Bump version -> 0.2.0

Refactor build and add Raspberry Pi 4/5 builds

.github/workflows/build.yml

@@ -20,7 +20,7 @@ jobs:
       env:
         CXX: clang++
       run: |
-        cmake .. -DCMAKE_BUILD_TYPE=$BUILD_TYPE -DBUILD_UTILS=ON -DBUILD_NAMCORE=ON -DBUILD_STATIC_RTNEURAL=ON 
+        cmake .. -DCMAKE_BUILD_TYPE=$BUILD_TYPE -DBUILD_UTILS=ON -DBUILD_INTERNAL_STATIC_WAVENET=ON -DBUILD_INTERNAL_STATIC_LSTM=ON -DBUILD_NAMCORE=ON -DBUILD_STATIC_RTNEURAL=ON 
         cmake --build . --config $BUILD_TYPE -j4
         cpack
 
@@ -51,7 +51,7 @@ jobs:
     - name: Build Plugin
       working-directory: ${{github.workspace}}/build
       run: |
-        cmake.exe -G "Visual Studio 17 2022" -A x64 -DBUILD_UTILS=ON -DBUILD_NAMCORE=ON -DBUILD_STATIC_RTNEURAL=ON -T ClangCL ..
+        cmake.exe -G "Visual Studio 17 2022" -A x64 -DBUILD_UTILS=ON -DBUILD_INTERNAL_STATIC_WAVENET=ON -DBUILD_INTERNAL_STATIC_LSTM=ON -DBUILD_NAMCORE=ON -DBUILD_STATIC_RTNEURAL=ON -T ClangCL ..
         cmake --build . --config=release -j4
 
     - name: Run ModelTest

.github/workflows/release.yml

@@ -31,7 +31,12 @@ jobs:
   build-linux-x64:
     name: Build Linux x64
     needs: create_release
+    strategy:
+      matrix:
+        native_arch: [ON, OFF]
     runs-on: ubuntu-latest
+    env:
+      ARCHIVE_NAME: neural_amp_modeler_lv2_linux_x64${{ matrix.native_arch == 'ON' && 'v3' || '' }}.tgz
     steps:
     - uses: actions/checkout@v3.3.0
       with:
@@ -42,12 +47,12 @@ jobs:
       env:
         CXX: clang++
       run: |
-        cmake .. -DCMAKE_BUILD_TYPE=$BUILD_TYPE -DLSTM_PREFER_NAM=OFF -DWAVENET_PREFER_NAM=OFF
+        cmake .. -DCMAKE_BUILD_TYPE=$BUILD_TYPE -DUSE_NATIVE_ARCH=${{ matrix.native_arch }} -DBUILD_NAMCORE=ON -DBUILD_INTERNAL_STATIC_WAVENET=ON -DBUILD_INTERNAL_STATIC_LSTM=ON
         cmake --build . --config $BUILD_TYPE -j4
 
     - name: Add LV2 Archive
       working-directory: ${{github.workspace}}/build
-      run: tar -czf neural_amp_modeler_lv2_linux_x64.tgz neural_amp_modeler.lv2
+      run: tar -czf ${{ env.ARCHIVE_NAME }} neural_amp_modeler.lv2
       
     - name: Upload Plugin Asset
       uses: actions/upload-release-asset@v1
@@ -55,30 +60,35 @@ jobs:
         GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
       with:
         upload_url: ${{ needs.create_release.outputs.upload_url }}
-        asset_path: ./build/neural_amp_modeler_lv2_linux_x64.tgz
-        asset_name: neural_amp_modeler_lv2_linux_x64.tgz
+        asset_path: ./build/${{ env.ARCHIVE_NAME }}
+        asset_name: ${{ env.ARCHIVE_NAME }}
         asset_content_type: application/tgz
 
