h0ffmann / ww3-gpu

ww3-gpu

without changing the answer

making WAVEWATCH III run faster with Kokkos and CUDA, without changing the Fortran's arithmetic: every ported routine gives the Fortran's bits.

the repository the port in rust: chaceon

the repository

Without Changing the Answer

WAVEWATCH III on GPUs with Kokkos and CUDA, bit for bit with the Fortran: github.com/h0ffmann/ww3-gpu

the code, the plans, the decisions and the measurements sit side by side. every claim carries (v) when it was checked and ⚠ when it was not, and every number comes with its command.

progress (pt-BR) · decision records (pt-BR)

the decisions

each one a choice the lab made, the context that forced it and what follows. an accepted record is never edited; a new one supersedes it.

ADR-0001No proof language for the GPU port; SMT queries and exhaustive sweeps insteadaccepted2026-10-08
ADR-0002Port order is measured wall time, most expensive firstaccepted2026-10-08
ADR-0003The repository is a book in progress on agentic coding and agentic researchsuperseded2026-10-08
ADR-0004The book's channels: GitHub as the source, a PDF per release on Zenodo, a free Leanpub edition generated by CIsuperseded2026-10-08
ADR-0005The Claude project's settings are snapshotted in the repository, dated, except memory, connectors and secretsaccepted2026-10-09
ADR-0007Graphify is adopted as an offline locator of WW3 module dependencies, built on preprocessed sources at the pin; it is never cited and never answers who calls whatproposed2026-10-11
ADR-0008O trabalho passa do projeto pessoal (Max) para a organização Team da colaboração com o LOSUFBA pelo repositório; threads, memória e rotinas não migram, e o projeto só muda de conta quando o Team tiver Projectsproposed2026-10-11
ADR-0009ww3-gpu adopts ai-memory from Fabio Akita, and with it the citation discipline, the second brain and the Gleam website of its sister repositoriesproposed2026-10-14
ADR-0010The Kokkos port has no tests of its own; WW3's regtests are its only testsaccepted2026-10-14
ADR-0011The port moves in large units, with discipline; WW3's regtests run at checkpoints, not after each routinesuperseded2026-10-15
ADR-0012Translate and wire now, with static parity and one review; the full WW3 regtest matrix of the translated units runs on the UFSC H100 nodeproposed2026-10-16
ADR-0013The parts of GFS-Wave's physics default to Kokkos once their checkpoint and a timing with parity passaccepted2026-10-16
ADR-0014A GPU run is judged by nccmp-tol's tolerances; the CPU backend stays bit for bitaccepted2026-10-17
ADR-0015Triton kernels are allowed as an exception to "no Python in lab code", under three conditionsaccepted2026-10-17
ADR-0016The goal is WW3's run time on Kokkos and CUDA; WW3's regtests move to a later validation phase; one person works in a short compound loopaccepted2026-10-17
ADR-0017The Kokkos port is single node; an MPI build of WW3 keeps the Fortran pathaccepted2026-10-18
ADR-0018Where ww_kokkos lives if WW3 upstream takes the Kokkos runtimeproposed2026-10-18
ADR-0019GoogleTest for the runtime infrastructure that ports no WW3 routineproposed2026-10-18
ADR-0020MPI enters the Kokkos port, with the exchange built on the resident spectrumproposed2026-11-01
ADR-0021Compound Engineering is dropped; engineering work stays sized to the changeaccepted2026-11-11
ADR-0022Every ADR is a section of one document, docs/ADRs.mdaccepted2026-11-11
ADR-0023The course and the book leave the repository; its texts are the preprint, the proposal and the thesisaccepted2026-11-11

the designs

a change is designed before it is built, tied to a deliverable of the project proposal, with its definition of done.

WFIP-0001WW3 samples driven by WeatherNext 3 wind on an RTX 4090 boxpostponed2026-10-08
WFIP-0002The lab as a book: compiled from the repository, built in CI, released with each versionsuperseded2026-10-08
WFIP-0003The Kokkos DIA through WW3's own regtest harness, on the h0ffmann/WW3 forkimplemented2026-10-14
WFIP-0005The device-resident time step: VA lives on the GPU, only forcing in and output outdraft2026-10-17
WFIP-0006The device computes glibc's libm bits: routing the port's per-point math through ww::libmpartially2026-10-17

the port, measured

how much of WW3 runs on Kokkos, then every number of every port report, step by step. a bar is the speed-up over its own baseline, on a log scale; the deeper the green, the larger it is. a step slower than its baseline is grey, and says so.

99.3 %of the run time, translated316 of 1 023 routinesa Kokkos translation exists in kokkos/src; the profile's self time, full (bench_small, 120 x 80, 9 520 sea points, 12 h)
99.3 %of the run time, ported72 of 1 023 routinesWW3 at the WW3/ pin calls the Kokkos code, behind its run-time flag; the profile's self time, full (bench_small, 120 x 80, 9 520 sea points, 12 h)
48.6 %of the run time, regtested14 of 1 023 routinesWW3's regtests ran with it on Kokkos and agreed with the Fortran; the profile's self time, full (bench_small, 120 x 80, 9 520 sea points, 12 h)
26.21 %of the lines of code, translated69 024 of 263 375Fortran lines of the translated routines, of every line in WW3's model/src, in 65 of 147 files
7.24 %of the lines of code, ported19 059 of 263 375Fortran lines of the ported routines, of every line in WW3's model/src, in 28 of 147 files
1.20 %of the lines of code, regtested3 151 of 263 375Fortran lines of the regtested routines, of every line in WW3's model/src, in 7 of 147 files
62 651lines of Kokkos C++in 355 filesunder kokkos/src: the port itself

AMUXD_KOKKOS ported, phase 1 (translated)

regtested no: translated only, as the owner asked on 2026-10-16 (no test run); its checkpoint (ADR-0011) needs a shim and its caller's hunk