On 25 Jan 2016 10:06 am, "Maximilian Trescher" <maximilian.trescher@fu-berlin.de> wrote:

Hi,

> The repetitive modes calculation can often be avoided by
> precalculating the modes and storing the result if the energy or the
> system parameters don't change. This calculation spends most time
> either in solving an eigenproblem or a more complex generalized Schur
> decomposition. As far as I can tell, the former is available in MAGMA,
> the latter hasn't been ported to GPUs yet.

Isn"t the QR algorithm basically a Schur decomposition?
Then it would be in Magma as well:
http://icl.cs.utk.edu/projectsfiles/magma/doxygen/group__magma__zgeqp3__comp.html


> A person with sufficient time and technical knowledge could try and
> replace the linear algebra currently used in Kwant with the
> GPU-accelerated ones, and this likely would result in a speedup.

last year I started writing a Python binding for Magma (using cython),
which I never completed beyond the 3 routines I needed at that time, but
it could be a good starting point to include Magma into python.
If anybody is interested I could share that code.


Best,
Max