primitives 0.1.0

Primitves Asm
docs.rs failed to build primitives-0.1.0
Please check the build logs for more information.
See Builds for ideas on how to fix a failed build, or Metadata for how to configure docs.rs builds.
If you believe this is docs.rs' fault, open an issue.

Experimentation to Optimize Alloy Primitives

(honestly idk what I'm doing, but you never know what you are doing until you try anyway, so who cares) don't tell anyone but this is just an excuse for me to learn simd stuff(idk what it's)

What's This All About?

This project is a wild ride into the world of primitive optimization. We're throwing ideas at the wall and seeing what sticks.

Design Choice

The Grand Plan (or lack thereof)

  1. Early support for arm neon
  2. Take a look at Alloy primitives
  3. Scratch our heads and wonder how to make them faster
  4. Remember that SIMD exists and seems cool
  5. Try to apply SIMD to... something. Anything, really.
  6. See what happens and hope for the best
  7. Implement Vector operations

How to Join This Chaos

  1. Clone this repo
  2. Install... things. (We'll figure out what exactly later)
  3. Run some code and see if it explodes
  4. If it doesn't explode, check if it's faster
  5. If it's faster, celebrate! If not, pretend you meant to do that

Results So Far

Benchmark Comparison

Primitive SIMD(Primitives) Alloy Primitives Performance Change
address/checksum 169.43 ns 201.41 ns Faster by ~23%
bytes/32 13.818 ns 15.818 ns Faster by ~14%
bytes/64 14.614 ns 17.667 ns Faster by ~21%
bytes/128 36.106 ns 36.859 ns Slightly faster by ~2%
bytes/256 42.191 ns 41.024 ns Slower by ~2.8%

SIMD Results for Parity Inversion

We've recently added SIMD optimizations for Parity inversion. Here are the benchmark results:

Input Size Alloy Primitives SIMD Version Performance Change
10 21.235 ns 23.258 ns Slower by ~9.53%
100 101.93 ns 85.767 ns Faster by ~15%
1000 974.07 ns 722.84 ns Faster by ~25%
10000 10.090 μs 8.007 μs Faster by ~20%

Key observations:

  • SIMD shows overhead for very small inputs(stick with the non simd version)
  • Performance gains become significant for larger inputs (1000+)
  • Consistent ~20-25% improvement for large inputs

Contribute

Got ideas? Throw them in! Know what you're doing? Even better, we could use the help!

License

MIT