01:10 <HdkR> Looks like Bifrost has either a CLZ or CTZ instruction. Found it while testing findMSB/findLSB

01:11 <HdkR> Need to read the shader to see wtf it is actually doing

01:11 <alyssa> HdkR: Probably same as Midg

01:12 <alyssa> HdkR: Namely, a clz instruction

01:12 <alyssa> And then they do 31-clz() or whatever

01:12 <HdkR> yea I see a sub 31-op

01:17 <HdkR> Yea, just a clz

01:18 <alyssa> This is really the biggest mystery of the lot:

01:18 <alyssa> ADD, (1 - src_alpha), src_alpha

01:18 <HdkR> Feels weird to have a clz instead of a lsb/msb op since GLSL doesn't define a way to clz

01:18 <alyssa> Requires a negate on the destination

01:18 <HdkR> bweh?

01:18 <alyssa> IKR

01:18 <alyssa> So either the clip bit doesn't do what I think

01:19 <alyssa> Or the negate doesn't

01:19 <alyssa> Maybe there's some special case, like...

01:19 <alyssa> If you complement the dominant via clip, it negates the non-dominant as well. For some reason.

01:20 <alyssa> Smells like errata more than intended functionality but I dunno

01:22 <alyssa> Regardless, though, with that bizarre assumption for one half the question, I'm passing add tests

01:24 <alyssa> It's so tempting to force blend shaders but that'll bite me, I know

01:27 <alyssa> Integers all ove

02:29 <alyssa> Blend shader stuff on the ML

02:30 <alyssa> Would love a CI run, hmm

02:30 <HdkR> ooo

02:31 <HdkR> Gonna have to take a look at that after dinner :D

02:31 <alyssa> :D

05:01 <tomeu> alyssa: ok, I'm going to schedule another exact run and see if the failures change

05:04 <alyssa> tomeu: Er, I just pushed, awkward timing

05:05 <HdkR> alyssa: I guess a limitation of the blend shader work is lack of support for blend_equation_advanced as well?

05:05 <HdkR> Theoretically easy to extend to support that in the future once anyone cares

05:05 <HdkR> Also, oh snap, called out

05:05 <alyssa> HdkR: There's already a lowering pass for that so I didn't bother yet

05:06 <HdkR> ah

05:06 <alyssa> Biggr issue I forgot about is color_mask, that's for v2 I guess

05:08 <alyssa> tomeu: Oh, wait, you're pushing to armhf and I to aarch64, different targets? :)

05:20 <tomeu> alyssa: oh, what did you push?

05:20 <alyssa> tomeu: Bunch of random fixes I had queued up... I want to know what the net score turns out to be

05:21 <tomeu> ah, ok

05:27 <HdkR> How is store_deref supposed to work? Looking in to it a little bit, does deref_var allow you to calculate a memory offset?

05:28 <alyssa> HdkR: Call lower_io

05:28 <alyssa> Your backend should never see a deref

05:28 <alyssa> If you're writing a nir pass that's different :p

05:29 <alyssa> Unless you mean from reviewing my pass

05:29 <HdkR> alyssa: I'm seeing it with SSBO accesses. Curious about it :P

05:29 <alyssa> Oh..

05:29 <alyssa> Dunno

05:29 <alyssa> Haven't got there

05:30 <HdkR> :D

06:23 <HdkR> Lyude: https://gitlab.freedesktop.org/sonicadvance1/mesa/tree/rhoudek/bifrost_compiler Here you go

06:24 <HdkR> I won't get until this weekend to work on it more

06:26 <HdkR> That roughly makes up for the week of work I missed as well :P

11:19 <urjaman> whoops, i didnt reboot the C201 at all during the week after i installed the rc7 kernel ... and now it's already too old

11:19 <urjaman> (that is, i "installed" it and was like meh it'll get tested whenever i get to rebooting ... nope, didnt get tested before now :P)

13:57 <alyssa> So today wasn't going to be a Panfrost work day, buuuuuuut

13:57 <alyssa> I woke up just now with a cosmic vision on how to implement that feature I tried to do the other day efficiently :p

13:57 <alyssa> Namely, csel with vectorized conditions... Midg can only do scalar conditions... but note the identity holds:

13:58 <alyssa> C ? A : B = (C ? A : 0) | (C ? 0 : B) = (C & A) | (~C & B)

13:58 <alyssa> iand r0, C, A

13:59 <alyssa> iandnot r0, B, C

13:59 <alyssa> er

13:59 <alyssa> iand r0, C, A

13:59 <alyssa> iandnot r1, B, C

13:59 <alyssa> ior out, r0, r1

13:59 <alyssa> 3 ops to do a vec4 compare

13:59 <alyssa> (Compare to 8 we presently emit, or 4 in the best case but that's probably not possible to hit)

14:57 <sphalerite> alyssa: I've been supporting you via liberapay, but apparently your account isn't configured to receive new donations or something? Are you in principle still accepting donations?

15:01 <urjaman> didnt liberapay just end the stuff ... or actually, i didnt check what happened after their problem way back

15:02 <urjaman> I remember sending her about a year worth of $1/month at some point when liberapay said they have payment processor problem of some sort

15:08 <alyssa> sphalerite: Something happened with liberapay last July, I haven't gotten around to figuring out how to reenable it since. I probably should have officially shut the account off but there hasn't been any real activity so it hasn't been hi-prio in either way

15:08 <sphalerite> ah ok

15:08 <alyssa> And donations were supposed to be tied to my blog more than anything \shrug/

15:09 <sphalerite> urjaman: yeah they have a new model which pays the money out directly in advance so they don't have to hold funds

15:09 <alyssa> sphalerite: ..So if I'm not configured, what's happening...?

15:09 <sphalerite> " Cannot be renewed because the account of the recipient isn't ready to receive new payments."

15:10 <alyssa> Ah.

15:13 <alyssa> FWIW the integer arithmetic alone is 3 ops and I don't think it's possible to do lower than that, but they're all scheduler-friendly ops

15:18 <urjaman> sphalerite: yeah i read that now (and to me them not holding funds kinda breaks the service...)