Phoenix is a highly-effective VQA (variational quantum algorithm) application-specifc compiler based on BSF (binary symplectic form) of Pauli exponentiations and Clifford formalism. Different from ZX-calculus-like approaches (e.g., TKet, PauliOpt) and local peephole optimization approaches (e.g., Paulihedral, Tetris), Phoenix exploits global optimization opportunities for VQA programs to the largest extent, when representing Pauli strings as BSF and employing Clifford formalism on the higher-level IR.
This repo includes benchmarking scripts and results with other SOTA baselines -- TKet, Paulihedral, Tetris, and Rustiq. Code of Paulihedral and Tetris are refactored and integrated in this repo.
If you make sure of Phoenix in your work, please cite the following publication:
@article{yang2025phoenix,
title={PHOENIX: Pauli-Based High-Level Optimization Engine for Instruction Execution on NISQ Devices},
author={Yang, Zhaohui and Ding, Dawei and Zhu, Chenghong and Chen, Jianxin and Xie, Yuan},
journal={arXiv preprint arXiv:2504.03529},
year={2025}
}- High-level compilation
- Global optimization
- ISA-independent
E.g., Hardware-agnostic compilation:
E.g., Hardware-aware compilation:
Basic library requirements are lists in requirements.txt.
- We align with the
1.2.4version ofqiskitacross the published benchmarking results, sinceqiskit's O1/O2 has different built-in workflows within its 0.xx.x versions and 1.xx.x versions. Version 1.0+ is suitable for Phoenix. - Originally, Paulihedral and Tetris require version 0.23.x and version 0.43.x of Qiskit. In this code repo, they can also be soomthly tested under Qiskit-1.2.4.
Benchmark suites:
- Hamlib: 100 generic Hamiltonian simulation programs ( a subset of the original Hamlib) from IBM's latest benchpress toolkit, including "binaryoptimization", "discreteoptimization", "chemistry", and "condensedmatter" four categories of Hamiltonian simulation programs.
- UCCSD: 16 molecule simulation programs from benchmarks from TKet benchmarking. We use this suite for fine-grain benchmarking and real system analysis.
Benchmarking scripts: (under ./experiments/scripts/)
bench_single.py: Benchmarking given a compiler and input file (.qasmfile for TKet,.jsonfile for Paulihedral/Tetris/Phoenix)bench_hamlib.py: Benchmarking given a compiler and a category of Hamlib benchmarksbench_uccsd.py: Benchmarking given a compiler and a physical-qubit topology type (E.g., all2all, manhattan, sycamore) (Since TKet/Phoenix does no make hardware-ware co-optimization, manually set topology execept "all2all" are invalid. Instead, the hardware-ware compilation are conducted by executing another scriptuccsd_all2all_to_limited.py)bench_utils.py: Utils used in benchmarking, within which the standard compilation passes of Phoenix and baselines are specifiedhamlib_cnot_to_su4.py: Rebase CNOT-based output circuits to SU(4)-based circuits given a compiler and a category of Hamlib benchmarksuccsd_all2lall_to_limited.py: Compile all-to-all compilation results from TKet/Phoenix for UCCSD benchmarks to limited-connectivity topology (manhattan, sycamore)uccsd_all2all_qiskit_opt.py: Further perform Qiskit O3 for logical-level synthesis results of Paulihedral/Tetris/Phoenix
Result files:
-
./experiments/output_hamlib/<compiler>/<category>/: Output circuits by some<compiler>(e.g., Tetris, Phoenix) from some<category>of the Hamlib benchmark suite -
./experiments/output_hamlib/<compiler>/<category>_su4/: Output circuits that are further rebased to SU(4) ISA (For Phoenix compiler, the SU(4)-based circuits can be natively generated that are the same as the rebased results from CNOT-based output circuits) -
./experiments/output_uccsd/<compiler>/<device>: Output circuits by some<compiler>(E.g., Tetris, Phoenix) for some kind of<device>(E.g., all2all, manhattan) from the UCCSD benchmark suite -
./experiments/output_uccsd/<compiler>_opt/all2all: Output circuits by some<compiler>(E.g., Tetris, Phoenix) when performing its logical-level synthesis with Qiskit O3 optimization procedure on logical circuits
Remarks: We developed in the early beginning of 2024. Suprisingly the BSF algorithm within Phoenix utilized similar data structures and heuristic-optimization philosophy to Rustiq. Despite the accidential similarity, our simplification goal and search strategy are distinct and involves a more sophisticated cost function. Even the BSF simplification algorithm alone outperforms Rustiq according to our further benchmarking in ./experiments/rebuttal.ipynb.
This project is licensed under the Apache License 2.0 -- see the LICENSE file for details.