Accelerating Complex Stencil Computations with Adaptive Fusion Strategy

Stencil computation is an important computational pattern widely utilized in various scientific applications, such as image processing, climate forecasting, and fluid dynamics. With the increasing demands for higher precision by scientific applications, stencil computations have become complex, containing a set of dependent stencil operators that may process multiple input grids. These stencils are referred to as complex stencils. For complex stencils, optimizing individual stencil operators is insufficient, and there is significant interest in developing optimization approaches across stencil operators. Existing stencil optimizations or compilers adopt the producer-consumer fusion of stencil operators to eliminate intermediate data access for better performance, however, neglecting the performance opportunity by exploiting inter-operator parallelism through parallelism fusion of stencil operators. To address the above limitation, we propose Plasticine, an adaptive fusion framework for complex stencil computations on GPU. We begin by introducing parallelism fusion, which enables the fusion of concurrent stencil operators for improved performance. Then, a novel multi-level complex stencil representation that effectively captures the characteristics of stencil programs is designed and a CNN-GNN-based model is utilized for fusion strategies selection. The experimental results show that our work, Plasticine, achieves consistently superior performance over state-of-the-art stencil compilers.