Faster classical simulation becomes essential for the validation of quantum computer, and tensor network contraction is a widely-applied simulation approach. Due to the memory limitation, slicing is adopted to help cutting down the memory size by reducing the tensor dimension, which also leads to additional computation overhead. This paper proposes novel lifetime-based methods to reduce the slicing overhead and improve the computing efficiency, including: interpretation for slicing overhead, an in place slicing strategy to find the smallest slicing set, a corresponding iterative method, and an adaptive path refiner customized for Sunway architecture. Experiments show that our in place slicing strategy reduces the slicing overhead to less than 1.2 and obtains 100-200 times speedups over related efforts. The resulting simulation time is reduced from 304s (2021 Gordon Bell Prize) to 149.2s on Sycamore RQC, with a sustainable mixed-precision performance of 416.5 Pflops using over 41M cores to simulate 1M correlated samples.
View on arXiv