Vector Processing on Scalar Architectures

A 64-bit processor must necessarily implement substantial parallelism in the movement and processing of data. Data movement is inherently parallel at the bit level, and many operations implemented in the integer unit exhibit bit-level parallelism. A microvector is an array of small data items or bit elds packed into a single word. Scalar operations performed on a microvector can be used to implement vector parallelism on its constituent elds. In this paper we present libuvec, a library for computing with data which is packed into bit elds. We present experimental results showing excellent performance, up to 30 times the speed of scalar arithmetic in certain applications.