High Performance Fortran Translator
		   -----------------------------------

                           Dr Bryan Carpenter
                Southampton University Computing Services.


Contents
--------

  Background
       Why HPF?				1 slide
       HPFF				1 slide

  HPF overview                          1 slide
       Parallel computation		n slides
       Distributed distributed data	n slides

  Southampton work
       PUMA				1 slide (1/2 ?)
       ADAPT and ADLIB			1 slide (1/2 ?)
       JISC and SERC projects		1 slide
       Mapping of computation		n slides
       Run-time library			n slides


1. Background
-------------


1.1 Why HPF?
------------

Some extant HPC architectures:

  * Single processor vector
  * Single instruction stream, distributed data (SIMD)
  * Multiple processor (MIMD)
    * Message-passing
    * Remote memory access
    * Shared memory

Standardisation efforts within some subclasses (eg MPI), but still
great diversity.

Parallelisation of sequential languages difficult/intractable.

HPF generalises SIMD dialects of Fortran---less baggage than
message-passing or general shared memory.
``Bulk-synchronous, parallel random access'' model.
Adds some of the flexibility of MIMD model.

Can be compiled to any of the above architectures ;-).


1.2 The High Performance Fortran Forum
--------------------------------------

Started meeting March 1992.  Participation of 40+ organisations,
including designers of Fortran D, Vienna Fortran, and manufacturers of
parallel computers (Cray, DEC, Fujitsu, HP, IBM, Intel, Maspar, Meiko,
nCube, Sun, Thinking Machines...).

Language is a development of SIMD Fortran dialects such as Connection
Machine Fortran, and distributed memory dialects such as Fortran D and
Vienna Fortran.

Standard distributed in May '93, published in September.

Still no full Subset implementations available.


2. HPF Overview
---------------

HPF extends Fortran in two ways:

  * It adds constructs which specify parallel execution of computations.

  * It adds directives which specify how program data is ``aligned''
    and distributed across multiple memories.


2.1 Parallel Computation
------------------------

HPF includes all the ``array syntax'' and ``array intrinsics'' of
Fortran 90.  It adds a FORALL construct.

Examples of Fortran 90 syntax:
\begin{verbatim}
  REAL X (10), Y (10), LEN (10)

  LEN = SQRT(X**2 + Y**2)
\end{verbatim}
All 10 lengths can be computed in parallel.
\begin{verbatim}
  REAL A (10, 10), B (10, 10), C (10, 10)

  DO I = 1, 10
    DO J = 1, 10
      A (I, J) = DOT_PRODUCT (A (I, :), B (:, J))
    END DO
  END DO
\end{verbatim}
The intrinsic {\em DOT_PRODUCT} can be computed by a parallel algorithm
(but {\em DO} loops remain sequential).  Note syntax for ``array
sections''.



By combining the F90 array intrinsics with section subscripts (and
``masking''), possible to express many parallel algorithms.

{\em But}, array syntax rapidly becomes clumsy for writing practical
parallel programs.  Can use {\em FORALL} instead
\begin{verbatim}
  FORALL (I = 1 : N, J = 1 : N)
    U (I, J) = (U (I, J - 1) + U (I, J + 1) +
                U (I - 1, J) + U (I + 1, J)) / 4
  END FORALL
\end{verbatim}
(a parallel iterative solver).
\begin{verbatim}
  FORALL (I = 1 : N, J = 1 : N)
    A (I, J) = DOT_PRODUCT (A (I, :), B (:, J))
  END FORALL
\end{verbatim}
(fully parallel version of earlier example).