Has anyone experimented with the workshare directive on whole array operations and ifort? I played a bit with this on gfortran, but found that it wasn't giving me any speedup, and I really need to resort to parallel do/end do loops to get a boost out of openmp.
Yes, that's what I found, with both gfortran and ifort. A more common situation for benefit from OpenMP is with an outer parallel do loop containing whole array operations. The -parallel option of ifort may be able to thread a whole array operation, but doesn't judge well how to benefit from the combination of vectorization and parallel.
You might note that you need a recent gfortran with options -mtune=barcelona -msse4 to get full benefit of vectorization on recent CPUs. On some of the older CPUs, there is less benefit from combining vectorization and OpenMP, but a more common reason why people forgo vectorization is to be able to start from a lower base when bragging about parallel performance scaling.
