Efficient Dynamic Local Enumeration for HPF

In translating HPF programs, a compiler has to generate local iteration and communication sets. Apart from local enumeration, local storage compression is an issue, because in HPF array alignment functions can introduce local storage inefficiencies. Storage compression, however, may not lead to serious performance penalties. A problem in semi-automatic translation is that a compiler should generate efficient code in all cases the user may expect efficient translation (no surprises). However, in current compilers this turns out to be not always true. A major cause for this inefficiencies is that compilers use the same fixed enumeration scheme in all cases. In this paper, we present an efficient dynamic local enumeration method, which always selects the optimal solution at run-time and has no need for code duplication. The method is compared with the PGI and the Adaptor compiler. Dynamic selection of enumeration orders Once a data mapping function is given in an HPF program, we know exactly which element of an array is owned by which processor. However, the storage of all elements owned by a single processor still needs to be determined. When a compiler is to generate code for local iteration or communication sets, it needs to enumerate the local elements efficiently. Efficient in the sense that only a small overhead is allowed compared to the work inside the iteration space. Because in both phases local elements need to be referenced, storage and enumeration are closely linked to each other. An overview of the basic schemes for local storage and local enumeration is given in [1]. In a cyclic(m) distribution the template data elements are distributed in blocks of size m in a round robin fashion. The relation between an array index i and the row, column, and processor tuple (r, c, p) is given by the position equation [1]. To avoid inefficient memory storage, local storage is compressed by removing unused template elements. There are various compression techniques, each with their own compression factor for rows (∆r) and columns (∆c). For a cyclic(m) distribution, the original (normalized) volume assignment is transformed into a two-deep loopnest. The outer loop enumerates the global indices of the starting points of the rows (order=row wise) or the columns (order=column wise), depending on the enumeration order as specified by order. In most compilers the order of enumeration is fixed. However, we have mofified the method outlined in [1] such that the generated code for both enumeration orders is identical, by adding a parameter ‘order’ to the run-time routines that