面向FPGA上高效SAT求解的布尔约束传播加速方法
Accelerating Boolean Constraint Propagation for Efficient SAT-Solving on FPGAs
AI总结:
本研究提出一种面向处理器/FPGA SoC的硬件加速SAT求解器,通过子句直接分配、公式分区热交换等设计加速BCP,在Xilinx Zynq平台实现了优于现有方案的加速效果。
摘要:
我们提出了一种面向处理器/现场可编程门阵列(FPGA)片上系统(SoC)的硬件加速SAT求解器。该方案通过细粒度的FPGA并行性,加速Davis-Putnam-Logemann-Loveland(DPLL)算法中计算开销最高的子例程——布尔约束传播(BCP)。与现有最优方案不同,本求解器通过将子句直接分配给FPGA上的子句处理器,并将大型公式拆分为FPGA可处理的更小分区,消除了开销高昂的子句查找操作。各分区可根据需要在运行时热交换,支持的公式规模仅受可用外部内存限制,而非FPGA片上内存。我们在Xilinx Zynq平台上对求解器进行了评估,结果表明,在不同公式规模下执行速度均有提升,且效果受公式分区策略影响。与现有最优方案相比,我们在2个代表性基准测试上的BCP速度分别提升了1.7倍和1.1倍,相较于纯软件实现的总加速比最高可达6倍。
英文摘要:
We present a hardware-accelerated SAT solver targeting processor/Field Programmable Gate Arrays (FPGA) SoCs. Our solution accelerates the most expensive subroutine of the Davis-Putnam-Logemann-Loveland (DPLL) algorithm, Boolean Constraint Propagation (BCP) through fine-grained FPGA parallelism. Unlike prior state-of-the-art solutions, our solver eliminates costly clause look-up operations by assigning clauses directly to clause processors on the FPGA and dividing large formulas into smaller partitions manageable by FPGA. Partitions are hot-swapped during runtime as required and the supported formula size is limited only by available external memory, not on-chip FPGA memory. We evaluate our solver on a Xilinx Zynq platform with results showing quicker execution time across various formula sizes, subject to formula partitioning strategy. Compared to prior state-of-the-art, we achieve 1.7x and 1.1x speed up on BCP for 2 representative benchmarks and up to 6x total speedup over software-only implementation.
机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。