Tiling for DMA-Based Hardware Accelerators

Explore a compiler pass implemented in MLIR that addresses the challenges of tiling for DMA-based hardware accelerators. Learn how polyhedral analysis is used to examine memory access patterns in loop nests and constrain tile sizes based on DMA chunk width limitations. Discover how this approach enables effective loop tiling for architectures with fixed-width DMA transfers, overcoming the limitations of current automatic loop tilers. Gain insights into optimizing compiler techniques for hardware accelerators in this 12-minute conference talk presented at LCTES'23 by ACM SIGPLAN.