HPF is an extension to the new standard Fortran (Fortran 90). It makes use of the fact that the array handling features of Fortran 90 are ideal for expressing parallel array operations. It also provides ways of describing parallel loops. The extensions to Fortran 90 include:

• Compiler directives: to guide the compiler to distribute the data in a given array across the available processors, or to label parallel DO loops
• A FORALL statement and construct to complement the Fortran 90 WHERE - both are parallel constructs in HPF
• New intrinsic and library procedures: to extend the already rich set in Fortran 90
• Subset HPF is a version of High Performance Fortran which omits from Full HPF some Fortran 90 features plus those compiler directives related to dynamic re-distribution of data at runtime. It remains much more powerful than FORTRAN 77

An HPF program should run effectively on a wide range of parallel machines, including distributed memory machines. On these, programmers have become used to writing "message passing" programs; and resigned to the difficulties of doing this. HPF programs contain no message-passing at all (though the compiler will inevitably insert some into the final code). An HPF program is as easy to write as a Fortran 90 program; and as easy to read too. So HPF provides a much more friendly programmming paradigm for users, who want their results fast, without needing to worry about details of the parallelism.

The HPF approach is based on two fundamental observations. First, the overall performance of a program can be increased if operations are performed concurrently by multiple processors. Second, the efficiency of a single processor is highest if the professor performs computations on data elements stored locally. Based on these observations, the HPF extensions to Fortram 90 provide a means for the explicit expression of parallelism and data mapping. An HPF programmer can express parallelism explicitly, and using this information the compiler may be able to tune data distribution accordingly to control load balancing and minimize communication.

An HPF program has essentially the same structure as a Fortran 90 program, but is enhanced with data distribution and alignment directives. When writing a program in HPF, the programmer specifies computations in a global data space. Array objects are aligned to abstract arrays with no data, called "templates". Templates are distributed according to distribution directives in "block" and "cyclic" fashions. All arrays aligned to such templates are implicitly distributed.

Parallelism can be explicitly expressed in HPF using the following language features:

• Fortran 90 array assignments
• masked array assignements
• WHERE statements and constructs
• HPF FORALL statemnets and constructs
• HPF INDEPENDENT directives
• intrinsic functions
• the HPF library
• EXTRINSIC functions

The compiler uses these parallel constructs and distribution information to produce code suitable for parallel execution on each node of a parallel computer.