TFLM Profiling

The example demonstrates several performance profiling facilities offered by neuralSPOT: * Basic Per-Layer Performance Profiling * Full PMU (performance management unit) Per-layer profiling (CortexM55-based platforms) * GPIO signaling for off-board for use by devices such as a Joulescope * Ambiq register dump to SWO

The example uses one of Ambiq's larger models (the arrhythmia detector from Heartkit) as the model-under-test

NOTE: This example works for TFLM and HeliosRT

---

Hardware Requirements

Any Apollo EVB (AP3, AP4, and AP5), though the profiling will be different for each.

Building and Running the Example

IMPORTANT: Note the make CLI flags (MLDEBUG and MLPROFILE) needed to build this example, which are not enabled by default for the rest of neuralSPOT. These flags do two things: * Link in the right TFLM/HeliosRT with an enabled microProfiler * Compile in statistics data structures (these can be fairly large, so they're not compiled in by default)

$> make clean # gets rid of any non-MLDEBUG/MLPROFILE object files which would otherwise lead to linking errors
$> make EXAMPLE=examples/tflm_profiling MLDEBUG=1 MLPROFILE=1 deploy
$> make view

Results

The SWO output can be divided into 3 sections: 1. Register printouts 2. Basic per-layer counters (single run showing 4 PMU (AP5), or DWT (AP3/AP4) per layer) 3. (AP5 only) Full PMU counters per layer