Design notes
Working notes on why the library is built the way it is. Each records what was measured, including the approaches that measurement ruled out.
The hierarchical filter
The cheap low-band pass, and how the tuning tables choose its coarse parameters from your reference and calibration.
Full correlation output
The full-lag path and the work saved by asking for peaks or screening pairs.
Batched matched filter design
The four-step transform, the split layout, and the fused peak scan.
The SIMD layer
How one source produces AVX-512, AVX2, SSE4 and NEON kernels.
A narrower coarse pass
Measurements on how far the first stage can be narrowed.
Zen 5 instruction notes
Measured issue rates, not vendor documentation.
The GPU backend
How the Vulkan path is built and dispatched, and the measurements that contradicted the plan.
What is left, and what is closed
Three live avenues, what shipped, and the many that measurement ruled out.