A Flexible Compute and Memory Infrastructure for High-Level Language to Hardware Compilation

We present a low-level infrastructure for use by high-level language to hardware compiler back-ends. It consists of the highly parameterizable, technology-independent module library Modlib and the LMEM framework for localizing variables in fast on-chip memories. Modlib not only supports all high-level language operators (including memory accesses), but also provides a wide spectrum of usage modes: covering static and dynamic scheduling, speculative predicated execution, pipeline balancing, and explicit canceling of mis-speculated computations. We examine the performance of the infrastructure for a number of automatically compiled kernels, including an MD5 kernel that significantly profits from using LMEM.