On Apr 19, 2013 3:49 AM, "Sean Owen" <[email protected]> wrote:
>
> Hey Mikio I posted the stack trace on the jblas-users list FYI. It
> happens on a random 100x100 matrix, for example -- but not every time.
I suspect that is due to jni calls being incredibky costly compared to fpu
ops. I see very similar behavior i my calls from R to java and back. And r
calls lapack too for certain things.

>
> PS Sebastian reminds me (though he clearly already said it) that, for
> this use case, in Ax = B, A is always symmetric, and taking advantage
> of that does in fact make it faster, and jblas packages up that
> routine quite easily. That makes the difference.
>
> I find that out of the box, pure Java is faster until about a 100x100
> matrix, which is still a scale that occurs in practice. With some
> digging I think it might be made faster, by maybe optimizing away some
> of the data copying.
>
> (And I don't get any SIGSEGV in this code path.)
>
> The only thing I "miss" right now versus a QR decomposition is that
> I'm using a rank-revealing implementation and it's useful to know the
> apparent rank when the matrix is near singular.
>
> This is quite a nice result overall.
>
> On Fri, Apr 19, 2013 at 10:44 AM, Mikio Braun <[email protected]>
wrote:
> > Hi everyone,
> >
> > I'll have a look at the segfault. Could you provide a small example. I
> > guess it has something to do with the dimensions of the matrix and
> > possible a bug in my code.
> >
> > Just to clarify, Solve.solve() uses dgesv which uses LU decomposition
> > to solve.
> >
> > Solve.solveLeastSquares() and Solve.pinv() use dgelsd which uses an
> > SVD decomposition, as you've pointed out.
> >
> > There are also solveSymmteric() for symmetric matrices and
> > solvePositive() for symmetric and positive definite matrices which
> > should again run a bit faster in that case.
> >
> > Concerning sparse matrices, that's something which is missing right
> > now. However, AFAIK there are no similarly performant sparse matrix
> > libraries in Fortran which you could plug in, meaning that you're
> > probably stuck with the performance of other Java libs.
> >
> > By the way, I'll be in the Bay Area next week, so if you want we could
> > meet and chat about jblas. I'm pretty open to adding new features
> > which users find necessary.
> >
> > Best,
> >
> > -M
> >

Reply via email to