[LukeP0WERS]

Shared memory is a common optimization used to store moderate amounts of data for shaders that is faster to write and load than spilling to local memory. It is significantly faster (3-4 cycles) than loading from the L1 and L2 caches (10-20 and 100-200 cycles respectively), not quite the speed of static registers (1 cycle) but its close. In fidgets case, switching to these improved performance massively, however only 16KiB is guaranteed to be supported by all devices (and for devices that support more it starts to introduce tradeoffs to performance due to lowering occupancy). This means only some of the shader variants can support shared memory registers:

REG_COUNT	Float (4 B)	Interval (8 B)	Gradient (16 B)
8	✓	✓	✓
16	✓	✓	✓
32	✓	✓	✗
64	✓	✗	✗
128	✗	✗	✗
192	✗	✗	✗
255	✗	✗	✗

I tested every model and every model that fit within the window benefited from using it. Each of these measurements came from the median of 3 runs:

Model	Mode	REG_COUNT	Shared Passes	Private	Shared	Change
quarter	2D	8	3/3	2.56 ms	2.57 ms	+0.4%
hi	2D	16	3/3	2.60 ms	2.58 ms	−0.8%
perlin	2D	16	3/3	3.72 ms	3.12 ms	−16.1%
prospero	2D	128	0/3	18.40 ms	18.44 ms	+0.2%
gyroid-sphere	3D	8	5/5	5.30 ms	4.19 ms	−20.9%
tanglecube	3D	8	5/5	3.05 ms	2.69 ms	−11.8%
cabin	3D	16	5/5	4.26 ms	3.26 ms	−23.5%
sponge	3D	16	5/5	15.53 ms	9.41 ms	−39.4%
bear	3D	32	4/5	63.33 ms	33.68 ms	−46.8%
terrain	3D	32	4/5	56.28 ms	26.44 ms	−53.0%
colonnade	3D	64	1/5	16.34 ms	15.64 ms	−4.3%

I was also curious whether I could squeeze out any additional perf by automatically increasing the shared memory limit to whatever the users device supported. Doing so actually worsened performance for some passes. Prospero took 127.6% longer with shared float eval, but that might have been a worse case scenario. Some of the 3D passes saw 5-10% further improvement, so that feature will come with tradeoffs that need to be documented, and I can make a follow up PR to add that later on. For now this is an improvement across the board, at least on the RTX 4070 Laptop I tested this with.

[LukeP0WERS]

This addresses the idea I talked about way back in #243, so I can close that separately once this goes through.