kws_infer — Keyword Spotting Inference Example¶

End-to-end TFLite Micro inference demo on Apollo510, demonstrating:

nsx_system modular init (one-call startup with preset config)
nsx_mem memory placement (tensor arena in TCM, model weights in TCM)
NsxPmuProfiler per-layer PMU profiling via TFLM's MicroProfilerInterface
ITM/SWO printf output at 1 MHz via JLink

Model¶

Google Speech Commands KWS (12-class): silence, unknown, yes, no, up, down, left, right, on, off, stop, go.

Input: [1, 49, 10, 1] int8
Output: [1, 12] int8
Ops: Conv2D, DepthwiseConv2D, AveragePool2D, FullyConnected, Reshape, Softmax
Arena: ~20 KB of 64 KB allocated
Inference: ~21 ms @ 192 MHz HP mode with I/D cache

Build¶

cd neuralspotx/examples/kws_infer
nsx lock      --app-dir .
nsx configure --app-dir .
nsx build     --app-dir .

Flash & View Output¶

nsx flash     --app-dir .
nsx view      --app-dir .

System Init¶

Uses nsx_system_init() with a modified development preset:

nsx_system_config_t cfg = nsx_system_development;
cfg.skip_bsp_init = true;   // skip BSP's 2-second delay
nsx_system_init(&cfg);       // → core, CPDLP, cache, HP mode, SpotMgr, ITM

This replaces ~60 lines of manual DCU/TPIU/ITM/cache/SpotManager init.

Key Files¶

File	Purpose
`src/main.cc`	Application entry — system init, TFLM setup, inference loop
`src/kws_model_data.h`	Model weights as C array (in TCM via default .data)
`src/nsx_pmu_profiler.h`	Per-layer PMU profiler (TFLM MicroProfilerInterface)
`src/nsx_pmu_profiler.cc`	PMU profiler implementation — cycles, dcache misses, icache misses
`CMakeLists.txt`	Build config — links `nsx::helia_rt` source module + NSX modules

Dependencies¶

ARM GCC 14.3 (arm-none-eabi-gcc)
helia-rt helia-rt-v1.16.0 via NSX source module (release-with-logs variant by default)
ns-cmsis-nn v7.26.0 via helia-rt's NSX dependency metadata
NSX modules: nsx-core, nsx-power, nsx-pmu-armv8m (with the PMU API from the standalone module and the local SDK shim supplied by nsx-ambiq-sdk)