Closed RobertPincus closed 1 year ago
Note that @MennoVeerman's comment is not addressed
Hi @RobertPincus, this looks great! I'm curious what sort of solution tolerance you're achieving with single precision. We're testing RRTMGP.jl with Float32, but the errors are quite large at the moment. cc @sriharshakandala
@charleskawczynski Feel free to review the PR :-).
In the test problems in our suite we see difference in single precision with respect to double precision of up to .13 W/m2. (See the steps labeled "Relax failure thresholds for single precision" in the updated CI workflow files.) @peterukk has a different formulation for the minimum value of k that reduces errors further. I may implement that and/or other alternatives now that automated testing of single precision is in place.
This is mostly a first implementation of single precision (finally addressing #156) following #39. The stand-alone tests are heavily revised - solutions to equivalent problems (e.g reversed vertical orientation) are tested with respect to numerical precision. Single precision fluxes differ from double precision in sample problems by as much as .13 W/m2.