alyssa changed the topic of #panfrost to: Panfrost - FLOSS Mali Midgard & Bifrost - Logs https://freenode.irclog.whitequark.org/panfrost - <daniels> avoiding X is a huge feature
raster has quit [Quit: Gettin' stinky!]
yann has quit [Ping timeout: 256 seconds]
yann has joined #panfrost
<hanetzer> one of you guys should test wayfire on panfrost :>
<HdkR> What's wayfire?
<HdkR> Is it as cool as Crossfire? (You'll get caught up in the...Crossfire~!)
vstehle has quit [Ping timeout: 258 seconds]
<hanetzer> its basically a simple wm with compiz built in :P
<hanetzer> for instance I can rotate the irc window this is in with it :P
<HdkR> How wobbly are those windows?
<hanetzer> extremely.
<HdkR> Perfect
<hanetzer> especially on a high refresh monitor (1080p240hz anyone? :P)
<HdkR> Need that new 360hz monitor
<hanetzer> my monitor does have a 360hz mode :>
<hanetzer> and a 480
<HdkR> I guess it sacrifices resolution?
<hanetzer> yeah.
<hanetzer> its 'best' mode is 4k120hz
<HdkR> I totally expect that 1080p 360hz monitor to be required to use DP 1.4 and have 420 encoding
<HdkR> Wacky Asus thing
<hanetzer> mine's a samsung with a zisworks kit installed :)
davidlt has joined #panfrost
vstehle has joined #panfrost
buzzmarshall has quit [Remote host closed the connection]
icecream95 has joined #panfrost
chewitt has quit [Quit: Zzz..]
_whitelogger has joined #panfrost
icecream95 has quit [Ping timeout: 246 seconds]
icecream95 has joined #panfrost
tomboy64 has quit [Ping timeout: 240 seconds]
tomboy65 has joined #panfrost
chewitt has joined #panfrost
chewitt has quit [Quit: Zzz..]
bshah has left #panfrost ["Konversation terminated!"]
chewitt has joined #panfrost
chewitt has quit [Quit: Zzz..]
icecream95 has quit [Ping timeout: 265 seconds]
raster has joined #panfrost
stikonas has joined #panfrost
raster has quit [Quit: Gettin' stinky!]
raster has joined #panfrost
kaspter has quit [Quit: kaspter]
kaspter has joined #panfrost
raster has quit [Quit: Gettin' stinky!]
raster has joined #panfrost
_whitelogger has joined #panfrost
robmur01 has quit [Ping timeout: 256 seconds]
robmur01 has joined #panfrost
buzzmarshall has joined #panfrost
chewitt has joined #panfrost
yann has quit [Ping timeout: 260 seconds]
yann has joined #panfrost
davidlt has quit [Ping timeout: 264 seconds]
<alyssa> bbrezillon: I'm noticing nontrivial time spent in bo_wait, even when dontwait is set (just the kernel<-->userspace calls)
<alyssa> I wonder if there's a clever way to avoid those calls for internal (non-shared) BOs
<alyssa> (I.e. by tracking batches in flight in userspace instead of deferring to kernel.)
<alyssa> Also some time spent in the bookkeeping (hash tables), at some point we might want a freedreno-like approach with numberings and then fixed bitsets
NeuroScr has joined #panfrost
<bbrezillon> alyssa: probably, or maybe batch wait requests
<bbrezillon> (pass more than one BO to check/wait-on)
<alyssa> bbrezillon: I'd rather not mess with the UABI at this point if we can help it.
adjtm_ has quit [Read error: Connection reset by peer]
cwabbott has quit [Ping timeout: 246 seconds]
* alyssa trying to debug poor perf
cwabbott has joined #panfrost
<HdkR> alyssa: Have you ever tried Clang/LLVM's xray?
<HdkR> It's instrumented profiling, so you end up with different information than sampling profiler
<alyssa> HdkR: Hm.
<HdkR> https://twitter.com/deanberris/status/938168146392883200 Gives fun visualizations like this
<HdkR> https://imgur.com/a/2rfiw2f Or here is a real app trace that I just captured :P
<alyssa> for my collective IRC brain - https://people.collabora.com/~alyssa/headless.txt for GPU-less benchmarking
<HdkR> :D
<bbrezillon> alyssa: well, you'll have to query the job state at least, which might indeed allow you to mark more than one BO "idle" at a time, but that's still one syscall you can't get rid of
<alyssa> bbrezillon: sure
<bbrezillon> extending the BO wait in a backward compatible way shouldn't be that hard though
<alyssa> I'm not convinced it's necessary, though
<alyssa> We'll see
<alyssa> (Edit: I hope it's not necessary)
<bbrezillon> sure, if you can do without it, that's all good, all I'm saying is that extending is not a big deal, it that's necessary
<alyssa> fair enough!
<bbrezillon> *extending it
<alyssa> aside - looking at the magic .barrier_buffer flag sometimes seen with texture ops
<alyssa> Staring at it long enough, I think it's a reuse flag
<alyssa> at first I thought if you used the same coord multiple times it'd set it, but now I see code like:
<alyssa> texture.2d.cont hr28', texture1, fsampler1, r29.xyxx,
<alyssa> ....
<alyssa> texture.2d.last.barrier_buffer /* XXX */ hr29', texture0, fsampler0, r29.zwzz,
<alyssa> *Not* the same coordinate. But it is the same register, and nowhere in the middle is r29 written to.
LinguinePenguiny has joined #panfrost
<alyssa> So I think the bit says that the texture unit's cached copy of the register is still accurate and can be reused, rather than paging out to the register file again.
<alyssa> also, uh, can't tell if it's my imagination but setting unknown_flags seems to help perf somehow
<alyssa> no, that's my imagination I think
stikonas has quit [Remote host closed the connection]
<alyssa> cubemaps could use some optimization, though... for cmdstream, we should be able to tile them I think
<alyssa> and for shaders, we shouldn't have to repeat the ld_cubemap_coord op
<alyssa> (Unclear if dealing with the latter requires CSE or just marginally better bookkeeping)
buzzmarshall has quit [Remote host closed the connection]
buzzmarshall has joined #panfrost
adjtm has joined #panfrost
<alyssa> (flags_lo & 0x200) is definitely a bit that needs more investigation as icecream discovered
<alyssa> perf improves 'mysteriously' with it unset
<alyssa> also, the blbo seems to disable helper invocs in some cases
<alyssa> fp16 varyings worth a look Soon