Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
19 commits
Select commit Hold shift + click to select a range
54daaa5
build(windows): link the core-linked C tests through IFX in the stati…
SMI-Lab-Inha Oct 30, 2026
4c1aadc
fix(deck): a number beyond the range of the kind is not finite on eve…
SMI-Lab-Inha Oct 30, 2026
d6c9333
fix(io): find a missing input file by INQUIRE before the OPEN
SMI-Lab-Inha Oct 30, 2026
8c266d5
build: automatic local variables with every compiler and OpenMP setting
SMI-Lab-Inha Oct 30, 2026
9021de6
fix(dynamics): no closures over a model in the line step callbacks
SMI-Lab-Inha Oct 31, 2026
46c2f83
test(threads): concurrent steps on every handle route and a source audit
SMI-Lab-Inha Oct 30, 2026
c9ba281
fix(io): no OPEN that would fail, so a long name cannot end the ifx r…
SMI-Lab-Inha Oct 30, 2026
c1b8548
build(windows): /fp:precise for the static Intel release
SMI-Lab-Inha Oct 30, 2026
585617a
build(openfast): automatic local variables in the Visual Studio project
SMI-Lab-Inha Oct 30, 2026
aa9250d
fix(diagnostics): per-system and per-workspace counts, no process-wid…
SMI-Lab-Inha Oct 30, 2026
e14d982
test(driver): a profile total of 0.000 s is 0.0 % of itself
SMI-Lab-Inha Oct 30, 2026
8ee30fc
build(intel): /fp:precise for every ifx build and the OpenFAST projects
SMI-Lab-Inha Oct 31, 2026
6d68a5f
build(intel): heap arrays above 1 MiB in every ifx build
SMI-Lab-Inha Oct 31, 2026
4465cfb
fix(release): openfast.exe names its OpenFAST base and CableDyn versi…
SMI-Lab-Inha Oct 31, 2026
8faa89a
perf(waves): block sine, cosine and exponential in the wave kinematics
SMI-Lab-Inha Oct 31, 2026
0e2002b
fix(waves): NaN through the block sine, cosine and exponential
SMI-Lab-Inha Oct 31, 2026
1fc5af9
build(waves): IEEE semantics for the block kernels under GNU fast-math
SMI-Lab-Inha Oct 31, 2026
e1271e8
test(waves): the block kernels over their whole range and at their zeros
SMI-Lab-Inha Oct 31, 2026
bc20c71
fix(capi): pause the OpenMP pool on Close only under GNU OpenMP
SMI-Lab-Inha Nov 2, 2026
File filter

Filter by extension

Filter by extension


Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 2 additions & 0 deletions .github/workflows/release-windows-static.yml
Original file line number Diff line number Diff line change
Expand Up @@ -32,6 +32,8 @@ jobs:
repository: OpenFAST/openfast
ref: 2895884d2be01862173c88d70f86b358d2f1a50a
path: openfast
# the history and tags: openfast.exe names its base by `git describe --tags`
fetch-depth: 0
persist-credentials: false

- name: Check out the pinned OpenFAST regression model
Expand Down
1 change: 1 addition & 0 deletions ARCHITECTURE.md
Original file line number Diff line number Diff line change
Expand Up @@ -446,6 +446,7 @@ src/
CableDyn_Hydro.f90 Morison loads, Airy and JONSWAP waves, currents
CableDyn_WaveSpectra.f90 spectra, spreading, and multi-train seas
CableDyn_StreamWave.f90 stream-function regular nonlinear waves
CableDyn_WaveMath.f90 block sine/cosine and exponential of the wave kinematics
CableDyn_Vessel.f90 vessel kinematics, 6-DOF motion, displacement RAOs
CableDyn_SeabedContact.f90 C1 normal contact and stick-slip friction laws
CableDyn_Bathymetry.f90 gridded bathymetry
Expand Down
76 changes: 76 additions & 0 deletions CHANGELOG.md
Original file line number Diff line number Diff line change
Expand Up @@ -294,6 +294,16 @@ All notable changes to CableDyn are recorded in this file. The format follows

