AI 中文总结
该研究针对格点手术量子编译提出资源感知方法Harvest，协同优化魔术态消耗、布局与路由，在基准套件上实现显著加速，提升调度长度并回收大量未使用资源。
AI 中文摘要
基于拓扑码的容错量子处理器通过格点手术执行程序，其中的操作必须在物理块的二维网格上进行映射、路由并配备魔术态。非克利福德（Clifford）操作需要这些魔术态，魔术态可通过蒸馏工厂或 cultivation（制备）产生，二者在占用空间与制备延迟之间存在权衡，且将魔术态输送至消耗它的数据块需要与其他所有操作共享同一布局的路由。然而，布局、路由、调度与魔术态供应无法单独优化：两个无电路级依赖的操作在放置后仍可能争夺相同端口、路由或魔术态终端，因此，若编译器将指令调度与魔术态生成解耦，或硬编码单一生成协议，将被迫在执行时间与布局占用空间之间权衡，而非跨协议协同优化二者。我们提出 Harvest，一种面向格点手术的资源感知编译方法，它在协议无关的资源模型下，协同优化魔术态消耗、电路感知布局与拥塞感知路由，随后在调度后回收未使用的布局占用空间。在标准基准套件（QAOA、QFT、QASMBench）上，与顺序执行相比，Harvest 实现了平均 4.83 倍的加速（最高达 17.8 倍），通过电路感知布局将调度长度提升至最高 1.35 倍，并回收了最高 72.0% 的未使用魔术态块与 33.9% 的未使用路由块。
英文摘要
Fault-tolerant quantum processors based on topological codes execute programs through lattice surgery, where operations must be mapped, routed, and supplied with magic states across a 2D grid of physical patches. Non-Clifford operations require these magic states, produced either by distillation factories or by cultivation, each trading footprint against preparation latency, and delivering a magic state to the data patches that consume it requires routing through the same shared layout as every other operation. Yet placement, routing, scheduling, and magic-state supply cannot be optimized in isolation: two operations with no circuit-level dependency can still contend for the same ports, routes, or magic-state terminals once placed, so a compiler that decouples instruction scheduling from magic-state generation, or hard-codes a single generation protocol, is forced to trade execution time against layout footprint instead of co-optimizing both across protocols. We present Harvest, a resource-aware compilation approach for lattice-surgery that co-optimizes magic-state consumption with circuit-aware placement and congestion-aware routing under a protocol-agnostic resource model, then reclaims unused layout footprint after scheduling. Across standard benchmark suites (QAOA, QFT, QASMBench), Harvest achieves an average speedup of $4.83\times$ (up to $17.8\times$) over sequential execution, improves schedule length by up to $1.35\times$ through circuit-aware placement, and reclaims up to $72.0\%$ of unused magic-state patches and $33.9\%$ of unused routing patches.
