stikonas has quit [Remote host closed the connection]
icecream95 has joined #panfrost
archetech has joined #panfrost
<kinkinkijkin>
i foresee a day when ARM chromebooks have 8GB of RAM! and radeon gpus! and 4k screens! it will be in 2037, when 8GB is even smaller than 4 is now, the radeon brand has been nonexistent for 10 years, and 4k is too few pickels
<alyssa>
kinkinkijkin: It's spelled pickles.
<kinkinkijkin>
i noticed as i sent it
<alyssa>
I thought it was part of the joke.
<kinkinkijkin>
i meant to write pickles yes ;p
vstehle has quit [Ping timeout: 260 seconds]
kaspter has joined #panfrost
<alyssa>
bbrezillon: I pushed a bifrost gles3 ci branch (on top of your MRT branch) to alyssa/mesa:boris-gles3
<alyssa>
CI isn't done yet but I'm seeing a lot of faults, ruh roh :(
<alyssa>
Hoping just a couple bugs we're blocking on... don't have time tonight to investigate though
<icecream95>
alyssa: BITSET_FOREACH_SET dereferences the pointer even if the size is zero, this causes crashes in tests with loops
<alyssa>
🤦♂️
<alyssa>
thanks
<alyssa>
RiiR
<alyssa>
OA
archetech has quit [Quit: Konversation terminated!]
<icecream95>
alyssa: See "pan/bi: Fix zero constants" for why many of the ES3 tests are faulting
<icecream95>
(from my cl branch)
<icecream95>
Something similar is needed to that in bi_schedule_clause
<alyssa>
---Forgot the MRT branch was on top of the scheduler. Could maybe be rebased out , not sure.
<icecream95>
Most of the failing 3D texture tests pass on G72
<icecream95>
Only 32 of the 251 fails on G52 also fail on G72
<alyssa>
We've broken G72 enough times, it's only fair :p
agrisis has quit [Ping timeout: 272 seconds]
davidlt has joined #panfrost
kaspter has quit [Ping timeout: 240 seconds]
camus has joined #panfrost
camus is now known as kaspter
davidlt has quit [Ping timeout: 240 seconds]
vstehle has joined #panfrost
<icecream95>
A full dEQP run gave me: Passed: 41776/45117 (92.6%)
chewitt has joined #panfrost
davidlt has joined #panfrost
<bbrezillon>
alyssa: hm, the constant slot selection bug needs to be fixed if we don't want to base GLES3 support on top of your scheduler branch
<icecream95>
alyssa: `unsigned lo = clause->constants[idx] & 0xF;`. Should that be `lo = cons & 0xF`?
kaspter has quit [Ping timeout: 256 seconds]
kaspter has joined #panfrost
warpme_ has joined #panfrost
<icecream95>
Hmm, it doesn't seem that simple. The code is definitely buggy, but I would just disable "fuzzy matching" in bi_lookup_constant (i.e. use `cons == (uint32_t) top`) for now
<bbrezillon>
icecream95: that's what I did
<icecream95>
bbrezillon: I was looking at that commit when you replied :)
<bbrezillon>
alyssa: I think there's a problem with blend shaders on Bifrost: we assume in_type == out_type, but that's not necessarily true
* macc24
is back
<alyssa>
bbrezillon: *blink*
Elpaulo has joined #panfrost
<alyssa>
bbrezillon: The good news is we now *know* the old lookup_constant function will neve be used for a clause with more than one constant/tuple so you can commit whatever hack you want to get the test to pass, the scheduler branch we'll fix later
robmur01 has quit [Remote host closed the connection]
robmur01 has joined #panfrost
<macc24>
woah
<macc24>
g72 works on 20.3.2
jolan has quit [Quit: leaving]
jolan has joined #panfrost
alpernebbi has quit [Quit: alpernebbi]
kherbst has joined #panfrost
karolherbst has quit [Ping timeout: 272 seconds]
<cphealy>
Is there a way of determining if the T720 does YUV color space conversion for free in the texture sampler instead of via shader instructions with NV12 (semi-planar)?
<alyssa>
bbrezillon: AFBC 3D is broken (hence why icecream95 reported a bunch of tests pass on g72 despite failing CI)
<alyssa>
We can disable that for the release (and/or backport a fix)
<alyssa>
and then a bunch of fails disappear
<alyssa>
bbrezillon: there was a lower fragout i8 nir pass somewhere? i'm not hallucinating, right?
<alyssa>
found it
<bbrezillon>
alyssa: ack for reverting 3D AFBC
<alyssa>
bbrezillon: thanks
<alyssa>
(not even reverting, just bailing in can_afbc. The difference is that, if we find and fix the issue a few weeks from now, we can backport the fix and retroactively enable)
<alyssa>
[I don't know if that's allowed but it's probably fine.]
archetech has joined #panfrost
archetech has quit [Quit: Konversation terminated!]
davidlt has quit [Ping timeout: 240 seconds]
icecream95 has joined #panfrost
jgmdev has joined #panfrost
<alyssa>
macc24: give !8427 a spin :)
<alyssa>
advertises ES3.0 (and GL3.0) on Bifrost
<alyssa>
so Midgard/Bifrost should more or less be at feature parity now, which is fun
<macc24>
gl3 and es3 on bifrost?
<alyssa>
yeah
<macc24>
yeah finalyl
<macc24>
when i'm done fixing void linux musl in cadmium
<macc24>
i heard that musl can be faster on slow machines
<alyssa>
🤨
<alyssa>
musl definitely has some of the best code quality of the libc's but raw performance?
<macc24>
idk
<macc24>
it's broken anyway so it needs fixing
<macc24>
wait i still have debian installed
<macc24>
and gpu devfreq somehow works now
<macc24>
alyssa: is there some midgard-only stuff in panfrost's opengl stuff?
<icecream95>
alyssa: "for eleventh debugging." Did you mean eleventh hour debugging?
<macc24>
maybe mark it in features.txt as midgard-only or bifrost-only if there is stuff like that?
<alyssa>
uhhh do I bang out a blog post for bifrost/3?
<alyssa>
trifrost
<alyssa>
bbrezillon: re INSTR_INVALID_ENC faults on gles3, those were genuine bugs I've fixed in !8427
<alyssa>
and also there was some memory corruption that ended up scribbling over shader bo's, in those cases things didn't even disassemble without crashing
<alyssa>
so I feel comfortable landing this, and worst case we backtrack but I agree that given the state of our userbase this is still a net win
<macc24>
>given the state of our userbase
<macc24>
?
<alyssa>
macc24: Nobody is 'accidentally' using panfrost on bifrost
<macc24>
alyssa: i was for 5 minutes
<alyssa>
accidentally, as in, unwittingly
<alyssa>
Got a laptop without paying attention to it having a mali, setup linux without knowing mesa packages drivers, magically have just works 3d until you install an upgrade and all of a udden supertuxkart is glitching
<macc24>
i mean
<macc24>
bifrost is working in currect debian sid
<alyssa>
true, but there still are few bifrost boards running mainline linux for end users
<alyssa>
I do forsee that changing within a year, and I think it already has started to :)
<macc24>
i'd like to get valhall laptop running mainline linux before 2022
<macc24>
s/get/see/
kherbst is now known as karolherbst
<alyssa>
macc24: re perf, I see panfrost as 6x slower on this benchmark for g52
<alyssa>
F'naaaa
<macc24>
alyssa: which benchmark
tomboy64 has quit [Ping timeout: 240 seconds]
<alyssa>
macc24: re bifrost parity -- it looks like we advertise everything that we do on t860, except for dual source blend