Skip to content

Toolchain Comparison#

Goal: run the same model/engine/board through GCC, armclang, and ATfE to see how compiler choice affects cycle counts.

Setup#

No config file needed — pass --toolchain on the CLI and keep everything else constant. Confirm all three toolchains are installed first (see Toolchains) and run hpx doctor to check the baseline tool checks pass.

Run#

Six commands, six result directories — three toolchains × two engines:

MODEL=examples/quickstart/kws_model.tflite

hpx profile $MODEL --engine helia-rt --toolchain gcc      \
    --output-dir results_compare/rt_gcc

hpx profile $MODEL --engine helia-rt --toolchain armclang \
    --output-dir results_compare/rt_armclang

hpx profile $MODEL --engine helia-rt --toolchain atfe     \
    --output-dir results_compare/rt_atfe

hpx profile $MODEL --engine helia-aot --toolchain gcc      \
    --output-dir results_compare/aot_gcc

hpx profile $MODEL --engine helia-aot --toolchain armclang \
    --output-dir results_compare/aot_armclang

hpx profile $MODEL --engine helia-aot --toolchain atfe     \
    --output-dir results_compare/aot_atfe

No engine config is needed by default — heliaAOT resolves its CMSIS-NN kernels from the NSX registry. To override that (or set other engine.config options), pass --engine-config aot.yml with a YAML file such as:

cmsis_nn_path: /path/to/ns-cmsis-nn

See Engines → heliaAOT.

What you get#

for dir in results_compare/*/; do
  total=$(jq -r '.total_cycles' "$dir/summary.json")
  printf "%-30s %12s cycles\n" "$(basename "$dir")" "$total"
done

rt_gcc                            2,014,841 cycles
rt_armclang                       1,874,429 cycles
rt_atfe                           1,881,203 cycles
aot_gcc                           1,965,501 cycles
aot_armclang                      1,869,210 cycles
aot_atfe                          1,872,977 cycles

(Illustrative — the relative gap between toolchains, not the absolute counts, is what to look at. Run the commands above to get your own numbers for your model and board.)

To see where the toolchain difference comes from, line up the per-layer CSVs for one engine:

paste -d, \
    <(cut -d, -f2,3 results_compare/rt_gcc/profile_results.csv) \
    <(cut -d, -f3   results_compare/rt_armclang/profile_results.csv) \
  | head

What this is not#

This does not isolate the engine choice from the toolchain choice — it varies both. To isolate the engine effect, hold toolchain constant and only flip --engine (see Engine Comparison). To isolate placement effects, also hold --arena-location / --weights-location constant (see Memory Placement Tuning).

Where to go deeper#