chaceon

WAVEWATCH III® in pure Rust, on the GPU, through Rust-CUDA and rust-gpu. It is compared with metocean/WW3 7.14 (7.14.mos-92e2cf6), and it follows the discipline of ww3-gpu without Kokkos.

State, 2026-11-04: a run with no Fortran program in it exists. Rust's ww3_grid, ww3_strt, ww3_prep (for the case with a forcing file) and ww3_shel, in sequence from the namelists, give mod_def.ww3, current.ww3, out_grd.ww3, out_pnt.ww3, track_o.ww3 and both restarts byte-identical to the reference on 96 files (v, just rust-only). ww3_prep takes every input type, and ww3_strt every start type (prep, strt). ww3_prnc reads NetCDF forcing for types AI and LL, byte-identical on 81 cases whose data are synthetic, because the regtests' NetCDF files are not in WW3's repository (report). ww3_shel takes water levels, IC0 ice, file winds and boundary input from nest.ww3, which Rust's ww3_bound writes: 214 comparisons identical over 34 boundary cases (v, just bound-all), on the host only (bound, W3IOBC). After the run, ww3_ounf and ww3_ounp write NetCDF files byte-identical to the reference's through the system libnetcdf (ounf, ounp), ww3_outf and ww3_outp write the same text files on 40 regtest inputs (outf, outp), and ww3_trck and ww3_trnc the same track files (v, just track-all, 74 comparisons). Spectral partitioning fields are byte-identical too (partitions). ww3_bounc is not ported. ww3_shel takes --threads N and, built with the feature mpi, mpirun -n R, with the same bytes (v, just mpi-pair). With W3SRCE, W3XYP3 and W3KTP3 as compiled SPIR-V on CPU Vulkan and VA resident on the device, the same files are byte-identical too. Batch 7, the alternative physics, is under way at the routine level: W3SBT4, MetOcean's bottom friction, is called from W3SRCE in a BT4 build (report). Thirty more routines, thirteen of ST1, FLX1, ST3, IC1 and IS1 and seventeen of ST2, ST6, FLX2, FLX3 and FLX4, have every captured call bit-identical on the host and, for the twenty that run on the device, in SPIR-V (plan, "Batches 7 and 8"). No model run uses any of them yet. Nothing has run on a real GPU, the PTX has never executed, and there are no timings: this is a verification result, not a speed one.

Port status

Generated by just port-summary from WW3/model/src at 92e2cf66 and rust/ports.tsv.

WW3 area	Files	Fortran lines	Not comments	Routines with Rust, lines	Share	Partial
Source terms	38	38 360	12 613	11 987	31%	0
Propagation, structured grids	7	12 190	4 783	3 059	25%	0
Propagation, unstructured grids	5	20 388	10 108	0	0%	0
Time loop, updates, state	20	45 309	20 815	7 657	17%	4 003
File I/O in the model	10	16 697	8 463	11 570	69%	629
Grid search and metrics	1	9 654	5 652	3 090	32%	30
Drivers	12	30 440	15 373	0	0%	2 921
Preprocessors	22	39 924	20 309	9 637	24%	8 064
Postprocessors	14	28 619	17 648	11 377	40%	7 295
Coupling and other	6	10 538	4 941	0	0%	0
All of model/src/*.F90	135	252 119	120 705	58 377	23%	22 942
Other model source: SCRIP/, PDLIB/, .f, .f90, .c	30	26 736	13 105	0	0%	0
All model source	165	278 855	133 810	58 377	21%	22 942

Routines with Rust sums the whole Fortran routine for each row of rust/ports.tsv with a verdict on the code reached, comments included, so a share is an upper bound on what was exercised, and it counts only the switches of switch_chaceon_shrd. Partial is translated: Rust exists and is not fully reached.

Rust by evidence, code lines	Lines
Real GPU: byte-identical whole runs	0
SPIR-V on CPU Vulkan (lavapipe); its kernels also natively on the CPU	5 189
Native on the CPU only	10 080
Every captured call identical, no whole-run verdict	3 090
Partly reached, or compiled and never run	10 870
Rust ported	29 229
Verification tools, not port code	5 936

Badges compare code lines on both sides: 133 810 Fortran and C lines of all model source that are not comments or preprocessor lines, every switch included, out of 278 855 lines with comments; WW3's regtest programs are not port targets. They are compared against Rust lines that are not blank or // comments. The two languages' lines are not the same unit, so the ratio is not a percentage of the port. CPU validated is the first two rows. SPIR-V on CPU Vulkan is compiled for a GPU API but runs on a CPU, so it does not count as GPU validated. A Rust file counts at the weakest ledger row pointing at it, which never overclaims and understates shel.rs and gridpre.rs, partly reached only through W3WAVE and W3NMLGRID.

Rust, all lines	Lines
ww3-kernels, runs on the device	6 492
ww3-libm, glibc's bits in plain Rust	1 273
ww3-host, the host side	20 557
GPU entry points and builders	1 900
Tools	9 415
All	39 637

Verdict, lines of Fortran routines	Lines
spirv: byte-identical whole runs on CPU Vulkan	7 828
host: byte-identical whole runs on the CPU	43 879
routine: every captured call identical	6 670
translated: exists, not fully reached	22 942

What still depends on Fortran. No Rust binary links Fortran. A run still reads files that Fortran programs wrote where Rust stops: mod_def.ww3 from ww3_grid for unstructured and SMC grids, formatted or unformatted data files (IDFM 2 and 3), the old ww3_grid.inp and tripole closure; the forcing file of ww3_prnc for input types AT, F1, F2 and DAT and of either forcing program on an unstructured grid; and nest.ww3 from NetCDF spectra, which only Fortran's ww3_bounc writes (Rust's ww3_bound writes it from text spectra). Rust's ww3_shel takes water levels, ice and boundary input on the host and refuses them on a device. ww3_prep and ww3_prnc are 4 470 lines, 2% of model/src. Every verdict also needs the Fortran reference, by design: it is what Rust is compared with.

Read	For
AGENTS.md	The rules
docs/RUST_GPU_PLAN_202610.md	The plan, in batches
docs/FLOAT_REPRESENTATION_202610.md	f32 and f64 against Fortran's kinds
docs/LEAN_202610.md	What Lean proves here, and what it cannot
docs/WW3_GPU_PARITY_202611.md	Feature parity with ww3-gpu, switch by switch
docs/ADRs/	The decisions
VERSIONS.md	What is pinned

Licence

MPL-2.0, translated files included (ADR-0006). WAVEWATCH III is a trademark of the NWS.