HPF compilers, in general, are fairly "young" as compared to compilers for other standardized languages such as Fortran or C. As a result, there is not yet a common body of practice, widely accepted, that is a standard yardstick for comparing all HPF compilers. This chapter makes the obvious comparison of which defined HPF language features are supported by each reviewed compiler. Then the capabilities of each compiler in handling FORTRAN 77 or Fortran 90 features, particularly with respect to parallelism and parallelization are tabulated. Each compiler has some associated performance analysis tools and some extensions beyond Standard Fortran or HPF. Some of these are tabulated. Finally some subjective perceptions as to the relative strengths of each compiler are given, derived from some application codes that have been implemented in or converted to HPF at the CTC, and from some inspection of and interaction with CTC user's codes.

Later chapters show more details of the generated code from each compiler for a few small kernels of HPF code. These are discussed to give a better feel for the compilation behavior of each compiler, as perceptible to a user of the compiler.

A complication in forming evaluation criteria for these compilers is the current distinction between HPF and Subset HPF. Indeed, more confusion is added to any feature comparisons based on the full and Subset languages because a number of detailed additions and subtractions were made to the contents of the Subset during the interpretations and corrections that led from "HPF 1.0" of May 1993 to the current "HPF 1.1" of November 1994. (Note that the HPF Forum has reconsidered the status of the "Subset HPF" definition in that the soon to be adopted proposals for "HPF 2.0" do not use that form of two-level designation.)

It would be desirable in a comparative review paper such as this to present carefully measured data on, for instance, compilation duration timing, size of generated executable files, runtime execution duration timing, etc., for a well studied collection of benchmark kernels and whole applications. This has not been done for this paper at this time.

None of these three compilers is completely compliant, or effective, for the full facilities of Fortran 90 combined with the full HPF language. Two compilers (xHPF and XL HPF) are designated as "Extended Subset HPF" compilers by their vendors, and each has significant extensions into facilities of the full language, and other departures from the Subset language. It is thus better to describe all these compilers by listing all the HPF features they will or will not process rather than to attempt to categorize them as "full" or "Subset".

The following table indicates some major aspects of HPF and briefly notes presence or absence of the feature in each of the compilers, or indicates restricted, related, or augmented facilities corresponding to the HPF feature.

Beyond details of HPF that are handled or not, each compiler has its own manner of handling of FORTRAN 77 and Fortran 90 features that may or may not contribute to the parallel execution of a program. The following table briefly notes some of these capabilities.

Our experience at CTC with HPF indicates that performance tuning (including discovering for which parts of a code the compiler has actually generated effective parallelism) is more laborious than correctness debugging, regardless of the HPF product. Thus, even the nature of a compiler report indicating parallel execution (or not) is an important "tool". The following table indicates some of the facilities, tools, and extensions that bear on those issues (and a few more minor ones).