The Case for OpenMP in GCC

[Scott Robert Ladd <coyote at coyotegulch dot com>]

Hello!

In recent discussions of 3.5 and 4.0, I haven't addressed why certain 
GCC enhancements may be desirable to both the user (i.e., someone who 
compiles code with GCC) and developer (i.e., someone who works on GCC 
itself) communities.

OpenMP (http://www.openmp.org) is one such enhancement.

The future of programming is clearly parallel. All of the major hardware 
vendors -- IBM, AMD, Intel, and Sun -- have announced multicore 
processors. Dual processor workstations, once only found in "scientific 
workstations", have now found their way into digital signal processing 
and data analysis applications. New supercomputer clusters are being 
constructed from multiprocessor commodity hardware.

Note that I am writing in terms of SMP and NUMA systems, and not about 
cluster applications. Yet even cluster-oriented code (built on MPI or 
PVM) may encounter the need for local parallelization, when a cluster is 
constructed from multiprocessor workstations.

Writing parallel code isn't simple; this is especially true if one wants 
to create *portable* parallel code. It's possible to write a parallel 
program using Windows or POSIX threads -- but doing so does not result 
in cleanly readable or portable code.

Enter OpenMP, an industry standard for parallelization in C, C++, and 
Fortran. OpenMP employs pragmas (in C and C++) or special comments 
(Fortran) to implement a parallel constructs. A single OpenMP directive 
can parallelize a loop, dramatically improving application performance. 
Several OpenMP directives, with appropriate arguments, can handle very 
complex parallel tasks.

OpenMP is scalable, automatically determining the number of threads to 
be used based on the underlying hardware architecture and environment 
settings. The syntax is simple, and OpenMP can be completely ignored 
when compiled for a serial execution environment.

You'll find OpenMP implementations in compilers produced by Intel, SGI, 
PGI, and Lahey. In my experience, most modern Fortran programmers expect 
OpenMP to be built into their compiler -- and implementing OpenMP would 
provide the tools required for automatic parallelization, a feature 
found in High Performance Fortran compilers.

GOMP -- the GNU OpenMP project -- is an effort to implement OpenMP in 
the tree-ssa (GCC 3.5/4.0) framework. The project is largely stalled due 
to a lack of funding and time on the part of its participants. While 
several of us have created or are working on design and library issues, 
this effort less likely to succeed without the broader support of the 
GCC community.

OpenMP support is, I believe, essential for GCC to be recognized as a 
credible compiler in certain realms, and the inclusion of OpenMP will 
benefit free software by making parallel development easier and more 
portable.


--
Scott Robert Ladd
Coyote Gulch Productions (http://www.coyotegulch.com)
Software Invention for High-Performance Computing