Skip to content

x86-64: black_scholes 2.00x, volume 1.71x, rt 1.42x vs ISPC at matched width #5

Description

@byeongjee

On the Zen4 box #[export] dispatches to the 16-lane avx512f target. Against
ISPC avx512skx-i32x16 at that same width, three kernels lose:

kernel ISPC x16 rustlane ratio
black_scholes 0.23 ms 0.46 ms 2.00
volume 1214 ms 2075 ms 1.71
rt 25.9 ms 36.8 ms 1.42

For contrast, mandelbrot is 0.67 and ao is 0.68 on the same run, and the
geomean over all seven is 1.16.

rt is the sharpest signal: it is 0.55 against neon-i32x8 on aarch64 but
1.42 here, from the same source.

Cause not investigated. Raw per-round timings come from
rustlane-bench/measure.sh + parse_measurements.py x86_64.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions