BENCH: set up benchmarks for the core transforms.#22
BENCH: set up benchmarks for the core transforms.#22DWesl wants to merge 5 commits intojswhit:masterfrom
Conversation
I was poking at the idea of modernizing the Fortran code, and there's a chance it could affect performance. There's a decent chance the compilers see if-then-else the same way as a block of gotos, but if not the more structured code might be easier to optimize.
|
Re-run of the benchmarks after passing complex coefficients to |
|
Thanks for all your recent contributions @DWesl! Not sure how I feel about touching the ancient NCAR fortran code though - seems like a slippery slope. The only reason I've found to use this instead of more modern libs (like SHTns) is the ability of SPHEREPACK to use a regularly spaced latitude grid (including the poles) without doubling the number of latitudes. |
My first thought was to wrap the files not exposed to python in Fortran 90 modules, but that broke the Python binding, so I'm inclined to agree with you.
I think there's one other, which uses Clenshaw-Curtis quadrature rather than Gaussian for that integration, but it seems it upscales the grid internally as well. I think there's a paper some while back showing exact transforms on a regular lat-lon grid can't happen without 2N latitudes, but frequently what we can get with N+1 latitudes is fine. (There's other papers investigating similar things with how well Clenshaw-Curtis quadrature deals with polynomials beyond the degree for which it is exact, compared to Gaussian quadrature, and extending that to certain rational functions). As partial evidence for that, the fast Legendre transform in the IFS is also approximate. |
@DWesl |
3 participants
I was poking at the idea of modernizing the Fortran code, and there's a chance it could affect performance. There's a decent chance the compilers see if-then-else the same way as a block of gotos, but if not the more structured code might be easier to optimize.
Benchmarks based on airspeed velocity. Run directions and results below.