Prerequisites: AVX2 CPU; clang; cmake; meson; python3-matplotlib;
probably more.

To download and benchmark everything and generate plot.pdf:

    ./do

Should finish in well under an hour. Exact time depends on your CPU.

The cycle counts here rely on RDTSC (and include RDTSC overhead). For
the most stable timings, you should disable Turbo Boost and pin the
clock speed to its maximum nominal frequency. Reliably measuring timings
with Turbo Boost takes more work, and might produce a different
comparison if the options being benchmarked can run at different boost
frequencies.

See also https://bench.cr.yp.to/supercop.html under "Reducing randomness
in benchmarks" for further measurement issues and workarounds.