[daip] Problem with 31DEC05 IMAGR with CMETHOD ''
Eric Greisen
egreisen at nrao.edu
Wed Dec 14 16:55:53 EST 2005
Subhashis Roy writes:
> It appears that with CMETHOD '', the present algorithm (07DEC05) is not correct to
> choose the fastest method when IMSIZE=128i. It is not selecting GRID at
> all in this case (It appears to be fine when IMSIZE=256).
> (this run is on i386 compiled with gcc-2.95.3).
>
> This is making it extremely slow (last step of removing ~10000 comp took
> 15 minutes) on my amd64 2.2GHz machine (~3.2 GHz P-IV). Clearly, FFT is
> much faster here.
>
Yes - there is an issue here not only of speed but of accuracy. We
discovered that our source models did not work correctly when we
provided a small subimage together with the CC file. The default
method was always gridded and the results were bad because it made a
small grid and then tried to interpolate over very large u-v cells.
We now force DFT when the image is not a power of 2 or when the images
average <= 128. That is pretty arbitrary and could be set < 128.
I should try things for a while before I am willing to change this
however. We did find another problem with the models which might have
been the real cause of the earlier trouble. As I say, I will have to
experiment.
You could experiment too - the subroutine is $APLSUB/FORCED.FOR. Edit
it, then
COMRPL $APLSUB/FORCED
COMLNK $APGOOP/IMAGR
Eric Greisen
More information about the Daip
mailing list