Use a vec4 instead of an array in canvas shader instance buffer

[clayjohn]

This avoids a pathological performance cliff on Adreno devices

I noticed a regression between dev2 and dev3 in 2D rendering performance. I finally took the time to investigate it last night. I found that on my Pixel 4 (Adreno 640) frame time roughly tripled (from 15 FPS down to 5 FPS) in the isometric 2D demo.

The regression was most likely caused by the move to batching which is very weird as batching barely affects the shader code.

I ran a profiler over the demo and found something very odd. The profiler showed shader processor memory bandwidth use go from 16mb/s to 6.5GB/s. I expected that this scene would be ALU bound due to the high number of lights.

After a lot of investigation, I accidentally stumbled on the cause. Adreno doesn't like having an array inside a struct inside of an array. We can easily fix that since vectors can be indexed like arrays. Simply by changing the lights member into a uvec4 we reclaim all the performance we lost. Non-Adreno devices don't care one way or the other, so this ends up being a very safe fix.

[stuartcarnie]

Nice find!

[Ansraer]

I hate that this works. Approved.

[Repiteo]

Thanks!