Custom Query (964 matches)

The changes made in response to ​src:ticket:195 (OpenMP undeclared variables) appear to have broken the initialisation process in serial mode.

Specifically, if I run BENCHMARK in serial mode with bounds-checking enabled, Lm=512, Mm=64 (the smallest benchmark domain), NTileI=1, NtileJ=4, I hit a bounds-checking error in LMD_SWFRAC_TILE, where some elements of MIXING(ng)%Jwtype are equal to zero and are used as indices into an array dimensioned 1:5. Going back to the place where these should be set, namely INITIALIZE_MIXING in mod_mixing.F, it appears that this is called only once, and sets only the values in the southernmost tile.

The BENCHMARK application now runs OK in OpenMP mode.

We are still testing the observation sensitivity driver. We corrected few problem in the mathematical formulation. Many thanks to Andy Moore for testing this driver. This is a tricky driver so we expect more adjustments.

I also upgraded the 4DVAR adjustment of surface forcing. What it is new? We are make the algorithm cleaner and considering additional strategies due the several surface forcing options available in ROMS. Many thanks to Gregoire Broquet for testing this algorithm.

I corrected a bug in mod_netcdf.F when reading a 2D floating-point array. I needed to use my_ncid instead of ncid.

We added the spectral and Ritz precoditioning to the incremental strong constraint 4DVAR (IS4DVAR) algorithm which uses the approximated Hessian eigenvectors and Lanczos vectors during minimization of the cost function. This precoditioning is still under testing and we will let you when it is ready to use. Many thanks to Andy Moore for coding this algorithm.

I also updated the 4DVAR adjustment of surface forcing to account for the different options in ROMS for surface tracer flux. Many thanks to Gregoire Broquet for bringing this to my attention.