-  build-linux-x64v3:
-    name: Build Linux x64v3
+  build-rpi:
+    name: Build Raspberry Pi
     needs: create_release
+    strategy:
+      matrix:
+        native_arch: [rpi4, rpi5]
     runs-on: ubuntu-latest
+    container: 
+      image: ghcr.io/tttapa/docker-arm-cross-toolchain:aarch64-rpi3-linux-gnu-gcc12
+    env:
+      ARCHIVE_NAME: neural_amp_modeler_lv2_${{ matrix.native_arch }}.tgz
     steps:
     - uses: actions/checkout@v3.3.0
       with:
         submodules: recursive
-        
+    
     - name: Build Plugin
       working-directory: ${{github.workspace}}/build
-      env:
-        CXX: clang++
       run: |
-        cmake .. -DCMAKE_BUILD_TYPE=$BUILD_TYPE -DUSE_NATIVE_ARCH=ON -DLSTM_PREFER_NAM=OFF -DWAVENET_PREFER_NAM=OFF
+        cmake .. -DCMAKE_BUILD_TYPE=$BUILD_TYPE -DBUILD_NAMCORE=ON -DBUILD_INTERNAL_STATIC_WAVENET=ON -DBUILD_INTERNAL_STATIC_LSTM=ON -DCMAKE_TOOLCHAIN_FILE=/home/develop/opt/x-tools/aarch64-rpi3-linux-gnu/aarch64-${{ matrix.native_arch }}-linux-gnu.toolchain.cmake
         cmake --build . --config $BUILD_TYPE -j4
 
     - name: Add LV2 Archive
       working-directory: ${{github.workspace}}/build
-      run: tar -czf neural_amp_modeler_lv2_linux_x64v3.tgz neural_amp_modeler.lv2
+      run: tar -czf ${{ env.ARCHIVE_NAME }} neural_amp_modeler.lv2
       
     - name: Upload Plugin Asset
       uses: actions/upload-release-asset@v1
@@ -86,14 +96,19 @@ jobs:
         GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
       with:
         upload_url: ${{ needs.create_release.outputs.upload_url }}
-        asset_path: ./build/neural_amp_modeler_lv2_linux_x64v3.tgz
-        asset_name: neural_amp_modeler_lv2_linux_x64v3.tgz
+        asset_path: ./build/${{ env.ARCHIVE_NAME }}
+        asset_name: ${{ env.ARCHIVE_NAME }}
         asset_content_type: application/tgz
         
   build-windows:
     name: Build Windows
     needs: create_release
+    strategy:
+      matrix:
+        native_arch: [ON, OFF]
     runs-on: windows-latest
+    env:
+      ARCHIVE_NAME: neural_amp_modeler_lv2_win_x64${{ matrix.native_arch == 'ON' && 'v3' || '' }}.zip
     steps:
     - uses: actions/checkout@v3.3.0
       with:
@@ -102,11 +117,11 @@ jobs:
     - name: Build Plugin
       working-directory: ${{github.workspace}}/build
       run: |
-        cmake.exe -G "Visual Studio 17 2022" -A x64  -T ClangCL -DLSTM_PREFER_NAM=OFF -DWAVENET_PREFER_NAM=OFF ..
+        cmake.exe -G "Visual Studio 17 2022" -A x64  -T ClangCL -DUSE_NATIVE_ARCH=${{ matrix.native_arch }} -DBUILD_NAMCORE=ON -DBUILD_INTERNAL_STATIC_WAVENET=ON -DBUILD_INTERNAL_STATIC_LSTM=ON ..
         cmake --build . --config=release -j4
 
     - name: Add LV2 Archive
-      run: Compress-Archive -Path ${{github.workspace}}\build\neural_amp_modeler.lv2 -Destination neural_amp_modeler_lv2_win_x64.zip
+      run: Compress-Archive -Path ${{github.workspace}}\build\neural_amp_modeler.lv2 -Destination ${{ env.ARCHIVE_NAME }}
       
     - name: Upload Plugin Asset
       uses: actions/upload-release-asset@v1
@@ -114,34 +129,6 @@ jobs:
         GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
       with:
         upload_url: ${{ needs.create_release.outputs.upload_url }}
-        asset_path: ./neural_amp_modeler_lv2_win_x64.zip
-        asset_name: neural_amp_modeler_lv2_win_x64.zip
-        asset_content_type: application/zip
-
-  build-windows-x64v3:
-    name: Build Windows x64v3
-    needs: create_release
-    runs-on: windows-latest
-    steps:
-    - uses: actions/checkout@v3.3.0
-      with:
-        submodules: recursive
-    
-    - name: Build Plugin
-      working-directory: ${{github.workspace}}/build
-      run: |
-        cmake.exe -G "Visual Studio 17 2022" -A x64 -DUSE_NATIVE_ARCH=ON -DLSTM_PREFER_NAM=OFF -DWAVENET_PREFER_NAM=OFF -T ClangCL ..
-        cmake --build . --config=release -j4
-
-    - name: Add LV2 Archive
-      run: Compress-Archive -Path ${{github.workspace}}\build\neural_amp_modeler.lv2 -Destination neural_amp_modeler_lv2_win_x64v3.zip
-      
-    - name: Upload Plugin Asset
-      uses: actions/upload-release-asset@v1
-      env:
-        GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-      with:
-        upload_url: ${{ needs.create_release.outputs.upload_url }}
-        asset_path: ./neural_amp_modeler_lv2_win_x64v3.zip
-        asset_name: neural_amp_modeler_lv2_win_x64v3.zip
+        asset_path: ./${{ env.ARCHIVE_NAME }}
+        asset_name: ${{ env.ARCHIVE_NAME }}
         asset_content_type: application/zip

