Andy Porter
Andy Porter
I was deliberately not using a tolerance to make this test more sensitive (although I had to back-off to -O1). For passthrough on CPU, I'd prefer to avoid a tolerance....
> Unless you know what could it be, we will need to repeat it with the debugger to see where it fails. I'm afraid I have no idea. Perhaps we...
> I am also finding false positive, which are more concerning because they lead to wrong arrayrange2loop transformations. For example for `a(:) = b(:)` we say its the same range...
There may already be a utility method to do this in fact.
The updates to the halo-exchange logic are complicated because, previously, redundant computation could only be performed to either a literal depth or to the maximum halo depth. If a kernel...
I've removing all of the 'literal_depth' related stuff - we have a single depth and it's always PSyIR. Unfortunately the multi-grid tests are failing (and probably lots of others).
@tommbendall this branch is ready for you to try. @sergisiso I think this is probably one for you as it touches a lot of things :-( Essentially it adds support...
> Finally, if I understand right, is the `halo_depth` argument the final argument at the algorithm layer, but the second argument (after `nlayers`) at the kernel layer? Yes that's right,...
CI-happiness permitting, this is ready for review again.
This is ready for another look now @sergisiso and @tommbendall. I'll fire off the integration tests again now too.