WIP: tflm test module for WOV

app/boards/intel_adsp_ace15_mtpm.conf

@@ -17,6 +17,10 @@ CONFIG_FORMAT_CONVERT_HIFI3=n
 CONFIG_SAMPLE_KEYPHRASE=y
 CONFIG_COMP_PHASE_VOCODER=y
 CONFIG_COMP_STFT_PROCESS=y
+CONFIG_SOF_STAGING=y
+CONFIG_CPP=y
+CONFIG_STD_CPP17=y
+CONFIG_COMP_TENSORFLOW=m
 
 # SOF / audio modules / mocks
 # This mock is part of official sof-bin releases because the CI that

app/boards/intel_adsp_ace30_ptl.conf

@@ -15,6 +15,12 @@ CONFIG_FORMAT_CONVERT_HIFI3=n
 CONFIG_COMP_GOOGLE_RTC_AUDIO_PROCESSING=m
 CONFIG_GOOGLE_RTC_AUDIO_PROCESSING_MOCK=y
 CONFIG_COMP_STFT_PROCESS=y
+CONFIG_SOF_STAGING=y
+CONFIG_COMP_KPB=y
+CONFIG_COMP_TENSORFLOW=y
+CONFIG_COMP_MFCC=y
+CONFIG_COMP_VOLUME=y
+CONFIG_COMP_GAIN=y
 
 # SOF / infrastructure
 CONFIG_KCPS_DYNAMIC_CLOCK_CONTROL=n
@@ -29,7 +35,7 @@ CONFIG_COLD_STORE_EXECUTE_DRAM=y
 CONFIG_INTEL_MODULES=y
 CONFIG_LIBRARY_AUTH_SUPPORT=y
 CONFIG_LIBRARY_MANAGER=y
-CONFIG_LIBRARY_BASE_ADDRESS=0xa0688000
+CONFIG_LIBRARY_BASE_ADDRESS=0xa0700000
 CONFIG_LIBRARY_BUILD_LIB=y
 CONFIG_LIBRARY_DEFAULT_MODULAR=y
 

app/llext_relocatable.conf

@@ -1 +1,2 @@
 CONFIG_LLEXT_TYPE_ELF_RELOCATABLE=y
+CONFIG_LLEXT_EXPORT_BUILTINS_BY_SLID=y

scripts/llext_link_helper.py

@@ -77,6 +77,8 @@ def main():
 
 	command = [args.command]
 
+	is_relocatable = '-r' in args.params
+
 	executable = []
 	writable = []
 	readonly = []
@@ -111,7 +113,8 @@ def main():
 				text_found = True
 				text_addr = max_alignment(text_addr, 0x1000, s_alignment)
 				text_size = s_size
-				command.append(f'-Wl,-Ttext=0x{text_addr:x}')
+				if not is_relocatable:
+					command.append(f'-Wl,-Ttext=0x{text_addr:x}')
 			else:
 				executable.append(section)
 
@@ -164,7 +167,8 @@ def main():
 
 		dram_addr = align_up(dram_addr, s_alignment)
 
-		command.append(f'-Wl,--section-start={s_name}=0x{dram_addr:x}')
+		if not is_relocatable:
+			command.append(f'-Wl,--section-start={s_name}=0x{dram_addr:x}')
 
 		dram_addr += section.header['sh_size']
 
@@ -177,7 +181,8 @@ def main():
 
 		dram_addr = align_up(dram_addr, s_alignment)
 
-		command.append(f'-Wl,--section-start={s_name}=0x{dram_addr:x}')
+		if not is_relocatable:
+			command.append(f'-Wl,--section-start={s_name}=0x{dram_addr:x}')
 
 		dram_addr += section.header['sh_size']
 
@@ -189,7 +194,8 @@ def main():
 
 		start_addr = align_up(start_addr, s_alignment)
 
-		command.append(f'-Wl,--section-start={s_name}=0x{start_addr:x}')
+		if not is_relocatable:
+			command.append(f'-Wl,--section-start={s_name}=0x{start_addr:x}')
 
 		start_addr += section.header['sh_size']
 
@@ -201,10 +207,11 @@ def main():
 
 		start_addr = align_up(start_addr, s_alignment)
 
-		if s_name == '.data':
-			command.append(f'-Wl,-Tdata=0x{start_addr:x}')
-		else:
-			command.append(f'-Wl,--section-start={s_name}=0x{start_addr:x}')
+		if not is_relocatable:
+			if s_name == '.data':
+				command.append(f'-Wl,-Tdata=0x{start_addr:x}')
+			else:
+				command.append(f'-Wl,--section-start={s_name}=0x{start_addr:x}')
 
 		start_addr += section.header['sh_size']
 

scripts/llext_offset_calc.py

@@ -47,6 +47,9 @@ def get_elf_size(elf_name):
 			if section.header['sh_addr'] + section.header['sh_size'] > end:
 				end = section.header['sh_addr'] + section.header['sh_size']
 
