User guide
The User guide is organised by what you are doing. Start with a task below, then follow its prerequisites and examples. If you have not run a profile yet, Getting started is the shorter path.
The blocks of a run configuration
Section titled “The blocks of a run configuration”Use these configuration blocks to find the guide for your next change. The generated configuration reference lists all keys and defaults.
| Block | Choose or inspect |
|---|---|
model |
Model format and arena size, memory placement |
engine |
Runtime or ahead-of-time compiler |
target |
Board and clock, toolchain, transport |
profiling |
Counters and sampling |
power |
Instrument, wiring and measurement window |
output |
Result files and model overlays |
timeouts, build |
Stage limits and dependency resolution |
Set up a run
Section titled “Set up a run”heliaRT, heliaAOT, TFLM and ExecuTorch: supported model formats, setup and configuration.
The model formats each engine accepts, the tensor arena, and how to size it from a measurement.
The board registry, what each SoC family can capture, the CPU clock and custom boards.
GCC, Arm Compiler 6 and the Arm Toolchain for Embedded: install, select, compare.
RTT, USB CDC, UART and SWO: what each needs plugged in and what it can lose.
Where the arena and the weights live, how automatic placement decides, and how to tune it.
Compare the ns and arm providers on a Cortex-M55 using the
documented operator measurements.
How the file and the command line merge, and the behaviour a schema table cannot express.
Measure
Section titled “Measure”Counter groups per SoC, multi-pass capture and overflow detection.
Joulescope wiring, INA228, measurement windows and on-device summaries.
Read results
Section titled “Read results”Every file a run writes, its schema and its units.
Read clean inference timing and instrumented layer costs with the right units and scope.
Re-analyse a bundle, compare two runs, and the rules that make a comparison fair.
Per-layer cost on the model graph.
Concepts
Section titled “Concepts”The pipeline, engine adapters, firmware generation, capture, field diagnostics and how to add an engine, each with a block diagram.