On Tue, Jul 3, 2012 at 3:48 PM, Bruce Fischl fischl@nmr.mgh.harvard.edu wrote:
I think that the theoretical ratio is much bigger than what you would be able to achieve in practice. It's hard to compute exactly, but for example for em_reg it should be the number of samples (which are processed independently), so something on the order of 1000. For ca_reg it should be bigger, but it's a bit more complicated as things aren't independent.
I _hope_ that the OpenMP version of mri_em_reg doesn't just try parallelising the energy evaluation - it would be better to split each transform off as a separate work item for handling by the available threads (similar to what I did on the extra-fast GPU version). This keeps the individual pieces of work big, which is good for CPUs
I think that what Akio's after are the numbers to plug into Amdahl's Law. I don't think that these are easy to work out for anything in Freesurfer. But you can put timers around the parallel sections, and see what speed up you get on those.
HTH,
Richard