Compiling array references with affine functions for data-parallel programs

• Main Authors: Wei Wen-Hsing, 韋文祥
• Other Authors: Sheu Jang-Ping
• Format: Others
• Language: en_US
• Published: 1995
• Online Access: http://ndltd.ncl.edu.tw/handle/47120745579309815449

碩士 === 國立中央大學 === 資訊及電子工程研究所 === 83 === Generating local memory access sequence and necessary communication sets is very important issue for compiling an array reference in a data-parallel program. In this thesis, we present a class table approach to efficiently enumerate the local memory access sequence and evaluate communication sets. For an given linear array reference, supposing the array is distributed with a cyclic(t) distribution across P processors, local memory access sequence of any processor for a computation involving the array references with one index variable is characterized by a class table of at most s items, where s indicates the memory stride of the array reference. We also take advantage of the class table to present the communication sets as closed forms from viewpoint of blocks when arrays are distributed block-cyclicly, and derive efficient algorithms for sending and receiving the necessary data between processors. Moreover, we present an algorithm for computing the essential information of the class table and the time that this algorithm takes is less than or equalto O(s). Finally, we adapt our approach to handle the array references with affine functions of index variables. In previous work involving generation of local memory access sequence and communication sets, they solved these problems for the model of linear array reference. In this thesis, we extend the array reference model of linear functions to affine functions and the overhead for construction of the class table is only O(s) time.