On Tue, Oct 15, 2019 at 04:10:34PM +0200, Tobias Burnus wrote:
> But for "omp update" with noncontiguous memory, it is not clear what's
> faster:
> * Simply updating one contiguous memory block, enclosing the gaps
> * Doing multiple updates of contiguous memory to take advantage of memory
> gaps.
omp target update with non-contiguous array sections isn't supported yet
even for C/C++, and when it is, it would be nice to handle it efficiently.
We already have omp_target_memcpy_rect function that can copy non-contiguous
data, but again there is no plugin API to allow at least for the common
cases that some hw is able to handle (perhaps 2 or 3 dimensions). Can at
least CUDA handle that? AMD GCN?
Anyway, I'm afraid that is left for next year.
Jakub
