Techniques for code and data management in the local stores of the cell processor

Abstract

The Cell Broadband Engine (CBE) contains one conventional PowerPC core with a standard cache hierarchy and eight Synergistic Processing Elements (SPEs), each of which contains a fast but small Local Store. These Local Stores are similar to ScratchPad memories, in that they are small, fast and directly addressed, but an additional constraint on programming the SPEs is that all Load/Store and instruction fetch accesses from the processor reference only the local stores. Access to the rest of the memory on a CBE system must be explicitly programmed using DMA commands. Techniques for fitting programs with large text and data footprints into the Local Stores, both manual and automatic will be described. Copyright 2007 ACM.