Run repeated quantum experiments

Repeated simulations execute a measurement loop without resetting quantum state between iterations. Replay mode reuses a verified control schedule instead of executing the CPU and TCU for every iteration.

Select a Python backend, install the experiments extra and enable QSBIT_PYTHON_BACKENDS. The AllXY example includes a program, configuration and numerical checks.

Configuration

Add simulation to a run configuration and launch it with qsbit-sim --config FILE. Use --check-config to validate the ELF, mappings and backend without executing rounds or writing outputs. Strategy runs accept no other CLI overrides.

{
  "simulation": {
    "execution": "replay",
    "quantum_execution": "direct",
    "region": {"begin": "round_begin", "end": "round_end"},
    "repetition_count_symbol": "repetitions",
    "repetitions": 1000
  }
}

Field	Meaning
execution	full executes every round on the control microarchitecture. replay records and checks two three-round runs, then repeats recorded operation batches.
quantum_execution	direct evolves and measures the backend for every shot. transition_probabilities uses a supporting backend’s single-qubit transition probabilities. Default: direct.
region.begin, region.end	ELF symbols delimiting the repeated instructions; begin is inclusive and end is exclusive.
repetition_count_symbol	File-backed writable 32-bit word loaded by the loop. The runner patches a working copy of the ELF.
repetitions	Integer from 1 through 4,294,967,295.

full requires quantum_execution: direct. The original ELF is unchanged. Program paths and output paths resolve relative to the run file. The watchdog must cover the requested total duration, including replayed rounds.

Program structure

Keep the ELF symbol table and use this RV32I loop:

.include "quantum.inc"
.section .text
.global _start, round_begin, round_end
_start:
  la t0, repetitions
  lw s0, 0(t0)
round_begin:
  wait.i 40000
  cw.i.i 0, 1
  wait.i 4
  cw.i.i 0, 6
  fmr zero, 0
round_end:
  addi s0, s0, -1
  bnez s0, round_begin
  sim_exit
.section .data
.global repetitions
repetitions: .word 1

The codeword numbers must match the configured mappings. The runner verifies the load, loop suffix and exit instructions. The region accepts only cw.i.i, positive wait.i and fmr zero; classical instructions, branches and measurement-result registers are rejected. This prevents measurement outcomes and loop-carried CPU data from changing the schedule.

Mappings must select unconditional gates, paired gate outputs or acquisitions without separate discriminator arms. Read every acquired target before another codeword or wait, and finish the region with FMR. Disable profile.fast_feedback. All operations must finish within their round, and the round duration must preserve both clock phases. Raw images, session resets, memory dumps and a trace path override are unsupported.

Replay and quantum state

The first round retains its startup timing. Later rounds use the steady event sequence, with absolute times advanced by the verified period. The backend is reset once, not once per round. Each replayed measurement receives a new identity. Direct replay evolves the final idle interval through the extrapolated controller stop tick after the last round.

The control checks run with fixed zero and one outcomes. The program restrictions exclude data-dependent control flow that could change the operation schedule. The recorded instructions, queue events and quantum operations must agree. The final two rounds must also have the same operation sequence after subtracting their time offset. Outstanding device operations at a region boundary cause an error.

transition_probabilities requires one qubit, time-independent, memoryless noise parameters and ideal computational-basis projective measurement. Aer supports this option for its circuit adapter. It calculates each measurement probability from the preceding zero or one result, including all intervening evolution and gates. Sampling preserves that dependence. This method has the same probability model as direct evolution but uses a different random sampling sequence; identical seeds need not produce identical counts across methods.

Outputs

The summary path is required. Its JSON records counts in measurement order within the round, backend options, execution settings, the resolved profile, program hash, repetition counts and host runtime. control_repetitions_executed distinguishes actual control runs from replayed statistics. control_stop_tick_source marks controller completion as observed or extrapolated.

Control ELF copies, call recordings, traces and summaries are written beside the result in <summary-stem>-control/. Calibration runs omit CpuStalled records; instruction, queue and device events retain their timestamps. Ordinary runs can set trace_stalls: false for the same trace filtering.