ctx->payloads[PIPE_SHADER_VERTEX].postfix.attribute_meta?
<bbrezillon>
looks like this BO is updated everytime ->bind_vertex_elements_state()/->set_vertex_buffers() are called
apol has quit [Remote host closed the connection]
<alyssa>
bbrezillon: Depends how often those functions are called :)
Lyude has quit [Read error: Connection reset by peer]