The Imitation Game: Leveraging CopyCats for Robust Native Gate Selection in NISQ Programs

• Published in: 2023 IEEE International Symposium on High-Performance Computer Architecture (HPCA) pp. 787 - 801
• Main Authors: Das, Poulami, Kessler, Eric, Shi, Yunong
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.02.2023
• Subjects: Gate nativization; NISQ Compilation; Quantum
• ISSN: 2378-203X
• DOI: 10.1109/HPCA56546.2023.10071025

Quantum programs are written in high-level languages, whereas quantum hardware can only execute low-level native gates. To run programs on quantum systems, each high- level instruction must be decomposed into native gates. This process is called gate nativization and is performed by the compiler. Recent quantum computers support a richer native gate set to reduce crosstalk by tackling frequency crowding and enable compilers to generate quantum executables with fewer native gates. On these systems, any two-qubit CNOT instruction can be decomposed using more than a single two-qubit native gate. For example, a CNOT can be decomposed using either XY, CPHASE, or CZ native gates on Rigetti machines. Unfortunately, two-qubit native gates have high-error rates and exhibit temporal and spatial variations, which limits the success-rate of quantum programs. Therefore, identifying the native gate that maximizes the success-rate of each CNOT operation in a program is crucial.Our experiments on Rigetti machines show that noise-adaptive gate nativization to select the native gate with the highest fidelity for each CNOT operation is often sub-optimal at the application level. This is because the performance of such nativization heavily depends on the correctness of the device calibration data which only provides the average gate fidelities and may not accurately capture the error trends specific to the qubit state space of a program. Moreover, the calibration data may go stale due to device drifts going undetected. To overcome these limitations, we propose Application-specific Native Gate Selection (ANGEL). ANGEL designs a CopyCat that imitates a given program but has a known solution. Then, ANGEL employs the CopyCat to test different combinations of native gates and learn the optimal combination, which is then used to nativize the given program. To avoid an exponential search, ANGEL uses a divide-and-conquer- based localized search, the complexity of which scales linear with the number of device links used by the program. Our evaluations on Rigetti Aspen-11 show that ANGEL improves the success-rate of programs by 1.40x on average and by up-to 2x.