(towards #75) Adds original tracer advection benchmark

[arporter]

• Original (serial) benchmark
• OpenACC version of original
• OpenMP version of original
• Compute moved to subroutine
• OpenACC version of compute-in-subroutine
• SIR-compliant Fortran version
• OpenACC added to SIR-compliant Fortran

[arporter]

OpenMP version runs on my desktop but doesn't show any performance benefit. However, OMP is not in the ESIWACE2 deliverable so I'm going to park that for now.

[arporter]

I need to add a checksum output to ease verification.

[arporter]

Now have version with compute moved to subroutine working on GPU. However, can see that we get managed-memory traffic at the start of each compute region:

I think this must be because a lot of the work arrays are done as automatic arrays and thus are re-allocated on the GPU each time the subroutine is called.

[arporter]

Made the automatic arrays into module-scoped allocatables that are allocated just once:

[arporter]

Presumably @rupertford, this solution won't work for SIR because I now have an allocate in the compute routine itself? I could move it out to an init method for the module?

[rupertford]

Presumably @rupertford, this solution won't work for SIR because I now have an allocate in the compute routine itself? I could move it out to an init method for the module?

I'm not actually sure. It may be OK as you can specify data as being local in the SIR which presumably means scoped within the code generated by SIR. But I've not looked at what gets generated.

[rupertford]

Looks great. Just a few comments inline.

[arporter]

Ready for another look from @rupertford now.

[rupertford]

Just a couple of docstrings that have not been updated. Please see inline.

[arporter]

OK, this should be ready to go now.

[rupertford]

All good now.