UC San Diego

With general-purpose compute stacks struggling to meet computational demands of emerging applications, there has been a shift in industry and the research community toward domain-specific architectures. Each of these specialized architectures is designed with domain-specific properties in mind, exploiting their algorithmic structure by using specialized hardware capabilities. The divergence of architectures designs makes them incompatible with mature compilation platforms which were designed to target general purpose processors, and do not incorporate the domain knowledge necessary for optimizing specific classes of program.

This challenge calls for rethinking of the compiler that has traditionally focused on generating fine-grained instruction sequences while the micro-architecture is abstracted away almost completely. Therefore, this proposal will discuss a two-part solution consisting of 1) a high-level programming language and intermediate representation called PolyMath (HPCA '20) which preserves multiple levels of operation granularity to flexibly enable compilation to different architectures and 2) An architecture abstraction to define and expose specialized microarchitectures to the compiler, allowing the same compilation workflow to be reused for different targets, accompanied by a low-level kernel abstraction which is automatically scheduled using the architecture abstraction. Together, PolyMath and the architecture abstraction allow domain-specific programs to be easily defined and analyzed through PolyMath, which generates an target-agnostic IR for coupling with information embedded in the architecture abstraction to compile into a valid schedule and generate executable binaries.