Repository navigation
Adding a Windows build - #1
Merged
Merged
Conversation
docs/windows.md sets out what already ports, what has to change, and the decisions that are not technical. The new Windows job is its first step: configure, build and run ctest with MSVC and Ninja, with no validation, packaging or release, to find out which of the expected problems are real. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A model prepared before the device was open reached prepare() with a host rate and block size of zero. The rate ratio was then infinite, its product with the zero block NaN, and NaN cast to int undefined: 0 on Apple silicon, INT_MIN on x64, where the Windows build threw "vector too long". With no host settings the resampler is now passthrough with nothing sized; AmpModel already refuses such a model and re-prepares it once the settings arrive. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Same validator and strictness as the macOS job. The standalone and the VST3 are uploaded as a run artefact for trying by hand; not an installer. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
pluginval.exe is a GUI-subsystem program and PowerShell does not wait for those: the step returned five seconds in, during the editor tests, and passed without a result. bash waits and propagates the exit code. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The host check is now built on Apple only: it hosts the AU and looks for the bundles in ~/Library, neither of which exists on Windows, and with it in the way AMPSIM_BUILD_TOOLS could not be turned on there. The benchmark runs at 48 kHz and at 44.1 kHz, the second with the resampler in the chain. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The footer said 48 kHz whatever rate the benchmark was given, which the 44.1 kHz run on the Windows runner made visible. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
An experiment to separate the runner's hardware from the code it runs: the benchmark alone, rebuilt with /arch:AVX2, run in the same job as the default build and interleaved with it. The step fails if the flag does not reach the compile lines, so a no-op comparison cannot pass for a result. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
/arch:AVX2 measured 27 to 38 % cheaper on the amp model on the same runner. It is set for every MSVC-style build ahead of nam_core and the plugin, so Eigen gets AVX2 and FMA. The SSE2 comparison is gone from CI; the benchmark step now fails if the flag stops reaching the compiler, since nothing else would notice. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Four more builds of the benchmark alone in the Windows job, each different from the normal build in one thing: NAM_ENABLE_A2_FAST, NAM_USE_INLINE_GEMM, NAM with whole-program optimisation, and clang-cl with lld-link. They run on the same runner in alternating order, and each is allowed to fail without costing the others their measurements. The benchmark now prints what it was built with and whether NAM took its A2 fast path, which it does silently. The bundled capture is a SlimmableContainer of two A2-shaped WaveNets, so the check looks inside containers. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
NAM's own CMake defines NAM_ENABLE_A2_FAST by default; this project builds NAM itself and had left it compiled out. The bundled capture is two A2-shaped WaveNets, and the path measured 3 to 7 % cheaper on the Mac. It is a second implementation of the same model, so the new test builds each bundled submodel both ways from the same weights and null-tests them: -122 and -132 dB apart. The -100 dB threshold fails a single weight changed by 1 % (-77 dB), and the test fails outright without the define. The CI experiments have answered their questions and are removed. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Avenir Next and Futura ship with macOS only, and are commercial faces that cannot be embedded. Figtree replaces Avenir for the interface and Jost replaces Futura for names printed on objects, on every platform. Both are OFL with no Reserved Font Name; the files are static Regular and Medium instances of the google/fonts variable fonts, with the licences beside them. Figtree is also the look-and-feel's default sans, so JUCE's own popups no longer fall back to the system font. Every height on the panel was chosen for the old faces, so each new one is scaled to put its capitals where theirs were, from cap height over ascent plus descent measured in the font files. Unscaled, the tempo reading was cut off as "120 BP..." and every printed name shrank; scaled, the three pages match the old ones closely. tools/PanelSnapshot.cpp renders each page to a PNG. It is how this was checked, and the Windows job now runs it and uploads the images. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…cally The C++ runtime is now linked statically on Windows, so neither the standalone nor the VST3 needs the Visual C++ redistributable; the packaging script refuses a binary that still imports it. packaging/windows/AmpSim.iss asks whether to install for everyone or for the current user, and puts the VST3 in the system or the per-user VST3 folder to match. It refuses a processor without AVX2, since the build would crash its host there, and requires Windows 10 2004, the first version that answers that question. The VST3 and the standalone are separate components, the licences always go with them, and uninstalling leaves the user's presets. packaging/package-windows.ps1 builds, checks and stages, and runs ISCC with the version read from CMakeLists.txt. CI builds the installer on every run, installs it both ways, installs it again over itself, and uninstalls it. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
JUCE's documentation for prepareToPlay says hosts may exceed the block size it announces. Every buffer in the chain is sized to that announcement, so a bigger block wrote past the end of them: a test that announces 64 samples and sends 1024 crashed with a segmentation fault, after an assertion in the drive pedal's oversampler. This is the likely cause of the corrupted output heard with the Windows standalone in WASAPI exclusive mode. processBlock now takes an oversized block in pieces no larger than it was prepared for. The metronome offsets the host's position by where each piece starts, so it stays on the grid. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Every Windows Audio mode goes through Windows' audio engine, which on the Audient iD4 tested by hand stopped at a 10 ms period and 44.1 kHz. ASIO uses the interface's own driver. The headers are those bundled with JUCE 9, from Steinberg's SDK under its GPLv3 option, compatible with the AGPLv3; the installer ships the SDK's licence, and CI checks that both it and JUCE_ASIO=1 make it into the build. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The macOS job used to publish on its own. A release job now runs on v* tags once the macOS and Windows jobs have both passed, checks the tag against CMakeLists.txt once, and attaches the .pkg, the .dmg and the Windows installer together, with notes covering both platforms' requirements. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The README, the user guide and packaging/README.md describe installing on Windows, its requirements, and playing through ASIO rather than Windows' shared audio path, with the guide's troubleshooting covering what came up on real hardware. CHANGELOG.md records Windows, the new typefaces, the A2 fast path and the two fixes. CLAUDE.md drops Windows from what is out of scope and describes the Windows build, CI and release job; the roadmap moves Windows to Done and leaves signing as the open question. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.