System-Level Design and Target Agnostic High-Level Optimizations for HLS
TimeWednesday, July 13th6pm - 7pm PDT
LocationLevel 2 Lobby
DescriptionThe generation of custom accelerators for applications implemented within high-level software frameworks requires considerable manual effort. To automate this process, we introduce a front-end compiler tool that extends the MLIR infrastructure. Our tool automatically searches, outlines, and pre-optimizes relevant code regions to generate high-quality accelerators through High-Level Synthesis.
Experimental results with kernels from the Polybench benchmark show that our high-level optimizations expose code structures resulting in accelerators with up to 114x faster runtime over an unoptimized baseline. We also demonstrate the applicability of the framework by exploring the automatic generation of accelerators for Deep-Neural-Networks with operations fused at arbitrary granularity.