### Changed

- Every Intel ifx build compiles CableDyn with the value-safe floating-point model,
`/fp:precise` (`-fp-model precise` off Windows): the static Windows release and wheel, every
other CMake build and configuration with ifx (with or without OpenMP, including the core that
the OpenFAST CMake route builds, and its adapter), and every configuration of the OpenFAST
Visual Studio project. The ifx default, `/fp:fast`, reassociates sums, divides by
multiplying with the reciprocal and flushes subnormals to zero, differently at every inlined
call site, so the gates that hold two routes to the same bits failed (a zero control command
against an uncontrolled cable, a held Coupled End B against a Fixed one, a committed against a
re-derived torque), subnormal deck values read as zero and the clamped-cable elastica secant
stalled. Results of ifx builds change at round-off; gfortran builds are unchanged.
- The macOS wheel needs macOS 15 or newer on Apple silicon (0.3.0: macOS 14). It is built on
the macOS 15 runner, as GitHub retires the macOS 14 runner image on 2 November 2026, and its
Homebrew gfortran and OpenBLAS support no older macOS. On macOS 14, build from source.
Expand Down Expand Up @@ -336,6 +346,23 @@ All notable changes to CableDyn are recorded in this file. The format follows
0.1 s drops from 18,000 components to 1,800. On the 300 s lazy-wave case both records then run
in 72 s, and every output channel's standard deviation moves by at most 0.002 %. A record with
no component at or below the cut-off is refused by name.
- Waves run faster under every compiler: the wave kinematics evaluate their phase sines and cosines
and depth exponentials in blocks of plain arithmetic that vectorise under any value-safe
floating-point model (`CableDyn_WaveMath`: one argument reduction for each sine-cosine pair,
polynomials, the power of two from its bits) in place of one library call per component, which ifx
makes under `/fp:precise`. At one thread, best of five, an irregular sea runs 31 to 33 % faster
with ifx (`chain_two_train_sea` 1.83 s to 1.22 s, `chain_torsethaugen_spread` 1.64 s to 1.13 s)
and 20 to 21 % faster with gfortran (4.09 s to 3.22 s, 3.56 s to 2.86 s); the VolturnUS line-break
decks 14 % (ifx, 31.0 s to 26.8 s) and 9 % (gfortran); `lazy_wave_vessel_rao` and
`lazy_wave_bend_stiffener` 9 % with ifx and within 1 % of before with gfortran; decks without
waves run as before. This recovers about two thirds of what `/fp:precise` cost the Windows release
(`chain_two_train_sea` +83 % over `/fp:fast` before, +24 % now). The sines and cosines are within
2.4 ulp (within 9e-26 next to their zeros) and the exponentials within 1.1 ulp of quadruple
precision, so wave decks change at round-off: 49 of the 53 example decks give identical files, two
lazy-wave decks differ by one unit in the last printed digit (7e-8 of a tension's scale), and in
the two VolturnUS line-break decks the drifting platform amplifies round-off after t = 860 s to
0.15 to 0.21 of the anchor-2 tension's scale, as much as the gfortran and ifx builds of the
unchanged code differ by.
- Coupled OpenFAST runs start from the equilibrium in SeaState's steady current (`CurrMod 1`),
as standalone decks start in their `current` option, instead of from the still-water shape.
The current is evaluated at SeaState's vertical grid levels for the static solves only; the
Expand Down Expand Up @@ -391,6 +418,20 @@ All notable changes to CableDyn are recorded in this file. The format follows
is now laid from End A, the line seen from its other end, and its nodes put back in order. The
deck starts at the same equilibrium as before on Windows (1e-7 m, 6e-4 N), and moving the fixed
rod by 1 ulp either way or 1e-12 m up leaves the rod's start unchanged (0 m).
- `CableDyn_Close` no longer pauses the OpenMP runtime of the whole process under the Intel and
LLVM OpenMP runtimes (ifx and other non-GNU builds with OpenMP). Every close called
`omp_pause_resource_all` (since 0.1.0) to join the calling thread's pool, a fix for GNU
OpenMP's process exit on MinGW-w64; libiomp5 and libomp pause the runtime of every thread
instead, so closing one handle paused the OpenMP of every other handle and of the host, and the
next parallel region on another thread deadlocked resuming it. An ifx build with OpenMP that
stepped a handle on a second thread after a close hung every time (the threading gate, 20 of
20 runs); it now passes 20 of 20 with 1 and with 4 OpenMP threads. GNU builds keep the pool join.
- A non-static ifx build (a plain CMake build with ifx, or the core of the OpenFAST CMake route)
kept run-time-sized arrays and array temporaries on the stack, as only the static Windows
