Advanced Optimization Techniques for High Performance Fortran

Abstract

With current commercial HPF compilers, users often have to restructure their codes extensively to obtain good performance, even for regular data-parallel applications on message-passing systems. The Rice dHPF compiler project is aimed at developing advanced optimization techniques that can provide consistently high performance for a broad spectrum of scientific applications with minimal restructuring of existing Fortran 77 or Fortran 90 applications. The foundation of the compiler is an abstract, integer-set-based approach for data-parallel program analysis and optimization. This framework has enabled us to implement a comprehensive collection of compile-time optimizations including advanced optimizations that either were never implemented before or implemented in a restricted form in a few compilers.

The project is currently focused on HPF versions of the NAS application benchmarks, which we have developed with minimal rewriting of the serial benchmarks (modifying less than 100 source lines per application). We have developed several essential compiler optimizations required by features common to such real-world benchmarks (in addition to the traditional data-parallel optimizations described in the literature). These optimizations include:

• a flexible computation partitioning strategy to avoid expensive inner-loop communication for privatizable arrays (arrays can be specified as privatizable within a loop by using the NEW directive);
• intelligent loop distribution to avoid inner-loop communication while preserving temporal locality due to intra-processor data reuse;
• parallelizing loops containing calls to non-pure functions; and
• multi-dimensional coarse-grain pipelining for wavefront computations.

Notes by Chuck Koelbel