Subword parallelism with MAX-2

MAX-2 illustrates how a small set of instruction extensions can provide subword parallelism to accelerate media processing and other data-parallel programs. This article proposes that subword parallelism-parallel computation on lower precision data packed into a word-is an efficient and effective solution for accelerating media processing. As an example, it describes MAX-2, a very lean, RISC-like set of media acceleration primitives included in the 64-bit PA-RISC 2.0 architecture. Because MAX-2 strives to be a minimal set of instructions, the article discusses both instructions included and excluded. Several examples illustrate the use of MAX-2 instructions, which provide subword parallelism in a word-oriented general-purpose processor at essentially no incremental cost.