Co-specialization of Hardware and Software

摘要

Hardware is becoming increasingly specialized because of the need for power efficiency. One way to gain efficiency is to use throughput-oriented processors (e.g. GPUs) optimized for data-parallel applications; these processors deliver more gigaflops per watt than CPUs optimized for single-threaded programs. On mobile and embedded platforms, where batteries limit the available energy, systems-on-a-chip contain multiple processors along with many specialized hardware units. Since most applications perform many different types of computations, the optimal platform will contain a heterogenous mixture of different types of processing elements. As the hardware has become more diverse, so have the programming models. This is because the low-level programming models reflect the underlying hardware abstractions. Unfortunately, the diversity in programming models makes it challenging for software developers to use emerging hardware platforms. The classic method for writing portable programs is to use a general-purpose programming language and an optimizing compiler. Unfortunately, current compilers are not powerful enough for todays platforms. Programs written for one class of machines will not run efficiently on another class of machines. My thesis is that the only practical method for writing programs for heterogeneous machines is to raise the level of the programming model. In particular, I advocate the use of domain-specific languages (DSLs). I will present the case for using DSLs, show how DSLs can cope with diverse hardware, and outline several areas of programming language research that may lead to better methods for building DSLs. Given the trend towards specialized hardware, it seems natural to co-specialize the software.