CMakeLists.txt

@@ -1,6 +1,6 @@
 cmake_minimum_required(VERSION 3.10)
 
-project(NeuralAmpModelerLv2 VERSION 0.1.9)
+project(NeuralAmpModelerLv2 VERSION 0.2.0)
 
 set(CMAKE_MODULE_PATH "${CMAKE_SOURCE_DIR}/cmake")
 set(LIB_INSTALL_DIR ${CMAKE_INSTALL_PREFIX}/lib CACHE PATH "The library install dir (default: lib)")
@@ -12,7 +12,7 @@ set(CMAKE_CXX_EXTENSIONS OFF)
 if (CMAKE_SYSTEM_NAME STREQUAL "Darwin")
 	include_directories(SYSTEM /usr/local/include)
 elseif (CMAKE_SYSTEM_NAME STREQUAL "Linux")
-	link_libraries(stdc++fs)
+	link_libraries( "$<$<AND:$<CXX_COMPILER_ID:GNU>,$<VERSION_LESS:$<CXX_COMPILER_VERSION>,9.0>>:-lstdc++fs>" )
 elseif (CMAKE_SYSTEM_NAME STREQUAL "Windows")
 	add_compile_definitions(NOMINMAX WIN32_LEAN_AND_MEAN)
 else()

README.md

@@ -1,27 +1,28 @@
 # neural-amp-modeler-lv2
 
-LV2 plugin for using neural network machine learning amp models.
+LV2 plugin for neural network machine learning amp model playback using the [NeuralAudio](https://github.com/mikeoliphant/NeuralAudio) engine.
 
 **There is no custom plugin user interface**. Setting the model to use requires that your LV2 host supports atom:Path parameters. Reaper does as of v6.82. Carla and Ardour do. If your favorite LV2 host does not support atom:Path, let them know you want it.
-If you are looking for a GUI version, @brummer10 [has one here](https://github.com/brummer10/neural-amp-modeler-ui) that works for Linux and Windows. You may also be interested in the the version shipped with the [MOD Desktop App](https://github.com/moddevices/mod-desktop-app), or my digital pedalboard app [Stompbox](https://github.com/mikeoliphant/StompboxUI).
+If you are looking for a GUI version, @brummer10 [has one here](https://github.com/brummer10/neural-amp-modeler-ui) that works for Linux and Windows. You may also be interested in the the version shipped with the [MOD Desktop App](https://github.com/moddevices/mod-desktop-app), or my digital pedalboard app [Stompbox](https://github.com/mikeoliphant/Stompbox).
 
 To get the intended behavior, **you must run your audio host at the same sample rate the model was trained at** (usually 48kHz) - no resampling is done by the plugin.
 
 For amp-only models (the most typical), **you will need to run an impulse reponse after this plugin** to model the cabinet.
 
-## Models and Performance
+## Models Supported
 
 The plugin supports both [Neural Amp Modeler (NAM)](https://github.com/sdatkinson/neural-amp-modeler) models and [RTNeural keras json models](https://github.com/jatinchowdhury18/RTNeural) (like those used by [Aida-X](https://github.com/AidaDSP/AIDA-X)).
 
-The best source of models is [ToneHunt](https://tonehunt.org/).
+The best source of models is [Tone3000](https://www.tone3000.com/).
+
+For more information on model type support, see the [NeuralAudio](https://github.com/mikeoliphant/NeuralAudio) repository, which is where the model handling code lives.
+
+## Performance
 
 NAM WaveNet models are generally quite expensive to run. This isn't (much of) an issue on modern PCs, but you may have trouble running on less powerful hardware.
 
