16:06 <alyssa> tomeu: BTW, if you need a break from Bifrost, something we'll need on Midgard is fp16 varyings

16:07 <alyssa> (And on Bifrost too, of course)

16:08 <alyssa> I need to figure out the exact compiler ABI for that, but conceptually if a varying is read as fp16/mediump from the fragment shader OR written as such in the vertex shader, the effective precision is only fp16

16:08 <alyssa> so the varying itself can be done in fp16, which saves memory / bw

16:08 <alyssa> Already we have pan_format_from_glsl but that's broken if one side does fp32 and the other does fp16

16:08 <alyssa> since varyings are linked between stages

16:09 <alyssa> (It also fails to recognize a lot of varyings could be fp16.)

21:27 <alyssa> Unrelated, but I have derivatives essentially working

21:27 <alyssa> The big outstanding problem is that MSAA isn't wired up so those tests fail. But the rest are ok, with the caveat that the cont/last flags are hacked on

21:27 <alyssa> Doing this the right way takes some more data flow analysis.

21:28 <alyssa> (To figure out if the thing we're differentiating is itself the result of a texture.)

21:28 <alyssa> Actually, texture(S, texture(T, uv).xy) is enough to repro the same bug- basically we need to keep helper invocations alive long enough

21:29 <alyssa> icecream95: ^^

21:31 <alyssa> More topically, my fp16 branch is cleaned up and just cleared CI, so yay.

21:31 <alyssa> Still more opts to do before it can be landed w/o regressing cycle count

22:29 <alyssa> Turns out folding conversions in *correctly* is hard