Thank you! Part 2 is going to largely be a rehash of some of the usual algorithmic tricks, although they do have to be adapted somewhat to using a pipeline.

There are several issues that I can see with trying this level optimization on a GPU - firstly, given the limited depth available, renders with doubles are going to be pretty fast already (as long as the GPU supports doubles, which some of the older ones or the ones in phones don't) so there is pretty limited bang for the buck there.

Secondly it is really easy to get good information about CPU microarchitecture and timings (on x86 anyway, thank you Agner Fog), whereas I'm not aware of that level of detail for graphics cards, so they are a bit of a black box. Compounding this, GLSL does compile down to an intermediate code (SPIR-V), but then each graphics card will have that converted to its internal structure in it's own opaque way. It's much harder to optimize when all you can do is try things and look at timings. If I get enough bandwidth, I may look into this.