-A Raspberry Pi 4 running a 64bit OS can run "standard" NAM models with plenty of room to spare for a cabinet IR and some effects. It is also capable of running two "standard" NAM models, but with less headroom for other effects.
+If you are having trouble running a "standard" model, try looking for "feather", or even "nano" (the least expensive) models. You can find a list of ["feather"-tagged models on Tone3000](https://www.tone3000.com/search?sizes=feather). Note that tagging models is up to the submitter, so not all "feather" models are tagged as such - you should be able to find more if you dig around.
 
-If you are having trouble running a "standard" model, try looking for "feather", or even "nano" (the least expensive) models. You can find a list of ["feather"-tagged models on ToneHunt](https://tonehunt.org/models?tags%5B0%5D=feather-mdl). Note that tagging models is up to the submitter, so not all "feather" models are tagged as such - you should be able to find more if you dig around.
-
-For more information on model type support, see the [NeuralAudio](https://github.com/mikeoliphant/NeuralAudio) repository, which is where the model handling code lives.
 
 ## Input Calibration
 
@@ -57,4 +58,6 @@ After building, the plugin will be in **build/neural_amp_modeler.lv2**.
 
 ```-DUSE_NATIVE_ARCH=ON```: If you have a relatively modern x64 processor, you can pass ```-DUSE_NATIVE_ARCH=ON``` on your cmake command line to enable certain processor-specific optimizations.
 
+```-DSMART_BYPASS_ENABLED=ON```: If enabled, this will bypass model processing if input has been silent (below -100 dB by default) for a sufficient number of samples (determined by the model's receptive field size).
+
 Also see the [NeuralAudio CMake options](https://github.com/mikeoliphant/NeuralAudio#cmake-options) - adding these to your neural-amp-modeler-lv2 cmake will pass them to the NeuralAudio build.

models/README.md

@@ -1,4 +1,4 @@
-These are some sample NAM models designed to be used for performance testing on CPU-limited devices. They are all based on a [LiveSpice model of a Boss SD-1 pedal](https://blog.nostatic.org/2023/04/this-boss-sd-1-pedal-does-not-exist.html).
+These are some sample NAM models designed to be used for performance testing on CPU-limited devices. They are all based on a [LiveSpice model of a Boss SD-1 pedal](https://blog.nostatic.org/2023/04/this-boss-sd-1-pedal-does-not-exist.html). They are provided here under the [CC BY-NC-ND 4.0 license](https://creativecommons.org/licenses/by-nc-nd/4.0/deed.en).
 
 All models were trained for 300 epochs using the NAM "v1_1_1.wav" capture signal.
 

resources/neural_amp_modeler.ttl.in

@@ -101,4 +101,12 @@ A large collection of models is available at https://tonehunt.org
 		lv2:minimum -20.0;
 		lv2:maximum 20.0;
 		units:unit units:db;
+	], [
+		a lv2:ControlPort, lv2:InputPort;
+		lv2:index 6;
+		lv2:symbol "quality_scale";
+		lv2:name "Quality";
+		lv2:default 1.0;
+		lv2:minimum 0.0;
+		lv2:maximum 1.0;
 	].

src/CMakeLists.txt

@@ -54,6 +54,12 @@ if (DISABLE_DENORMALS)
 	add_definitions(-DDISABLE_DENORMALS)
 endif (DISABLE_DENORMALS)
 
+option(SMART_BYPASS_ENABLED "Enable auto-bypass on silence" OFF)
+
+if (SMART_BYPASS_ENABLED)
+	add_definitions(-DSMART_BYPASS_ENABLED)
+endif (SMART_BYPASS_ENABLED)
+
 set_target_properties(neural_amp_modeler
 	PROPERTIES
 	CXX_VISIBILITY_PRESET hidden

src/nam_plugin.cpp

@@ -7,12 +7,18 @@
 
 #define SMOOTH_EPSILON .0001f
 
+#ifndef BYPASS_DB_THRESHOLD
+#define BYPASS_DB_THRESHOLD -100
+#endif
+
 namespace NAM {
 	Plugin::Plugin()
 	{
 		// prevent allocations on the audio thread
 		currentModelPath.reserve(MAX_FILE_NAME + 1);
 
