Skip to content

VAR v1.2.0 – 17 Nov 2025

Latest
Latest

Choose a tag to compare

View all tags
@boonzy00 boonzy00 released this 17 Nov 12:51
v1.2.0
0963932

VAR v1.2.0 – 17 Nov 2025

What changed

  • Runtime CPU feature detection in VAR.init(null)
    Checks for AVX2 on x86_64, NEON on aarch64, falls back to scalar otherwise. No compile-time flags needed anymore.
  • Added NEON implementation for aarch64 (routeBatch uses 4×f32 vectors when available)
  • Batch functions now dispatch to the correct implementation at runtime (scalar / AVX2 / NEON)
  • Added optional auto-tuning of the GPU threshold
    When .auto_tune = true, raises the threshold slightly on machines with >16 cores to reduce cache pressure on large servers. Default remains off (fixed 1 %).
  • New small example in README: 1000-drone swarm collision avoidance using cone volumes
  • Fixed benchmark executable name in run_bench.sh and added a --force-path flag for manual testing
  • Updated performance table with real numbers from my Ryzen 7 5700
    On this particular CPU the vector path ended up at ~0.17 B/sec (same as scalar). No measurable speedup here—keeps the numbers honest.

Usage is unchanged

const router = VAR.init(null);  // automatically picks the best available path

All existing safety behaviour (divide-by-zero guard, negative volumes → CPU, etc.) still applies on every code path.

Benchmark table (run_bench.sh, same methodology as before)

Machine Scalar Vector path
Ryzen 7 5700G ~0.17 B/sec ~0.17 B/sec (AVX2)

(NEON numbers will be added once I get clean runs on ARM hardware)

Install / upgrade exactly as before:

zig fetch --save https://github.com/boonzy00/var/archive/v1.2.0.tar.gz

Feedback welcome, especially from anyone running on recent ARM boxes or bigger Zen CPUs.

That’s all for this release.

Assets 2
Loading