Per-layer cost on the model graph

Goal

Look at a hot layer in context. You will learn which files to load, how the colours map to values, and how to switch between counters to tell a compute-bound layer from a memory-bound one.

Hardware and wiring

Any supported board for the profile run; none for the viewing step, which is a browser on the host.

Software prerequisites

hpx, a completed profile with output.model_explorer left at its default of true, and Model Explorer (pip install ai-edge-model-explorer, then model-explorer), or its hosted build.

Config file

Any profile config; examples/quickstart/hpx_full_sweep.yml sets output.model_explorer: true explicitly and captures the counters that make the overlays interesting. The quickstart hpx_rt.yml writes overlays for its default counters as well.

Run

cd examples/quickstart
hpx profile --config hpx_full_sweep.yml
ls results/results_full_sweep/model_explorer/
model-explorer kws_model.tflite

In Model Explorer, open the model, choose Add overlay and pick results/results_full_sweep/model_explorer/me_overlay_ARM_PMU_CPU_CYCLES.json.

Expected output

Not yet captured: the shipped bundles were produced with model_explorer: false. A run with overlays enabled writes one file per captured counter:

model_explorer/me_overlay_ARM_PMU_CPU_CYCLES.json
model_explorer/me_overlay_ARM_PMU_INST_RETIRED.json
model_explorer/me_overlay_ARM_PMU_STALL_FRONTEND.json
model_explorer/me_overlay_ARM_PMU_STALL_BACKEND.json

Each is Model Explorer’s node-data format keyed by graph id:

{"main": {"name": "ARM_PMU_CPU_CYCLES", "results": {"0": {"value": 337348.0}, "1": {"value": 216205.0}}, "gradient": [{"stop": 0, "bgColor": "#22c55e"}, {"stop": 0.5, "bgColor": "#eab308"}, {"stop": 1, "bgColor": "#ef4444"}]}}

How to read it

• Green is the minimum value in that overlay, yellow the midpoint, red the maximum. The scale is relative to the one file, so a colour says nothing about absolute cost and nothing across two overlays or two runs.
• Switch overlays on the same graph: a layer red on cycles and red on ARM_PMU_L1D_CACHE_MISS_RD warrants investigating memory placement. Low cache misses alone do not prove a compute bottleneck; inspect stalls and instruction counts too.
• Node matching: heliaRT and TFLM overlays match by the runtime’s sequential operator index; heliaAOT overlays use the model’s original operator index, which the compiler preserves even when it skips operators (Model Explorer overlays).
• jq '.main.results' me_overlay_ARM_PMU_CPU_CYCLES.json reads the values without the viewer; the same numbers with more context are in profile_results.csv.

Variations

• --no-model-explorer or output.model_explorer: false skips the export.
• Compare two runs visually by opening each overlay in its own tab; for a numeric comparison use hpx compare (Compare engines).

Validated on

Not yet validated. No shipped bundle contains a model_explorer/ directory; the file layout and JSON shape above are quoted from report/model_explorer.py at the source tree this site was built from.