12:44 <enunes> linkmauve: I experimented a bit with the gtk4 shaders, I noticed that pretty much all shaders end up fairly big and suboptimal also due to gsk/resources/glsl/preamble.fs.glsl

12:45 <enunes> linkmauve: for example: compiling the fragment shaders from super mario 64 average 10 instructions, shaders from glmark2 30-100 instructions

12:45 <enunes> linkmauve: shaders from gtk4, 300-700 instructions

12:47 <enunes> still on my system I get 20-30 fps with gtk4 and I dont get the glitch in your screenshot

12:48 <enunes> but also hacking around to simplify the gtk4 shaders doesn't change that much, there may be some bottleneck somewhere else and that would require a closer look, maybe the way gtk4 is rendering causes too many framebuffer reloads which kills performance with the mali

14:06 <linkmauve> enunes, if the functions in the preamble aren’t used they will be optimised away, so I would suppose that’s a red herring.

14:06 <linkmauve> Do you have a env var to dump annotated shader assembly btw?

14:07 <linkmauve> I’ve used that in iris in the past to know what to optimise in my shaders.

14:10 <enunes> linkmauve: in lima there is LIMA_DEBUG=pp,gp,shaderdb for fragment shader, vertex shader and shaderdb-like summary statistics respectively

14:11 <enunes> for a quicker turnaround on shader analysis I dumped shaders with https://docs.mesa3d.org/shading.html#capture and you can test compile the output with https://gitlab.freedesktop.org/mesa/shader-db

14:16 <linkmauve> Neat!

14:16 <linkmauve> That way I don’t need to rebuild GTK on the Switch, scp the package, install it on the PinePhone or the Olimex, and run it with Lima!

14:17 <linkmauve> Speaking of Olimex, I get the jumpy vertices on both A10 and A20, despite the different GPU revisions. :/

14:18 <enunes> jumpy vertices you mean the glitched output from the screenshot?