search

likebike / fasteval

Fast and safe evaluation of algebraic expressions
https://crates.io/crates/fasteval/
MIT License
267 stars 26 forks source link

Support f32 as calculation & output type #6

Open aabizri opened 4 years ago

aabizri commented 4 years ago

First of all, congrats and thanks for the crate !

I'm currently using fasteval in a computer graphics project, and in that world, f64 or double-precision numbers are very rarely used: not only is the precision often unnecessary but it makes GPUs much slower (often by 2x) as GPUs are big SIMD/T machines and can pack 2 times more f32 instructions than f64 per cycle, and I suppose this is the same for SIMD in CPUs.

This could be an easier step towards full SIMD use within the crate, though I haven't implemented it as a PoC.

Currently I'm converting the f64 to f32 before sending it to the GPU but I think this inefficiency may be a nice optimization, and a good step towards the milestone of supporting arbitrary-precision numbers.

likebike commented 4 years ago

Great idea. I'll add f32 to the list in #7 .

adamsky commented 3 years ago

I've had a go at this recently, using num-traits to create a more generic representation of the float to support both f32 or f64. Let me know if this is something you would consider pulling in, or what you think might be a better approach to tackle it. Here's the repo: https://github.com/adamsky/fasteval/tree/f32

likebike commented 3 years ago

Thanks Adam! I'll review the changes when I get some time. They look really nice. When I review, I will mainly be thinking about whether the same thing can be achieved without adding a dependency.