#
✕ nick:alyssa
#asahi-gpu · 2023-04-10 21:55
alyssathank you^^
#asahi-gpu · 2023-04-10 21:52
alyssa(hopefully neither of the regressions are hard to solve, ITMT i've reverted both commits on next which sucks)
#asahi-gpu · 2023-04-10 21:52
alyssa2. the regression of magenta rectangles from SCANOUT modifier changes (commented on gitlab how to reproduce)
#asahi-gpu · 2023-04-10 21:52
alyssa1. the regression in supertuxkart from compressed z/s (trace sent)
#asahi-gpu · 2023-04-10 21:52
alyssalina: for wednesday's stream, do you mind looking into
#asahi-gpu · 2023-04-09 21:03
alyssaand no rush, the corrsponding Mesa support won't be flipped on for a while
#asahi-gpu · 2023-04-09 21:03
alyssathank you!
#asahi-gpu · 2023-04-09 21:01
alyssa(for an instruction using a sample register source, r1l in that example)
#asahi-gpu · 2023-04-09 21:01
alyssa14:40 <@alyssa> 2e: 6185014002400300 TODO.iter r1, perspective, cf1, cf0, 2, no_forward, sample, centroid, 0b0
#asahi-gpu · 2023-04-09 21:00
alyssahttps://github.com/dougallj/applegpu/pull/35 for the PR
#asahi-gpu · 2023-04-09 21:00
alyssaTellowKrinkle: mind figuring an extend field? o:)
#asahi-gpu · 2023-04-09 18:05
alyssabluetail4: Probably not, I just ate and am quite full :D
#asahi-gpu · 2023-04-09 18:05
alyssaHell yeah!
#asahi-gpu · 2023-04-09 18:05
alyssaPass: 44167, Warn: 38, Skip: 184, Duration: 5:20, Remaining: 0
#asahi-gpu · 2023-04-09 16:52
alyssathink I'm down to a single fail then
#asahi-gpu · 2023-04-09 16:52
alyssaok, another bug down
#asahi-gpu · 2023-04-09 16:02
alyssaHmm, yeah that sounds like progress for dEQP-GLES3 :~)
#asahi-gpu · 2023-04-09 16:02
alyssaPass: 44164, Fail: 3, Warn: 38, Skip: 184, Duration: 5:42, Remaining: 0
#asahi-gpu · 2023-04-08 18:10
alyssa:q
#asahi-gpu · 2023-04-08 18:10
alyssaso what's wrong with punch through + MSAA4
#asahi-gpu · 2023-04-08 18:08
alyssabut only for samples4, not samples2
#asahi-gpu · 2023-04-08 18:08
alyssaso... why is punchthrough + msaa causing faults on the linux side..
#asahi-gpu · 2023-04-08 17:30
alyssabut yeah nothing special there
#asahi-gpu · 2023-04-08 17:28
alyssanvm
#asahi-gpu · 2023-04-08 17:28
alyssaoh er wait there
#asahi-gpu · 2023-04-08 17:28
alyssaget_sample_position seems broken on Metal but whatever
#asahi-gpu · 2023-04-08 17:22
alyssahttps://backtick.town/~bloom/0001-wip.patch
#asahi-gpu · 2023-04-08 17:21
alyssaok, so bary have bary + perspective + fragcoord z + gouraud
#asahi-gpu · 2023-04-08 17:14
alyssaspiciness += 2;
#asahi-gpu · 2023-04-08 17:14
alyssaand a new output select field
#asahi-gpu · 2023-04-08 17:14
alyssaas for the coefficient register setup struct, ooh spicy there's a new unknown bit
#asahi-gpu · 2023-04-08 17:13
alyssaso the coefficient register just gives you the vec2 barys you wanted
#asahi-gpu · 2023-04-08 17:13
alyssa<x, y, 1 - x - y>
#asahi-gpu · 2023-04-08 17:13
alyssathen the barycentric coordinates (the vec3 returned to Metal) are
#asahi-gpu · 2023-04-08 17:12
alyssaiterproj at centre with a special coefficient register produces a vector <x, y>
#asahi-gpu · 2023-04-08 17:12
alyssaThat one is in Metal, fortunately
#asahi-gpu · 2023-04-08 17:11
alyssaRelatedly, how are barycentric coordinates loaded?
#asahi-gpu · 2023-04-08 17:10
alyssaso even though the hardware should be capable we need to infer a lot of these details from how Apple's driver implements other features (namely, interpolateAtOffset)
#asahi-gpu · 2023-04-08 17:10
alyssathis is extra annoying because Metal does not support this feature (though Vulkan does with an extension, as does Direct3D)
#asahi-gpu · 2023-04-08 17:09
alyssaPerspective would be hopefully similar but I haven't thought in enough detail. Worst case you can emulate perspective because at this point we can do our own interpolation
#asahi-gpu · 2023-04-08 17:09
alyssaFor linear interpolation, you should be able to (roughly) interpolate at the current pixel and then subtract an appropriate delta vector based on the barycentric coordinate of the current pixel
#asahi-gpu · 2023-04-08 17:08
alyssabut I haven't worked out the math yet
#asahi-gpu · 2023-04-08 17:07
alyssaI suspect the answer is yes
#asahi-gpu · 2023-04-08 17:07
alyssa(Equivalently, can we interpolate with user-specified barycentrics?)
#asahi-gpu · 2023-04-08 17:07
alyssaSo then the big question is, setting up the coefficient registers as shaded rather than flat, can we load the values at the individual vertices of the triangle?
#asahi-gpu · 2023-04-08 17:06
alyssa(The provoking vertex is in the coefficient register setup, mind. So you can't get the non-provoking vertex for something flat.)
#asahi-gpu · 2023-04-08 17:05
alyssaFor the special case of flat shading, you just take the third value, since the value at the top left of a tile is also the value everywhere
#asahi-gpu · 2023-04-08 17:04
alyssa(where the third component is linearly interpolated)
#asahi-gpu · 2023-04-08 17:04
alyssaand from that vector you can implement whatever interpolation you want
#asahi-gpu · 2023-04-08 17:04
alyssaSo the hardware writes <x delta, y delta, value at the top left of the tile>

plus ›