NAM Oversampler v1.5.0

Release notes

• Added multi-core support: this splits by phase the model processing and parallelize it across the available number of threads.
  You can enable it via the ON/OFF setting in the 'Oversampling' page and only works when oversampling is enabled as well.
  Number of threads can be set to 'Auto' or manually to a specific number, 'Auto' should work pretty well but feel free to experiment to find the best balance on your system.

• Improved filters response and added guard filter to prevent processing/creating signal above the Nyquist of the model's sample rate when the host sample rate is greater than the model.

• Improved CPU usage of filters, 'Minimum Phase' now uses a 0 latency polyphase IIR half-band decimator.
  The new filters processing is parallelized too when 'Multi-Core' is ON.

• Improved mono/stereo switch handling: this avoids unnecessary loading/reloading of the model(s) when going from mono to stereo or viceversa.

• Optmized CPU usage for stereo processing.

• Improved slimmable model handling by routing size changes through the ResamplingNAM wrapper.

• Several other small optimizations to CPU usage.

• Fixed AU view class collision with NeuralAmpModeler by @honkkis in #8

• Linux build momentarily deprecated (sorry but can't test it personally and cannot actively support it at the moment)

New Contributors

• @honkkis made their first contribution in #8

Full Changelog: 1.4.1...v1.5.0

NAM Oversampler v1.5.0-rc5

Changes since 1.5.0-rc4

• Possible fix for crackling issue on Apple Silicon Macs when multi-core is enabled.

Full Changelog: 1.5.0-rc4...1.5.0-rc5

NAM Oversampler v1.5.0-rc4

Changes since 1.5.0-rc3

• Improved multicore oversampling efficiency.

• Combined stereo processing with a single sync per block.

• Optimized scheduling on Apple Silicon, with performance core decoration and Workgroup Audio support.

• Lighter worker dispatcher with zero CPU consumption when idle.

• More efficient cloning of WaveNet and Slimmable models.

• Reduced buffer copies by directly processing compatible phases.

• Optimized minimum phase and linear phase anti-aliasing filters.

• Preallocated critical buffers to reduce allocations and instability during switching.

• Corrected declared latencies for all filters up to 32x.

Full Changelog: 1.5.0-rc3...1.5.0-rc4

NAM Oversampler v1.5.0-rc3

Changes since 1.5.0-rc2

• Improved the filters' response, now they don't over-cut above 20 kHz and still show minimal aliasing residual (Linear Phase short might show a bit more residual).

• Improved filtering for mismatched sample rates between host and model.

• Additional CPU usage improvements, especially with Multi-Core ON.

• Optimized CPU usage for stereo processing

• The plugin should now correctly prioritize the use of P-cores on Apple Silicon CPUs with Multi-Core ON.

• Reduced model loading time with Multi-Core ON.

Full Changelog: 1.5.0-rc2...1.5.0-rc3

NAM Oversampler v1.5.0-rc2

Changes since 1.5.0-rc1

• Improved multicore worker scheduling behavior.

• Added macOS worker QoS hints for Apple Silicon.

• Added Windows worker anti-EcoQoS handling for hybrid Intel CPUs.

• Tuned the phase-multicore worker pool so it uses the effective requested thread count instead of waking a maximum-size hardware pool.

• Reduced the chance that the audio thread waits on low-priority or efficiency-core scheduled worker threads.

• Phase-multicore workers are now explicitly marked as realtime-relevant work.

On macOS, worker threads now use:

QOS_CLASS_USER_INTERACTIVE

This helps reduce the chance that Apple Silicon schedules phase workers as lower-priority background work while the audio thread is waiting for them.

On Windows, worker threads continue to use MMCSS / Pro Audio scheduling, and now also explicitly opt out of execution-speed throttling where supported. This is intended to improve behavior on modern Intel hybrid CPUs with P-cores and E-cores.

Full Changelog: 1.5.0-rc1...1.5.0-rc2

NAM Oversampler v1.5.0-rc1

Release notes

• Added multi-core support (EXPERIMENTAL - WORK IN PROGRESS): this splits by phase the model processing and parallelize it across the available number of threads.
  You can enable it via the ON/OFF setting in the 'Oversampling' page and only works when oversampling is enabled as well.
  Number of threads can be set to 'Auto' or manually to a specific number.
  'Auto' sets it to "available number of logical cores -4" on Windows and Intel Macs, this showed the best performances in my tests but can be different for your specific CPU.
  On Apple Silicon the 'Auto' setting is more conservative to avoid overloading all P-cores.
  Feel free to experiment to find the best balance on your system.

