Performance Management Guide

Optimizing Preprocessors for FORTRAN and C

The KAP and VAST preprocessors for the FORTRAN compiler can restructure FORTRAN source code to better use the POWER family, POWER2, and PowerPC processing unit resources and memory hierarchy. A version of the KAP preprocessor is also available for restructuring the code of C programs. The preprocessors perform memory management optimizations, algebraic transformations, inlining, interprocedural analysis, and other optimizations that improve the performance of FORTRAN or C applications.

Performance tests indicate improvements in the range of 8 to 18 percent, on average, when a suite of programs is compiled with the preprocessors, compared to compiling with the same optimization options for the unpreprocessed version.

The KAP and VAST preprocessors attempt to transform source-level algorithms into algorithms that can take full advantage of the optimizing capabilities of the compiler. The preprocessors also generate listings that identify the transformations performed and areas of your code that prevent transformations from being carried out. The preprocessors analyze source code, and perform transformations that can improve a program's performance.

Any transformation done by the preprocessors can also be accomplished through hand-tuning. The advantages of using a preprocessor rather than hand-tuning are as follows:

• In many cases, the preprocessors yield programs that perform as efficiently as, or more efficiently than, their hand-tuned equivalents, without a significant investment of programmer time. If you use the preprocessors, you may not require as thorough an understanding of the architecture or of tuning techniques discussed elsewhere in this book.
• For certain programs, you can get code that is highly optimized, simply by selecting appropriate command-line preprocessor options and by adding a small number of directives to the source code of your program. In cases where the preprocessors do not yield a noticeable improvement, work with the preprocessor listings to see what areas of the source code prevent optimization.
• Some of the transformations done by the preprocessors involve considerable expansion of source code. While these expansions can improve your program's efficiency, implementing them through hand-tuning would increase the likelihood of algorithmic and typographical errors, reduce the readability of the source code, and make program maintenance more difficult.
• The preprocessors can generate code that is tuned for a particular architectural configuration, even one that is not available on POWER family, POWER2, and PowerPC systems. You can maintain a single version of source code, and produce transformed versions that are tuned for different POWER, POWER2, and PowerPC models or for machines with other cache and processor characteristics.
• The preprocessors can often improve on hand-tuned code. Although it is possible to tune your programs by hand to as great a level of efficiency as the preprocessors do, some of the more complicated transformations can lead to coding errors when attempted by hand.