+		if start == 0xffffffff:
+			return 0
+
 		size = end - start
 
 	return size

src/audio/tensorflow/README.md

@@ -1,22 +1,154 @@
-# TensorFlow Lite Micro (TFLM) Architecture
+# TensorFlow Lite Micro (TFLM) & Wake-on-Voice (WoV) Architecture
 
-This directory acts as the bridge for running ML models.
+This directory provides the TensorFlow Lite for Microcontrollers (TFLM) classification module (`TFLMCLY`) for Sound Open Firmware (SOF), including integration with MFCC feature extraction, mtrace logging, IPC host notifications, and Key Phrase Buffer (KPB) Wake-on-Voice (WoV) trigger infrastructure.
+
+---
 
 ## Overview
 
-Integrates TensorFlow Lite for Microcontrollers into the SOF audio pipeline. Evaluates pre-trained neural network topologies inline with the audio stream for tasks like wake-word, noise cancellation, or sound classification.
+The TFLM module evaluates pre-trained micro speech neural network models inline within the SOF audio processing graph. It receives pre-processed audio feature tensors (e.g. 40-bin mel spectrograms from the MFCC component), runs model inference, logs keyword detections to `mtrace`, issues IPC4 notifications to the host audio driver, and signals the KPB module to drain buffered pre-keyword audio for Wake-on-Voice.
+
+---
+
+## Architecture & Data Flow
 
-## Architecture Diagram
+### Dual-Path Wake-on-Voice (WoV) Architecture
+
+To allow continuous keyword evaluation without streaming audio to the host until a keyword is detected, the pipeline separates real-time keyword detection from host PCM draining via KPB:
 
 ```mermaid
-graph LR
-  Feat[Audio Features] --> TFLM[TFLM Inference Engine]
-  SubGraph[FlatBuffer Model] -.-> TFLM
-  TFLM --> Out[Inference Labels/Scores]
+graph TD
+    DAI[HDA Mic DAI] --> Gain[Gain Component]
+    Gain --> KPB[KPB Buffer Module]
+
+    subgraph "Real-Time Detection Path (KPB Pin 1)"
+        KPB -- Live Audio Stream --> SRC[SRC: 48kHz -> 16kHz]
+        SRC --> MFCC[MFCC Feature Extractor]
+        MFCC -- 40-bin Mel Tensors --> TFLM[TFLM Classifier: tflmcly]
+        TFLM --> VSink[Virtual Sink: virtual.tflm_sink]
+    end
+
+    subgraph "Host Draining Path (KPB Pin 2)"
+        KPB -- History Draining Stream --> Host[Host Copier: PCM Capture]
+    end
+
+    TFLM -- "1. Log to mtrace (comp_info)" --> MTrace[mtrace / SOF Trace Log]
+    TFLM -- "2. IPC4 Host Notification" --> IPC[Host Audio Driver]
+    TFLM -- "3. KPB_EVENT_BEGIN_DRAINING (notifier_event)" --> KPB
+```
+
+---
+
+## Pipeline Execution & Event Flow
+
+1. **Continuous Real-Time Listening**:
+   - Live microphone audio is captured by the HDA DAI and passed into `KPB` (`kpb.2.1`).
+   - KPB Output Pin 1 continuously streams audio to `SRC` (resampling 48kHz $\to$ 16kHz), `MFCC` (generating 40-bin `int8_t` mel spectrogram features), and `TFLM` (`tflmcly.1.1`).
+   - KPB stores the raw PCM audio continuously in its circular history buffer (e.g. 2100ms – 3000ms history).
+
+2. **Inference & Keyword Detection**:
+   - `tflm_process()` feeds 1960-byte feature tensors ($40 \text{ features} \times 49 \text{ windows}$) into the Micro Speech TFLM interpreter (`[1, 49, 40]` `int8_t` input tensor).
+   - Upon `TF_ProcessClassify()`, predictions are evaluated across categories:
+     - `0`: `silence`
+     - `1`: `unknown`
+     - `2`: `yes` (Keyword)
+     - `3`: `no` (Keyword)
+
+3. **Wake-on-Voice Trigger & Notification**:
+   When a keyword (`yes` or `no`) is detected with $\ge 0.70$ confidence:
+   - **mtrace Logging**: Logs detection with keyword label and confidence score via `comp_info()`:
+     `"TFLM keyword detected: yes (confidence 0.852)"`
+   - **Host IPC Notification**: Sends an IPC4 module notification (`SOF_IPC4_MODULE_NOTIFICATION`) to inform the host driver.
+   - **KPB Draining Signal**: Fires a system notification event (`NOTIFIER_ID_KPB_CLIENT_EVT`) with `KPB_EVENT_BEGIN_DRAINING`.
+   - **Host PCM Draining**: KPB opens Output Pin 2 to `host-copier`, draining pre-keyword history buffer audio followed by live mic audio to the host capture stream.
+
+---
+
+## Topology v2 Integration & Usage
+
+### 1. Component Widget (`include/components/tflm.conf`)
+
+Defines `Class.Widget."tflmcly"`:
+- **UUID**: `42:c6:1d:c5:e1:a2:df:48:a4:90:e2:74:8c:b6:36:3e` (`c51dc642-a2e1-48df-a490e2748cb6363e`)
+- **Type**: `effect`
+
+### 2. Detection Pipeline Template (`include/pipelines/cavs/host-gateway-src-mfcc-tflm-capture.conf`)
+
+Instantiates the real-time detection graph:
+```conf
+Object.Widget {
+    virtual."1" { name "virtual.tflm_sink" }
+    src."1"     { ... }
+    mfcc."1"    { ... }
+    tflmcly."1" { ... }
+}
 ```
 