• Added integration support for the optimized A2 Fast model backend.
  For supported A2 models, the plugin can use a frame-parallel backend instead of launching many tiny phase-model jobs, reducing multicore overhead and improving scalability at high oversampling factors.

• Improved cpu usage of filters, 'Minimum Phase' now uses a 0 latency polyphase IIR half-band decimator.
  The new filters processing is parallelized too when 'Multi-Core' is ON.

• Improved mono/stereo switch handling: this avoids unnecessary loading/reloading of the model(s) when going from mono to stereo or viceversa.

• Improved slimmable model handling by routing size changes through the ResamplingNAM wrapper.

• Several other optimizations to CPU usage on Windows

• Fixed AU view class collision with NeuralAmpModeler by @honkkis in #8

• Linux build momentarily deprecated (sorry but can't test it personally and cannot actively support it at the moment)

New Contributors

• @honkkis made their first contribution in #8

Full Changelog: 1.4.1...v1.5.0

NAM Oversampler v1.4.1

Release notes

• Turn on LTCG for mac release builds by @jlwitthuhn in #6
  This should reduce cpu usage a bit on MacOS

New Contributors

• @jlwitthuhn made their first contribution in #6

Full Changelog: v1.4.0...1.4.1

NAM Oversampler v1.4.0

Release notes

• Improved filters, now there are 3 options available:

1. Minimum Phase: cascaded half-band FIR with 0 latency (default for realtime)
2. Linear Phase (short): cascaded linear phase half-band FIR with 127 taps per stage
3. Linear Phase (short): cascaded linear phase half-band FIR with 511 taps per stage (default for offline rendering)
   Number of stages depends on the oversampling: 1 stage at 2x, 2 stages at 4x, 3 at 8x, 4 at 16x and 5 at 32x

• Fixed offline rendering alignment issues.

• Fixed filter-mode/default handling between realtime playback and offline rendering.

• Improved host reset/order handling for oversampling and latency reporting.

• Reduced risk of incorrect latency compensation when switching render modes.

Full Changelog: v1.3.1...v1.4.0

NAM Oversampler v1.3.1

Release notes

• Fixed occasional corruption of stereo processing after loading a new .nam file
• A few minor graphical tweaks

Full Changelog: v1.3.0...v1.3.1

NAM Oversampler v1.3.0

Release notes

• Added separate filter choice for offline rendering.
• Added additional filter types: Min Phase FIR and Polyphase FIR.
  The former is slightly heavier on the cpu than IIR but creates less phase rotations in the audible portion of the spectrum, the latter is a more efficient version of the linear phase FIR.
• Improved filter calculation logic: now it takes into account also the NAM model's sample rate.
  Example: if the project sample rate is 96 kHz but the model sample rate is 48 kHz, the filter doesn't let pass anything above 24 kHz as there's no realistic info contained in the model for frequencies above 24 kHz.
  If instead the model sample rate is 96 kHz too, the filter will let signal extend up to 48 kHz, as in this case the model was trained on those frequencies as well and they should be represented when the project sample rate allows it.
  NOTE: this change doesn't affect aliasing reduction
• Added current oversampling indicator on the main page, next to the logo.
• Fixed stereo mode not being engaged in some DAWs/hosts.
• Fixed a bug that might cause a wrong convolution time-scaling, thus a shifting of the model's frequency response, when the project sample rate is a non-integer multiple of the model sample rate (e.g. 48 kHz model in a 44.1 kHz project)
• DAW/host presets now correctly recall all oversampling and filter parameters
• Missing output on Mac/AU on 96 kHz projects should be fixed (cannot test this personally)
• Plugin not loading in Reaper on Mac should now be fixed (cannot test this personally)
• Added VST3 version for Linux (experimental - no GUI - cannot test this personally)

Known issues

• Sometimes, after loading a new model, the stereo processing may be corrupt (dry signal and/or much louder output on one channel).
  Usually just relaoding the same .nam file fixes the issue.

Full Changelog: v1.2.0...v1.3.0