Hacker News Reader

Comment by radarsat1 | Hacker News Reader

Depending on how many individual tweaks are necessary for hardware variants of course... but at this level of code & complexity it actually seems pretty reasonable to write 3 or 4 versions of things for different vendors. More work yes, but not pointless.

View on HN

A nice example of this is fftw which has hundreds (if not thousands) of generated methods to do the fft math. The whole project is a code generator.

It can then after compilation benchmark these, generate a wisdom file for the hardware and pick the right implementation.

Compared with that "a few" implementations of the core math kernel seem like an easy thing to do.

Not exactly comparable, as you said, the FFTW implementations are auto-generated but it doesn't sound like these few implementations will be.

ATLAS was an automatically tuned BLAS, but it’s been mostly supplanted by ones using the hand-tuned kernel strategy.

Apache TVM does something similar for auto-optimization and last time I checked it wasn't always a win against OpenVINO (depending on the network and batch-size) and it came with lots of limitations (which may have been lifted since) - stuff like dynamic batch size.

I wish we had superoptom

Metalibm[1,2] is a different idea, but kind of related: if you need a special (trigonometric, exponential, ...) function only with limited accuracy or only on a specific domain, you can have an approximation generated for your specific needs.

[1] https://github.com/metalibm/metalibm

[2] https://indico.cern.ch/event/166141/sessions/125685/attachme...

Reply on news.ycombinator.com