luhenry commented on pull request #30810: URL: https://github.com/apache/spark/pull/30810#issuecomment-747405702
I've run some benchmarks overnight comparing the implementation of the native, f2j, and vectorized implementations. You can find the results at https://gist.github.com/luhenry/2cda93cb40f3edef76cb499c896608a9 Some things I noted which are noteworthy or which I need to investigate further: 1. for `daxpy` and `ddot`, the vectorized version is equivalent or faster than the native implementation 2. for `dscal`, `dspr`, `dsyr`, and `sdot`, the vectorized version is faster than native on small vectors/matrices but slower on large ones; that is something I need to investigate in the Vector API implementation itself 3. for `dgemm` and `dgemv`, the performance is at best equivalent to f2j and native; I'm focusing on that now since `dgemm` is a major component of neural net training and logistical regressions; the problem is with my naive implementation of the algorithms and not necessarily with the Vector API itself. ---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