-## Configuration and Scripts
+### 3. Top-Level Topology Configuration (`sof-hda-tflm.conf`)
+
+Instantiates the complete HDA Mic WoV topology with dual-path KPB routing:
+```conf
+Object.Base.route [
+    # DAI -> Gain -> KPB
+    { source "dai-copier.HDA.Analog.capture"; sink "gain.2.1" }
+    { source "gain.2.1"; sink "kpb.2.1" }
+
+    # KPB Pin 1 -> Real-time Detection Path
+    { source "kpb.2.1"; sink "src.1.1" }
+
+    # KPB Pin 2 -> Host WoV Draining Path
+    { source "kpb.2.1"; sink "host-copier.0.capture" }
+]
+```
+
+---
+
+## Building and Testing Topologies
+
+### Pre-processing and Compiling with `alsatplg`
+
+Build `sof-hda-tflm.tplg` using the Topology v2 pre-processor:
+
+```bash
+ALSA_CONFIG_DIR=tools/topology/topology2 \
+  tools/bin/alsatplg \
+  -I tools/topology/topology2/ \
+  -p -c tools/topology/topology2/sof-hda-tflm.conf \
+  -o build/sof-hda-tflm.tplg
+```
+
+### Inspecting Decoded Topology Graphs
+
+Decode and verify the compiled `.tplg` binary:
+
+```bash
+tools/bin/alsatplg -d build/sof-hda-tflm.tplg -o decoded.txt
+grep -A 20 "SectionGraph" decoded.txt
+```
+
+Expected graph output:
+```text
+SectionGraph {
+    set0 {
+        gain.2.1 <- dai-copier.HDA.Analog.capture
+        kpb.2.1 <- gain.2.1
+        src.1.1 <- kpb.2.1                 (Real-Time Detection Path)
+        host-copier.0.capture <- kpb.2.1   (Host WoV Draining Path)
+    }
+    set1 {
+        mfcc.1.1 <- src.1.1
+        tflmcly.1.1 <- mfcc.1.1
+        virtual.tflm_sink <- tflmcly.1.1   (Detection Sink Termination)
+    }
+}
+```
+
+---
+
+## Source Files
 
-- **Kconfig**: Enforces requirements for C++17 support and core framework staging logic (`COMP_TENSORFLOW`).
-- **CMakeLists.txt**: An intricate build specification linking the Tensilica neural network library block computations (`nn_hifi_lib`) and the TensorFlow Lite micro core engine (`tflm_lib`). Also hooks the `tflm-classify.c` SOF adapter via compiler flags explicitly enforcing memory, precision, and XTENSA optimizations.
-- **tflmcly.toml**: Topology definition for the specific TFLM Classifier implementation binding the engine against the UUID `UUIDREG_STR_TFLMCLY`.
+- **[tflm-classify.c](file:///home/lrg/work/sof-ptl/sof/src/audio/tensorflow/tflm-classify.c)**: SOF module adapter implementation for TFLM.
+- **[speech.cc](file:///home/lrg/work/sof-ptl/sof/src/audio/tensorflow/speech.cc)** / **[speech.h](file:///home/lrg/work/sof-ptl/sof/src/audio/tensorflow/speech.h)**: TFLM C++ API bridge & micro speech tensor wrapper.
+- **[tflm.conf](file:///home/lrg/work/sof-ptl/sof/tools/topology/topology2/include/components/tflm.conf)**: Topology v2 widget class definition.
+- **[host-gateway-src-mfcc-tflm-capture.conf](file:///home/lrg/work/sof-ptl/sof/tools/topology/topology2/include/pipelines/cavs/host-gateway-src-mfcc-tflm-capture.conf)**: Detection pipeline template.
+- **[sof-hda-tflm.conf](file:///home/lrg/work/sof-ptl/sof/tools/topology/topology2/sof-hda-tflm.conf)**: Top-level WoV topology configuration.