Pipelined multiprocessing with upstream processor concurrently writing to local register and to register of downstream processor

Abstract

A multiprocessor system comprises a series of processors arranged to process data in an assembly-line fashion. Each processor includes an executor (execution unit, instruction decoder, and program counter) and a set of registers. Each set of registers is divided into two banks. At any given time, one bank is the “active” bank that is accessible by the local processor, and the other is the “shadow” bank, inaccessible to the local processor. Each processor but the last writes in parallel to its active bank and to the shadow bank of the immediate downstream processor. When all processors have completed working the data in their respective possession, a context-switch is performed switching register banks so that former active banks become shadow banks and former shadow banks become active banks. This makes data that was being processed by an upstream processor virtually immediately available to a local processor. This saves the latency that would be involved in transferring register data after the data is processed. Accordingly, system throughput is improved.