jernej has quit [Read error: Connection reset by peer]
jernej has joined #panfrost
Net147 has quit [Read error: Connection reset by peer]
Net147 has joined #panfrost
dstzd has joined #panfrost
icecream95 has joined #panfrost
<icecream95>
alyssa: I wonder if we need a debug flag to do a memset on panfrost_pool_alloc results, as there is a non-zero number of bugs resulting in the GPU reading uninitialised data and causing graphical sessions to hang…
<icecream95>
Not necessarily with zero, maybe a value that is easier to grep for
<HdkR>
414141414141414141
<HdkR>
That one is nice since you get ram filled with 'A' :D
<alyssa>
kinkinkijkin: geometry shaders are a mistake
Net147 has quit [Read error: Connection reset by peer]
Net147 has joined #panfrost
<alyssa>
icecream95: Eek.
Net147 has quit [Read error: Connection reset by peer]
Net147 has joined #panfrost
Net147 has quit [Read error: Connection reset by peer]
Net147 has joined #panfrost
<bbrezillon>
First full deqp-GLES3 run => "Pass: 40269 Fail: 2257 Skip: 183 ExpectedFail: 0 UnexpectedPass: 0 Crash: 1247 Timeout: 2 Missing: 0 Flake: 154"
<bbrezillon>
not bad
<bbrezillon>
if I omit the fact it took 1h30 because of all the faults :-)
<alyssa>
bbrezillon: Nice!
<alyssa>
91% pass rate, that's an A! ;P
<alyssa>
bbrezillon: By the way, I spent some time working on builder support for a new IR
<kinkinkijkin>
while it is true that geometry shaders are a mistake, they're also usable for entirely gpu-performed animation, and me likey functional offloading
<alyssa>
Which should make NIR->BIR a lot less painful
<alyssa>
The bad news is that unless I am extremely careful it will end up being one giant commit rewriting a huge chunk of the compiler "at once"
<alyssa>
which I was hoping to avoid but there's not a good way around it.
<alyssa>
Options to proceed are:
<alyssa>
* Start a bifrost-ng compiler from scratch and work my way up, since everything is getting changed anyway.
<alyssa>
* Merge all the outstanding bifrost MRs, do the change as quickly as possible (and possibly on a weekend), and move on in our brave new Bifrost world
<alyssa>
* Some combination, I don't know
<alyssa>
I'm sort of leaning towards just pushing one big change but keeping development of new compiler stuff upstream. Since I imagine you'd prefer builders too :)
<kinkinkijkin>
I mean, it's not exactly too late to merge a massive change to the core of how it works that will need its own individual and mixed testing
<kinkinkijkin>
driver isn't exactly what one would call fully "production-stable" yet :p
<alyssa>
Indeed.
<alyssa>
But I don't want upstream mesa to regress basic Bifrost support, that's really silly.
<kinkinkijkin>
true
<kinkinkijkin>
perhaps have it included as a compile-time option? if that's even feasible
<kinkinkijkin>
i wouldn't know since i don't know how the code is structured
<bbrezillon>
alyssa: do as you think is easier for you
<bbrezillon>
rebasing pending MRs shouldn't be a big deal anyway
<bbrezillon>
and yes, if the builder is similar to what we have in NIR, I'm all for it :)
italov has quit [Ping timeout: 256 seconds]
<alyssa>
kinkinkijkin: Yeah, that'd be nice but I don't know how hard it would be, see: symbol duplication.
<alyssa>
bbrezillon: Fair enough. And yes, bir' is taking direct inspiration from NIR :)
<kinkinkijkin>
i am really getting impatient for my duet to be here, im still waiting for the order to be placed again after a technical issue