Skip to Main content Skip to Navigation

Polyhedral Code Generation: Reducing Overhead and Increasing Parallelism

Harenome Razanajato 1, 2
2 CAMUS - Compilation pour les Architectures MUlti-coeurS
Inria Nancy - Grand Est, ICube - Laboratoire des sciences de l'ingénieur, de l'informatique et de l'imagerie
Abstract : This thesis proposes new extensions to the code generation phase in polyhedral compilers. The main focus of recent work on polyhedral compilation focus is the optimizations leveraged by polyhedral transformations while state-of-the-art code generation algorithms are considered satisfactory. We show that state-of-the-art polyhedral code generation can still be further improved. We explain how splitting polyhedra can reduce the control overhead introduced by polyhedra scanning in the code generated by a polyhedral compiler. Synchronizations in parallel code can drastically impede a program’s performance. We propose a method to detect and lift unnecessary synchronization barriers. Finally, we introduce pipelined multithreading, a transformation that introduces parallelism in a class of programs that was, until now, ignored by polyhedral parallelizers.
Complete list of metadata
Contributor : Vincent Loechner Connect in order to contact the contributor
Submitted on : Friday, January 29, 2021 - 4:14:07 PM
Last modification on : Thursday, December 2, 2021 - 3:17:04 AM


  • HAL Id : tel-03125702, version 1


Harenome Razanajato. Polyhedral Code Generation: Reducing Overhead and Increasing Parallelism. Distributed, Parallel, and Cluster Computing [cs.DC]. Université de Strasbourg (Unistra), FRA., 2020. English. ⟨tel-03125702⟩



Les métriques sont temporairement indisponibles