This is embodied in 2 classes of operations:
-
INTRINSICS - 74 library routines
-
these would be coded in most efficient fashion possible for target machine - whether global address or message passing, eg., ALL, ANY, CSHIFT, SUM, ...
-
represent a set of primitives higher level than message passing which implement functionality needed in most (any) parallel processing systems
-
PARALLEL STATEMENTS - FORALL / array assignments, INDEPENDENT
-
here compile time analysis leads to efficient parallel implementation
-
HPF will again invoke a set of runtime library routines
-
need most efficient, not most convenient routines
|
