Re: GNAT vs Matlab - operation on multidimensional complex matrices

[vincent.diemunsch@gmail.com]

Le mardi 31 mars 2020 21:20:45 UTC+2, Simon Wright a écrit :
> Shark8 <onewingedshark@gmail.com> writes:
> 
> > Ok then; the following should have convention Fortran IIRC:
> >
> > type tCubeReal is array (1..NumChannels, 1..NumAngles, 1..NumRanges) of mReal
> >   with Convention => Fortran;
> 
> I wondered about this; the code ran considerably slower with Convention
> => Fortran.

In Row Major (Convention Fortran) one needs to iterate on the first dimension, then the second and finally the third, to stay on contiguous regions in memory (and hence in the cache). So the loop must put the first dimension in the center :

for p in M’Range(3) loop
  for n in M’Range (2) loop
     for m in M’Range (1) loop
         ... M(m, n, p)

This is somehow counter-intuitive. But Row Major (like Little-endian with which it is closely related) has more properties (like extracting a submatrix or conforming to classical linear algebra) and is therefore preferred by scientists.

Regards,

Vincent