Automatic Parallelization - New Approaches to Code Generation, Data Distribution, and Performance Prediction

This paper introduces the Weight Finder, an advanced profiler for Fortran programs,which is based on a von Neumann architecture. Existing Fortran codes are generally too large toanalyze fully in depth with respect to performance tuning. It is the responsibility of the WeightFinder to detect the most important regions of code in the program, as far as execution time isconcerned. Program transformation systems, compilers and users may then subsequently con-centrate their optimization efforts upon these areas in code.Furthermore program unknowns, such as loop iteration counts, true ratios and frequency infor-mation, are derived. Analysis and prediction systems require concrete values for these unknownsin order to provide reasonable accurate results.Animation, simulation, debugging and trace based tools may use the profile data as beingderived by the Weight Finder in order to detect program parts which are never executed, simulatethe program’s control flow, etc.This tool is based on an optimized instrumented profile run. Several optimizations are shownwhich eliminate large portions of the instrumentation code, thus decreasing profile run-time andmemory requirements, and improving the measurement accuracy. It is shown how the profile datacan be adapted for program transformations to the original Fortran program without redoing theprofile run.The Weight Finder is a 10000 line software package which is fully integrated under the ViennaFortran Compilation System.