3 ms·

Depending on how many individual tweaks are necessary for hardware variants of course... but at this level of code & complexity it actually seems pretty reasona

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.

treffer 3y ago
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.
- naasking 3y ago
  Not exactly comparable, as you said, the FFTW implementations are auto-generated but it doesn't sound like these few implementations will be.
- bee_rider 3y ago
  ATLAS was an automatically tuned BLAS, but it’s been mostly supplanted by ones using the hand-tuned kernel strategy.
- touisteur 3y ago
  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
- mananaysiempre 3y ago
  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 https://github.com/metalibm/metalibm [2] https://indico.cern.ch/event/166141/sessions/125685/attachments/201415/282783/DeDinechin-2012-CERN.pdf https://indico.cern.ch/event/166141/sessions/125685/attachme...