HPFC at a glance

Input language

• Fortran 77
• HPF static and dynamic mapping directives
• prescriptive and descriptive mappings for arguments
• HPF independent and new directives
• private local, pure and io directives
• assumes all sizes as known at compile time
• arrays in commons can be distributed statically

Output

• 2 programs: host for I/Os and SPMD node for computations
• Fortran 77 + HPFC Runtime support + PVM 3

Implementation

• data structures based on NEWGEN. Linear/C3 library for linear techniques.
• 18,000 lines of ANSI C for the compiler (HPFC, a PIPS module)
• 1,000 lines of shell and C for the driver and other utilities
• 3,000 lines of m4/fortran for the runtime support library (expanded to +8,000)
• PIPS as a whole is around 200,000 lines of code.

Standard Optimizations

• run-time resolution to handle general cases.
• reduced allocation of distributed arrays on nodes.
• overlaps, partial shifts and reductions for block-distributed stride-1 aligned arrays.
• simple copies for any aligned arrays are recognized.

Advanced Optimizations

• for I/O-related communications (abstract, paper)
• for array remappings (abstract, paper)
• and more linear algebra-based techniques to come (abstract, paper).

Experiments

• +100 small or medium size codes for non regression test.
• SOR on a network of SUN workstations (abstract, paper).
• wave propagation code from IFP (French Oil Institute) on the SEH/ETCA CM5.
• matrix transpose on a DEC Alpha farm (see remappings).