Repository navigation
[FEATURE] Instantiate (linear) models from .wav files - #337
Merged
Merged
Conversation
dfernandes83
added a commit
to dfernandes83/NeuralAmpModelerCore
that referenced
this pull request
Sep 21, 2026
Brings in NAM Core 0.6.0 from sdatkinson/NeuralAmpModelerCore: - models with variable sample rate (sdatkinson#333): Linear reports SupportsArbitrarySampleRate() and resamples in Reset() - 1-to-n and n-to-1 channel configurations in Linear models (sdatkinson#335, breaking: a Linear model now needs one kernel per output/input path) - Linear models instantiated from .wav files through get_dsp() (sdatkinson#337) - version bump to 0.6.0 (sdatkinson#336) Conflicts: - NAM/get_dsp.cpp: kept all three includes (<algorithm> and <cctype> from upstream, <cmath> from our validation code). - NAM/linear.cpp: took upstream's rewritten direct-convolution loop. Our side only had a static_cast to silence an MSVC conversion warning in the code that upstream replaced. Consumers now have to compile NAM/wav.cpp too: get_dsp.cpp calls nam::detail::load_wav_ir().
This was referenced Sep 24, 2026
squatchlr
added a commit
to Squatch-Stack/NeuralAmpModelerCore
that referenced
this pull request
Oct 15, 2026
Closes sdatkinson#327 ## Summary Each architecture registered itself from a static object in its own translation unit (`static ConfigParserHelper _register_X(...)`). When Core is linked from a static archive, nothing references those units. The linker drops them, their initializers never run, and every load fails with `No config parser registered for architecture: …`. Even in a normal link, the registry was empty until those units had been initialized. `ConfigParserRegistry::instance()` now registers the six built-ins itself, from one list in `get_dsp.cpp`, inside its function-local static. Naming each `create_config` there means any program that uses the registry links every built-in, however Core is linked. The initialization is thread-safe, so concurrent first calls register once. The per-file registrars are removed, since they would now register duplicates. This is option 2 from the issue. Calls to `registerParser` are real references, so no `volatile` array is needed, and callers don't have to call an init function. `get_dsp.cpp` was chosen over a new file so that projects listing Core's sources by hand need no change. External registration (`factory::Helper`, `ConfigParserHelper`) works as before, and `test_extensible` still passes. sdatkinson#228 takes the same approach as part of a larger library PR. If `get_dsp` moves into its own library (sdatkinson#294), the list moves with it. ## Tests `tools/test/test_builtin_registration.cpp`: - all six built-ins are registered; - `WaveNet` is already registered when the registry is first read during static initialization. That fails on `main`, because `run_tests.cpp` is initialized before the architecture files: `Assertion failed: (kWaveNetRegisteredDuringStaticInit)`; - registering a built-in name again still throws. The issue's own reproduction links the same objects directly, then through `ar`. We extended it to more model types: | AppleClang 21 | `main`, via `ar` | this PR, via `ar` | |---|---|---| | `A2.nam`, `slimmable_container.nam` | SlimmableContainer not registered | loads | | `wavenet.nam`, `wavenet_condition_dsp.nam` | WaveNet not registered | loads | | `lstm.nam` | LSTM not registered | loads | | a 64-tap `.wav` IR (sdatkinson#337) | Linear not registered | loads | The PR also loads with `-O3 -flto -Wl,-dead_strip`. On GCC 14.2 (Pi 5), `main` fails and the PR loads the same six cases, with `-ffunction-sections -fdata-sections -Wl,--gc-sections` and with `-flto -Wl,--gc-sections`. ## Verification - **AppleClang 21 and Linux clang 18.1.3 (libstdc++ and libc++):** CI's steps pass, in Debug and Debug with `NAM_USE_INLINE_GEMM`. - **GCC 14.2 (Raspberry Pi 5):** `run_tests` passes, and the new test fails first there too. - **Concurrency:** ThreadSanitizer, with 8 threads making their first `get_dsp` call through a static archive at once: clean in 5 of 5 runs, and all 8 threads load every time. - **Output:** 36 renders are byte-identical to `main`. Nothing on the audio path changes. ## Not covered MSVC and Windows. Programs that never touch the registry are unaffected either way. Co-authored-by: Squatch Stack <321440025+squatch-stack@users.noreply.github.com>
silvercobraa
added a commit
to Insapiable-Inc/NeuralAmpModelerCore
that referenced
this pull request
Oct 18, 2026
Brings the FiLM fix for non-blended gating (sdatkinson#321, sdatkinson#322), .nam file validation (sdatkinson#315), the hot-path allocation removal (sdatkinson#312), cached A2 prewarm (sdatkinson#319), faster long linear convolutions (sdatkinson#324), variable-rate and 1-to-n / n-to-1 Linear models (sdatkinson#333, sdatkinson#335), Sequential models (sdatkinson#323) and Linear models from .wav files (sdatkinson#337). Conflict: test_wavenet_configurable_gating — upstream fixed the same short weight list (sdatkinson#316); their fix kept. Upstream's new tests are moved to the checked weights_iterator (assert(it.at_end()) instead of comparing with weights.end()). 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.
A .wav file is often used to hold the impulse response of a system. This can be used to define a linear model (as is super-common in digital audio).
This implements a loading routine that does a (lossless) conversion from a .wav IR to a linear model.
Resolves #157