In-Place-Folding of Non-Scalar Hyper-Planes of Multi-Dimensional Arrays

Abstract

Memory management plays a key role when trying to compile functional programs into efficiently executable code. In particular when using flat representations for multi-dimensional arrays, i.e., when using a single memory block for the entire data of a multi-dimensional array, in-place updates become crucial for highly competitive performance. This paper proposes a novel code generation technique for performing fold-operations on hyper-planes of multi-dimensional arrays, where the fold-operation itself operates on non-scalar sub-arrays, i.e., on vectors or higher-dimensional arrays. This technique allows for a single result array allocation over the entire folding operation without requiring the folding operation itself to be scalarised. It enables the utilisation of vector operations without any added memory allocation or copying overhead. We describe our technique in the context of SaC, sketch our implementation in the context of the compiler sac2c and provide some initial performance measurements that give an indication of the effectiveness of this new technique.