Customizing the Computation Capabilities of Microprocessors.

Abstract

Designers of microprocessor-based systems must constantly improve performance and increase computational efficiency in their designs to create value. To this end, it is increasingly common to see computation accelerators in general-purpose processor designs. Computation accelerators collapse portions of an application's dataflow graph, reducing the critical path of computations, easing the burden on processor resources, and reducing energy consumption in systems. There are many problems associated with adding accelerators to microprocessors, though. Design of accelerators, architectural integration, and software support all present major challenges.

This dissertation tackles these challenges in the context of accelerators targeting acyclic and cyclic patterns of computation. First, a technique to identify critical computation subgraphs within an application set is presented. This technique is hardware-cognizant and effectively generates a set of instruction set extensions given a domain of target applications. Next, several general-purpose accelerator structures are quantitatively designed using critical subgraph analysis for a broad application set.

The next challenge is architectural integration of accelerators. Traditionally, software invokes accelerators by statically encoding new instructions into the application binary. This is incredibly costly, though, requiring many portions of hardware and software to be redesigned. This dissertation develops strategies to utilize accelerators, without changing the instruction set. In the proposed approach, the microarchitecture translates applications at run-time, replacing computation subgraphs with microcode to utilize accelerators. We explore the tradeoffs in performing difficult aspects of the translation at compile-time, while retaining run-time replacement. This culminates in a simple microarchitectural interface that supports a plug-and-play model for integrating accelerators into a pre-designed microprocessor.

Software support is the last challenge in dealing with computation accelerators. The primary issue is difficulty in generating high-quality code utilizing accelerators. Hand-written assembly code is standard in industry, and if compiler support does exist, simple greedy algorithms are common. In this work, we investigate more thorough techniques for compiling for computation accelerators. Where greedy heuristics only explore one possible solution, the techniques in this dissertation explore the entire design space, when possible. Intelligent pruning methods ensure that compilation is both tractable and scalable.