release compiled with `/heap-arrays`: the 2835-node mixed IEA-15 MW deck and a fine-mesh deck of
the statics matrix overflowed the 1 MiB stack of an MSVC-linked program. Every ifx build now
compiles with `/heap-arrays:1024` (`-heap-arrays 1024` off Windows), as the static release and
the OpenFAST Visual Studio project did; the static release is unchanged.
- A rod numbered as a body no longer joins the body's attachment points. A `Rod<N>A/B` end point
keeps its rod's id where a `Body<N>` point keeps its body's, and the body's point list took
every point with its id: a rod end POINT row listed before the body's points became the body's
Expand Down Expand Up @@ -460,6 +501,36 @@ All notable changes to CableDyn are recorded in this file. The format follows
share: a free spar rod with a 55 m pinned cable (EI 2e4 N m^2) started 5.7e-2 m off its
equilibrium, drifted 7e-2 m in 3 s and moved a leg tension by 2.3 %. Such a line now takes the
static passes too (4 here); the rod holds its pose to the output's 8 digits.
- Diagnostic counts that concurrent models shared: the adaptive-substep fallbacks of an `EI = 0`
system and the assembly and modified-Newton counts of the finite-EI Cosserat solver were
process-wide (and lost increments without OpenMP); they are now kept per system and per solver
workspace. The count of finite-EI intervals recovered by substeps is reported only by the opt-in
run profile (`CABLEDYN_PROFILE`), whose figures stay process-wide by design; each model keeps its
own. No result changes.
- A failing `OPEN` of a file name of about 600 characters or more (the extended-length spelling of
a path beyond the Windows limit) ended an ifx-built program or the wheel's DLL with
`STATUS_STACK_BUFFER_OVERRUN` inside the Intel Fortran run-time, for example an output root in
a folder that does not exist. Every `OPEN` of a user-supplied name now first asks the system
whether the file opens (`CD_Path_Openable`) and reports the failure without attempting the
`OPEN`; the messages are unchanged.
- Concurrent steps of separate handles crashed the ifx builds with access violations: an `EI = 0`
line model passed its load and added-mass callbacks to the integrator as internal procedures,
and ifx calls such a closure through a trampoline kept in memory every thread shares, so two
threads stepping distinct models called each other's half-written trampolines (a 40-line
reproducer crashes with two threads under every ifx option). The model now hands its loads to
the integrator as an object (`CD_Cable_Step_Loads`); results are unchanged. Seven decks
stepped together on 28 threads (`c_api_threads_mixed`) failed every run before and pass 50 of
50.
- Separate handles used on several threads at once failed in the Windows wheel and the release
binaries (Intel ifx without OpenMP): ifx kept the fixed-size local arrays and `CHARACTER` locals
of every procedure in static storage, which all threads share, so an initialisation that ran
while other handles stepped stopped with "tangent assembly failed" or "band scatter" errors
(gfortran builds were not affected: `-fopenmp` makes every local automatic). Every build now
makes local variables automatic, `/Qauto` (`-auto`) with ifx and `-frecursive` with gfortran,
which `-fopenmp` already implied, so the default gfortran build is unchanged, and so does every
configuration of the OpenFAST Visual Studio project (`LocalVariableStorage`). The release
benchmark decks run as fast or faster (0.3 to 21 %); ifx results change at round-off (with
`/fp:precise` every example is bit-identical).
- In OpenFAST, Free Rigid6 bodies and Free, Pinned and CoupledPinned rods started their first
step from zero acceleration: the module restores its state mirror after the output call at
`t = 0`, which marked the objects' accelerations as committed before the start accelerations
Expand Down Expand Up @@ -536,6 +607,11 @@ All notable changes to CableDyn are recorded in this file. The format follows
side, or by the sign of a round-off component: in OpenFAST about half the rows of a cable in
the x-z plane read 359.9x. A line whose chord heads within 45° of +X now reports its azimuths
in (−180, 180], every other line in [0, 360) as before.
- The CableDyn-enabled `openfast.exe` named itself `OpenFAST-v5.0.0-dirty`: OpenFAST stamps its
version with `git describe --dirty`, and the integration patches modify the checkout. Built
through the Visual Studio solution or through CMake it now reads
`OpenFAST-v5.0.0-CableDyn-<CableDyn version>`, the describe of the unpatched OpenFAST base and
the CableDyn version, and the release build fails unless `openfast.exe -v` prints exactly that.

## [0.3.0] - 2026-10-23

Expand Down
Loading
Loading