Extension of PrismStencil to Sparse Tensor Cores

Extend the PrismStencil stencil tensorization algorithm to exploit sparse Tensor Cores while preserving its low computational inflation and data-reuse benefits.

Background

PrismStencil is designed for dense Tensor Core matrix-multiply-and-accumulate operations and reduces tensorization overhead through multidimensional vector mapping, unified gather–scatter tensorization, and cross-product data reuse. The paper observes that stencil-to-matrix transformations naturally produce zero entries, making them potentially suitable for sparse Tensor Core acceleration.

The discussion compares PrismStencil with SPIDER, a sparse-Tensor-Core approach that exploits 2:4 structured sparsity. Sparse Tensor Cores require larger MMA shapes than dense Tensor Cores, which can introduce additional alignment overhead. Although PrismStencil’s lower computational inflation may make it suitable for these larger sparse shapes, the paper does not implement or evaluate such an extension and explicitly leaves it for future work.