Hi,
I have no interest in pursuing this at the moment.
That's no problem. We have a permanent record here
(http://thread.gmane.org/gmane.science.nmr.relax.scm/22044) for this
issue. It should be a useful reference for the future.
If numpy should provide a universal function, which is faster,
like einsum is a replacement for dot, for multiple dimension, then we can
have a look.
We will have to wait for a future version I'm afraid. Probably a very
distant future. There doesn't seem to be any developers interested in
this problem on the numpy or scipy mailing lists, as far as I can
tell.
But I wont use scipy.
This is best. It was just a suggestion for understanding the problem
and possible solutions, and more for reading the Scipy code for these
functions. I.e. just a learning exercise.
The thing I would like more, is to eliminate the looping when doing the
matrix exponential.
I have just completed a test/profiling script, which strides through the
data.
It took it me quite long to grasp the striding thing, but I think it is
worth my time to learn how to access data in any possible way.
We can gain 1.5 in speed ! :-)
On top of the current gains, this will be quite significant. I can't
help you with the striding concept as I'm not too familiar with it.
This seems similar to pointer arithmetic in C for accessing matrix
data. Is this essentially collapsing the rank-7 [NE, NS, NM, NO, ND,
x, y] structure into a rank-3 structure and looping over the first
column. This combined with maybe a Pade approximation instead of the
much slower eigenvalue decomposition method might be the fastest.
Regards,
Edward
Re: r24294 - in /branches/disp_spin_speed: lib/dispersion/ns_r1rho_2site.py target_functions/relax_disp.py