Spatial software pipelining on distributed architectures for sparse matrix codes

Abstract

Wire delays and communication time are forcing processors to become decentralized modules communicating through a fast, scalable interconnect. For scalability, every portion of the processor must be decentralized, including the memory system. Compilers that can take a sequential program as input and parallelize it (including the memory) across the new processors are necessary. Much research has gone towards the ensuing problem of optimal data layout in memory and instruction placement, but the problem is so large that some aspects have yet to be addressed. This thesis presents spatial software pipelining, a new mechanism for doing data layout and instruction placement for loops. Spatial software pipelining places instructions and memory to avoid communication cycles, decreases the dependencies of tiles on each other, allows the bodies of loops to be pipelined across tiles, allows branch conditions to be pipelined along with data, and reduces the execution time of loops across multiple iterations. This thesis additionally presents the algorithms used to effect spatial software pipelining. Results show that spatial software pipelining performs 2.14x better than traditional assignment and scheduling techniques for a sparse matrix benchmark, and that spatial software pipelining can improve the execution time of certain loops by over a factor of three.