+		bypassThresholdLinear = powf(10, BYPASS_DB_THRESHOLD * 0.05f);
+
 //		NeuralAudio::NeuralModel::SetLSTMLoadMode(
 //#ifdef LSTM_PREFER_NAM
 //			NeuralAudio::PreferNAMCore
@@ -122,7 +128,7 @@ namespace NAM {
 					{
 						lv2_log_trace(&nam->logger, "Staging model change: `%s`\n", msg->path);
 
-						model = NeuralAudio::NeuralModel::CreateFromFile(msg->path);
+						model = nam->loader.CreateFromFile(msg->path);
 					}
 
 					if (model != nullptr)
@@ -181,6 +187,18 @@ namespace NAM {
 		nam->currentModelPath = msg->path;
 		assert(nam->currentModelPath.capacity() >= MAX_FILE_NAME + 1);
 
+		if (nam->currentModel != nullptr)
+		{
+			int receptiveFieldSize = nam->currentModel->GetReceptiveFieldSize();
+
+			if (receptiveFieldSize > -1)
+			{
+				// A newly loaded model is prewarmed to have a silent sample history
+				nam->silentSamples = receptiveFieldSize;
+				nam->smartBypassed = true;
+			}
+		}
+
 		// send reply
 		nam->schedule->schedule_work(nam->schedule->handle, sizeof(reply), &reply);
 
@@ -194,7 +212,7 @@ namespace NAM {
 	{
 		maxBufferSize = size;
 
-		NeuralAudio::NeuralModel::SetDefaultMaxAudioBufferSize(size);
+		loader.SetDefaultMaxAudioBufferSize(size);
 	}
 
 	void Plugin::process(uint32_t n_samples) noexcept
@@ -234,6 +252,18 @@ namespace NAM {
 			}
 		}
 
+		if (*(ports.quality_scale) != qualityScale)
+		{
+			qualityScale = *(ports.quality_scale);
+
+			loader.SetDefaultQualityScaleFactor(qualityScale);
+
+			if (currentModel != nullptr)
+			{
+				currentModel->SetQualityScaleFactor(qualityScale);
+			}
+		}
+
 		float level;
 
 		float modelInputAdjustmentDB = 0;
@@ -241,6 +271,44 @@ namespace NAM {
 		if (currentModel != nullptr)
 		{
 			modelInputAdjustmentDB = currentModel->GetRecommendedInputDBAdjustment();
+
+#ifdef SMART_BYPASS_ENABLED
+			int receptiveFieldSamples = currentModel->GetReceptiveFieldSize();
+
+			if (receptiveFieldSamples > -1)
+			{
+				for (unsigned int i = 0; i < n_samples; i++)
+				{
+					if (abs(ports.audio_in[i]) <= bypassThresholdLinear)
+					{
+						silentSamples++;
+					}
+					else
+					{
+						silentSamples = 0;
+					}
+				}
+
+				if (silentSamples >= (uint32_t)receptiveFieldSamples)
+				{
+					silentSamples = (uint32_t)receptiveFieldSamples;	// Prevent silentSamples growing and eventually overflowing uint32
+
+					if (smartBypassed)
+					{
+						for (unsigned int i = 0; i < n_samples; i++)
+						{
+							ports.audio_out[i] = ports.audio_in[i];
+						}
+
+						return;
+					}
+
+					smartBypassed = true; // If we aren't already, we'll be bypassed on the next process call
+				}
+				else
+					smartBypassed = false;
+			}
+#endif
 		}
 
 		// convert input level from db

src/nam_plugin.h

@@ -60,6 +60,7 @@ namespace NAM {
 			float* audio_out;
 			float* input_level;
 			float* output_level;
+			float* quality_scale;
 		};
 
 		Ports ports = {};
@@ -70,16 +71,19 @@ namespace NAM {
 		LV2_Log_Logger logger = {};
 		LV2_Worker_Schedule* schedule = nullptr;
 
+		NeuralAudio::NeuralModelLoader loader;
 		NeuralAudio::NeuralModel* currentModel = nullptr;
 		std::string currentModelPath;
 		float prevDCInput = 0;
 		float prevDCOutput = 0;
+		float qualityScale = 1.0f;
 
 		Plugin();
 		~Plugin();
 
 		bool initialize(double rate, const LV2_Feature* const* features) noexcept;
 		void set_max_buffer_size(int size) noexcept;
+		void activate() noexcept;
 		void process(uint32_t n_samples) noexcept;
 
 		void write_current_path();
@@ -120,5 +124,8 @@ namespace NAM {
 		float inputLevel = 0;
 		float outputLevel = 0;
 		int32_t maxBufferSize = 512;
+		float bypassThresholdLinear = 0;
+		uint32_t silentSamples = 0;
+		bool smartBypassed = true;
 	};
 }