[S-SS09] Seismic wave propagation: Theory and Application

[SSS09-P01] Innovative FP16-SVE Optimization for Accelerating Seismic Simulations

*Wenqiang Wang¹, Juepeng Zheng^2,1 (1.National Supercomputing Center in Shenzhen, Shenzhen, China, 2.School of Artificial Intelligence, Sun Yat-sen University, Zhuhai, China)

Keywords:Seismic simulation, Seismic wave propagation, FP16-SVE algrithms, Low storage seismic numerical method

Seismic simulation plays a crucial scientific role in understanding the propagation of seismic waves and their impact mechanisms. The core challenge of seismic simulation lies in numerically solving the elastic wave equation. Traditional methods relying on single-precision (FP32) and scalar operations suffer from high memory consumption and low efficiency. Compared to FP32, half-precision floating-point format (FP16) reduces memory usage by half and doubles memory access efficiency. Meanwhile, compared to scalar operations, scalable vector extension (SVE) operations utilize a single instruction to process multiple data elements (SIMD), significantly accelerating computation. Therefore, to address these challenges in memory consumption and computation efficiency, this paper develops FP16-based seismic simulation methods, SVE vectorization acceleration algorithms, and an innovative FP16-SVE hybrid algorithm on Kunpeng architecture. The FP16-SVE hybrid algorithm combines the memory efficiency of FP16 with the vectorization capabilities of SVE, optimizing performance for seismic simulations. Based on these methods, three seismic simulation solvers were implemented, validated, and benchmarked. The results demonstrated significant performance improvements, with the FP16-SVE hybrid solver achieving nearly 3x speedup and halving memory usage compared to the standard solver. This innovative solution enhances computational efficiency and resource utilization, offering transformative potential for large-scale and real-time seismic applications.