parameter_selection

parameter_selection#

Full path: iqm.qrisp_iqm.passes.routing.core.parameter_selection

Parameter selection for the SABRE layout-and-route transpiler.

Maps a single effort integer to the internal tuning parameters (init_layout_attempts, C, layout_iterations) that control how much classical compute is invested in finding a good mapping.

Design rationale#

The transpiler has three independent knobs:

  • init_layout_attempts (ILA) — how many random starting layouts to evaluate. Up to cpu_count layouts are evaluated in parallel; each additional batch of cpu_count adds one parallel-wave.

  • layout_iterations — how many back-and-forth refinement passes each layout candidate receives.

  • C — routing diversity multiplier. C * cpu_count independent routing runs are performed for the best layout.

The relative importance of ILA vs C depends on the number of 2-qubit gates in the circuit:

  • Shallow circuits (few 2-qubit gates): layout choice dominates — invest more in ILA.

  • Deep circuits (many 2-qubit gates): routing stochasticity dominates — invest more in C.

The crossover follows the empirically fitted relationship ILA_importance% = -11.14 * log2(num_2qb_gates) + 114.5 (R² = 0.69).

Layout iterations saturate at 3 for 14/16 tested regimes.

Fitting methodology#

The formula was calibrated on an 11,520-point parameter sweep covering grid sizes 3×3 to 11×11 and circuits with 30 to 1000 random 2-qubit gates. Per-regime response surfaces depth ~ 1/(log ILA + 1) + 1/(log C + 1) achieved R² = 0.72–0.90.

Validation (Formula C in fit_parameters_v3.py):

  • Median efficiency vs model-optimal = 100%

  • Mean efficiency = 77%

  • Worst-case absolute depth penalty < 0.7 gates (< 0.5%)

Functions

compute_parameters(effort, num_2qb_gates[, ...])

Derive transpiler parameters from a single effort level.
Contents