netlib-java relaunch... MTJ updates soon!

[Sam Halliday]

Hi all,

I'm the author of https://github.com/fommil/netlib-java and maintainer of https://github.com/fommil/matrix-toolkits-java

I'm near the end of a major (API preserving) rewrite of netlib-java that will make it a *lot* easier for end users to use native implementations of BLAS/LAPACK/APACK and to build their own machine optimised versions. Also included are offset parameters. These have been the two biggest RFEs in almost 5 years.

SNAPSHOT releases are available for testing (I have to get my RPi into a fit shape to check the ARM HF build).

One thing I'm really missing are tests and perf tests. I rely on upstream developers to run their tests against my releases to check that the methods they use are giving the results they expect.

I have a ticket to use F2J to port the official netlib tests

  https://github.com/fommil/netlib-java/issues/21

but that is going to be a major piece of work.

I also have a couple of very small performance tests (basically ddot, dgesvd and dsygv). Results for ddot are on the project page (with some interesting performance from the pure Java impl!).

Next on my list is to get MTJ to use this new lib, and to implement GPU acceleration in MTJ for dense matrix/vector multiplication! Hopefully you'll be able to re-run your benchmarks when that happens.

I was also hoping you might have some more standard BLAS/LAPACK benchmarks that I could use ? I'm tracking this under

 https://github.com/fommil/netlib-java/issues/37

so feel free to respond here on directly on the ticket.

Best regards,

Sam

[Peter A]

What sort of benchmarks are you looking for?  JMatBench is more focused on only a few of the most common functions and does a fairly rigorous set of monte carlo tests.  It could be expanded relatively easy to include more functions, but it would take a long to run through the 100's of BLAS/LAPACK functions.

Now for the 20 or so functions it does it, it does a very good job picking up numerical problems (often missed by unit tests) and changes in runtime performance.  I have caught problems missed by the original author and I use it as a final step before releasing my own library.  MTJ is already included so you can see if anything got worse/better.  You could also add a native BLAS/LAPACK test using JNI which will tell you if the performance is comparable.  I've been thinking of doing that for a bit.

I'm interested in seeing how easy it is to do GPU stuff with MTJ after your changes and how much of a performance improvement there is.  Where on the project webpage are the results you talked about?  I looked briefly and didn't see any link.

- Peter

--
You received this message because you are subscribed to the Google Groups "java-matrix-benchmark-discuss" group.
To unsubscribe from this group and stop receiving emails from it, send an email to java-matrix-benchmar...@googlegroups.com.
For more options, visit https://groups.google.com/groups/opt_out.

--
"Now, now my good man, this is no time for making enemies."    — Voltaire (1694-1778), on his deathbed in response to a priest asking that he renounce Satan.