Reducing Instruction Issue Overheads in Application-Specific Vector Processors

The traditional approach to IP core design is to use simulations with test vectors. This is not feasible when dealing with complex function cores such as the Image Segmentation case-study algorithm in this paper. An algorithm developer needs to carry out experiments on large real-world data sets, with fast turn-around times, and in real time to facilitate performance tuning and incremental development. Previously we proposed a methodology called Application-Specific Vector Processor (ASVP). The ASVP approach first constructs a programmable architecture customized for a given application, then employs software techniques to develop firmware that implements the algorithm. In our setting we employ an embedded simple scalar CPU (8-bit PicoBlaze 3) to control a floating-point vector processing unit (VPU) by issuing wide (horizontally encoded) instructions to it. In this work we dramatically reduce the overhead of the wide-instruction issue (in one case by 13x) by implementing a new two-level configuration table. The table stores frequently used vector definitions (in Level 1) and vector instructions (in Level 2), pre-loading them quickly into the issue buffer. A configuration in the issue buffer can be further modified before being sent to the processing unit. This ensures the architecture stays general and fully customizable.