A general transposition method for a matrix on auxiliary store

Abstract

An algorithm is developed and described for transposing a matrix larger than available working storage. If an (n×m)-matrix is stored in row-major order, and blocks ofn elements may be transferred to and from working storage at a time, the algorithm needsw=(5[m/n]+8)·n elements to be present in working storage at a time and requires [log₂(2mn/w)] passages over the matrix. The algorithm is as efficient as earlier methods but needs no extra backing storage space. An algebra for mixed radix notation and a generalization of mixed radix notation is introduced for the description and verification of transposition algorithms, and earlier algorithms are briefly certified or disproved.