* alyssa
wonders what happens to her Kodi if she disabled tiling
<anarsoul>
it's gonna be slow?
<alyssa>
it's already slow \o/
<anarsoul>
slower
<anarsoul>
however I'm not sure where your bottleneck is
<anarsoul>
maybe it'll change nothing
<urjaman>
between the bottle body and the cap, i would assume :P
<alyssa>
okay both "force tiled" and "force untiled" work equally badly >_>
<urjaman>
sorry, i'll just go back to whatever i was supposed to be doing :P
<alyssa>
but force tiled -is- noticably slower
<alyssa>
urjaman: lol
<anarsoul>
alyssa: do you have anything like PLBU in midgard/bifrost?
<alyssa>
PLBU?
<anarsoul>
polygon list building unit
<alyssa>
Maybe?
* alyssa
reads lima code
<anarsoul>
alyssa: see lima_draw.c, lima_update_pp_stream()
<anarsoul>
read comment at the very beginning
<anarsoul>
alyssa: honestly, I don't understand how PLBU works, so you may want to ask Qiang
<alyssa>
no, I don't think we do
<alyssa>
although maybe we do and I've just been misusing it the whole time -- I'd believe it :v
<hanetzer>
remember, if it aint broke, then you don't know how to use it ;P
<anarsoul>
alyssa: that would explain terrible performance
<anarsoul>
well, I'm not sure if it's terrible to be honest
<anarsoul>
haven't tried it yet
<alyssa>
anarsoul: I'm itching to get access to the perf counters
<anarsoul>
:)
<alyssa>
But... I will note on purely GPU-bound workloads, we're... just under the blob in perf
<alyssa>
anarsoul: btw, what do you guys need the untiling routines for?
<anarsoul>
alyssa: in case if cpu wants to read texture back?
<anarsoul>
i.e. transfer_map() untiles it into linear buffer ("staging buffer")
<alyssa>
1) Don't you keep an untiled copy around in CPU memory or anything?
<anarsoul>
transfer_map() tiles it back
<anarsoul>
alyssa: we keep untiled copy until transfer_unmap() is called
<alyssa>
2) ...fair enough, I guess. We have dedi hw for doing it but that's probably significantly slower than doing it in software given the roundtrip
<anarsoul>
alyssa: well, we can sample tiled texture and render it into untiled fb, but I'm not sure how much faster it'll be
<alyssa>
On that note... is there any logical way to get that file in common code? I'm not sure where that would go but I want to keep down minimising
<anarsoul>
alyssa: I think so. Whoever merges last will have to figure that out
<alyssa>
lol
<anarsoul>
yeah, that's gonna be lima
<alyssa>
Depends what happens with this kernel~
_whitelogger has joined #panfrost
chewitt has quit [Quit: Zzz..]
cwabbott_ has joined #panfrost
cwabbott has quit [Ping timeout: 268 seconds]
cwabbott_ is now known as cwabbott
<urjaman>
okay lol^2
<urjaman>
i did a really quick(TM) hack of making pan_nondrm.c talk to the new kernel module ... I was expecting this to crash, but instead es2_info has allocated 1 GB of memory so far :P
<urjaman>
okay, and it eventually printed the info strings and quit lol
<alyssa>
urjaman: Pretty sure we allocate like 2GB of memory regardless
<alyssa>
That's So We Don't Run Out (TM)
<urjaman>
i _have_ 2GB of memory
<urjaman>
it runs a bit faster once i closed firefox so it doesnt need to push it entirely into swap before doing anything, but yeah i kinda missed half of the problem :P (es2gears says "Error submitting" ... aka i forgot that file also uses ioctls :P)
<urjaman>
ok, now es2gears was "running" at something 50-100fps range (without anything on the screen) whilst dmesg got flooded with gpu page faults
<urjaman>
i'll take a look at this (more indepth than just squishing the square peg in the round hole ;p) after I've slept
rhyskidd has quit [Quit: rhyskidd]
rhyskidd has joined #panfrost
indy has quit [Remote host closed the connection]
indy has joined #panfrost
pH5 has quit [Quit: bye]
chewitt has joined #panfrost
pH5 has joined #panfrost
chewitt has quit [Ping timeout: 245 seconds]
rhyskidd has quit [Quit: rhyskidd]
rhyskidd has joined #panfrost
afaerber has quit [Quit: Leaving]
afaerber has joined #panfrost
afaerber has quit [Ping timeout: 252 seconds]
chewitt has joined #panfrost
anarsoul|2 has joined #panfrost
BenG83 has joined #panfrost
pH5 has quit [Quit: bye]
anarsoul|2 has quit [Ping timeout: 268 seconds]
pH5 has joined #panfrost
anarsoul|2 has joined #panfrost
afaerber has joined #panfrost
<Lyude>
urjaman: it did the sam thing for me
<Lyude>
*same thing
<Lyude>
but I suspect it's because if we're freeing resources and we don't have the ioctls for that converted over, they'll just get ignored
<Lyude>
because there isn't really much error handling anywhere in panfrost right now
<Lyude>
although I have at least hooked up the resource allocator in my branch (which I'll post in a second), but not the deallocator. Whereever that happens to be