HPF and Data Parallelism

In an HPF program we:

• define conceptual processor grids,
• distribute data onto processors,
• perform calculations,

The conceptual processor grids do not have to bear any relation to the underlying physical processor configurations -- these are merely the way in which the HPF program views the system.

Each of the executable mechanisms in the following list specify some form of parallelism. The compiler will spot when constructs such as these are being used and can the parallelise the code at this point:

• Fortran 90 contains array syntax and intrinsics to express parallelism
• FORALL is a generalised array assignment statement (and is in Fortran 95)
• INDEPENDENT loops are DO loops which have the property that the order in which the iterations are executed is unimportant. This means that they can all be executed at the same time.
• PURE procedures are `side-effect free' procedures. They can be called in the body of FORALL and it is clear that one instance of a PURE procedure will not interfere with another.
• EXTRINSIC s are procedures written in a different language (ie non-HPF)
• The HPF Library module and the new intrinsics are comparable to the Fortran 90 intrinsics.