Per-Layer Breakdown#

Goal: capture every available PMU counter (CPU, memory, MVE) for each layer in a model, not just the default basic_cpu set.

Setup#

hpx_full.yml

model:
  path: my_model.tflite
  arena_size: 131072

engine:
  type: helia-rt
  config:
    variant: release-with-logs
    dist_path: path/to/helia_rt_v1_7_0

target:
  board: apollo510_evb

profiling:
  pmu_counters:
    cpu: all
    memory: all
    mve: all
  per_layer: true
  iterations: 5
  warmup: 2

output:
  format: csv
  dir: ./results/full_sweep
  model_explorer: true
  detailed: true             # get per-preset CSVs and memory.json

Run#

hpx profile --config hpx_full.yml

Requesting all for three groups runs multiple PMU passes to cover every counter — see Multi-pass profiling for how passes are scheduled and merged.

What you get#

cat results/full_sweep/summary.json | python -m json.tool

{
  "layers": 13,
  "total_cycles": 2016376,
  "cache": {
    "l1d_hit_rate_pct": 91.2
  },
  "top_layers": [
    {"op": "CONV_2D", "cycles": 338176, "pct": 16.8}
  ]
}

profile_results.csv has one row per layer with every requested counter as a column. With --detailed, detailed/memory.json adds per-layer cache counters and arena allocation.

Where to go deeper#

PMU Counters — counter groups, aggregation, derived metrics (L1D hit rate, MVE instruction share), and how to read profile_results.csv with pandas.
Model Explorer Overlays — visualize per-layer hot spots on the model graph.
Basic Profiling — the minimal-counter starting point.