Code-Size Conscious Pipelining of Imperfectly Nested Loops

Mohammed Fellahi (1) , Albert Cohen (1) , Sid Touati (2)

1 ALCHEMY - Architectures, Languages and Compilers to Harness the End of Moore Years
2 PRISM - Parallélisme, Réseaux, Systèmes, Modélisation

This paper is a step towards enabling multidimensional software pipelining of non-perfectly nested loops on memory-constrained architectures.We propose a method to pipeline multiple inner loops without increasing the size of the loop nest, apart from an outermost prolog and epilog. We focus on the domain of media and signal processing, where short inner loops are common and where embedded constraints drive the selection of code-size conscious algorithms. Our first results indicate that the additional constraints associated with the method do not impede the extraction of significant amounts of instruction-level parallelism. In addition to preserving precious scratch-pad or cache memory, our method also avoids the performance overhead of prologs and epilogs resulting from pipelined inner loops with short trip count.

Mots clés: Loop nest optimization; software pipelining; multidimensional retiming; code size; startup time

Domaines: Langage de programmation [cs.PL]

Mohammed Fellahi, Albert Cohen, Sid Touati. Code-Size Conscious Pipelining of Imperfectly Nested Loops. MEDEA Workshop (MEmory performance DEaling with Applications, systems and architecture), held in conjunction with PACT 2007 Conference, Sep 2007, Brasov, Romania. ⟨10.1145/1327171.1327177⟩. ⟨hal-00646688⟩