Andy Porter comments

Results 753 comments of


                                            Andy Porter

FFSL transport in LFRic depends on a kernel that loops over halo cells

Finally getting back to this @tommbendall. I've updated the doc as per your suggestion. Hopefully I'm correct in thinking that a kernel that operates on HALO_CELL_COLUMN or OWNED_AND_HALO_CELL_COLUMN can't have...

FFSL transport in LFRic depends on a kernel that loops over halo cells

Aha! So much for my thinking :-)

FFSL transport in LFRic depends on a kernel that loops over halo cells

TODO: - [ ] Ensure that 'special' kernels (inter-grid, CMA) reject an OPERATES_ON that goes into the halos? - [x] Update loop bound generation for these new kernels (see below);...

FFSL transport in LFRic depends on a kernel that loops over halo cells

Another question for @tommbendall : what should we do if distributed-memory support is switched off? Can we 'optimise out' any calls to kernels that have `operates_on=HALO_CELL_COLUMN`? Should everything else 'just...

FFSL transport in LFRic depends on a kernel that loops over halo cells

I'm starting to get there now but am painfully aware that this is probably going to clash horribly with what @sergisiso is doing in moving LFRic to use the PSyIR...

FFSL transport in LFRic depends on a kernel that loops over halo cells

I had a brainwave and thought that perhaps I could use the existing `Dynamo0p3RedundantComputationTrans` to implement the necessary functionality. However, that only accepts an integer for the depth of the...

FFSL transport in LFRic depends on a kernel that loops over halo cells

I was going to allow halo kernels to be optimised out of the PSy layer if distributed-memory is disabled. However, that causes a lot of issues as we can then...

FFSL transport in LFRic depends on a kernel that loops over halo cells

@tommbendall just a heads-up that argument_mod.F90 in LFRic core will need to have the new iteration spaces added to it. In the copy in our test infrastructure I have: ```...

FFSL transport in LFRic depends on a kernel that loops over halo cells

Another point for consideration by @TeranIvy and @tommbendall : my 'bright' idea of changing "cell_column" to be "owned_cell_column" has more consequences than I thought. If I change PSyclone so that...

FFSL transport in LFRic depends on a kernel that loops over halo cells

Also, a question for @tommbendall: we need to think about what state the halo regions will be in after these new kernel types have been called. Are the rules the...