EETimes

Embedded Systems September 2000 Vol13_10

Issue link: http://dc.ee.ubm-us.com/i/71837

Contents of this Issue

Navigation

Page 8 of 229

Parity Bit Organized Dissent I n their articl e, "Portable Vector Math Libraries" Oune 2000, p. 104), William Wright and J ames Metzger de cri be an approach to program- ming for portabili ty across multiple platforms and vecto r libraries. They were kind eno ugh to mention tJ1e Vector, Signal, and Image Processing Library (VSIPL) effort. However, their mentio n of this effo rt included some fac tual inaccuracies which we would like to take iliis oppo rtunity to co.-rect. The article concludes its section on VSIPL by stating that "no implementa- tions of the library are widely available." Actually, implementations of tl1e library are avai lable on embedded multicom- puter hardware fi·om Mercury, SKY, and CSP. MPI Software Technology has an implementation available fo r the PowerPC G3 and G4 processors that is also disnibuted with hardware from DY 4 Systems. Links to the Web sites for these vendors can be obtained from tl1e "users and products" section of the VSIPL Web site ( wurw. vsijll.org). In addition, an open source version of the library is available for download from the site. This library, written by Randy Judd of the U.S. Navy Space and Naval Warfare Systems com- mand, hould compile on most plat- forms that have an ANSI C compiler. The open VSIPL tandard and multiple commercial implementations free ilie user from having to develop his own API, and then to update it for each new target architecture. Using VSIPL has seve ral advantages over the approach advocated by the autl10rs. VSIPL has a well-defin ed , portable in te rface, designed with sub- stantial input from maj or embedded and workstation computing vendors and users, and supported on all the platfo• ·ms listed above. Regression test- ing of ilie library is supported directly by tlle freely available VSIPL test suite. As th e authors correcLly poin t o ut, individuals definin g their own API are required to write and support these tests tJ1emselves. Memory alignment issues fo r vector librari es are handled by the VSIPL implementation witho ut inte• of complex data types, the approach advocated by the authors will work as long as the library expects the real and imaginary portions of the data to be stored in an alternating (inte rleaved ) format. However, VSIPL also supports a split fo• ·mat fo r complex data, which is advantageous on vecto r processors such as tl1e Motorola G4 PowerPC. The authors state that a maj o r advantage of designing your own vector library API is that one can "choose tlle sacrifices that are approp1iate in terms of execution speed and memory usage to achieve portabili ty." The VSIPL stan- dard supports such tra.de-offs by parti- tioning tlle functionali ty of the library into well-defined profiles. The Core- Lite profile is focused on vector signal processing and includes FIT and FIR filter operations. The full Core profi le includes matrix algebra operations for more complex signal processing requirements. Users can write their programs using tl1e functionality pro- file tl1at best sui ts Ll1ei•· needs. In addi- tion, the participation of multiple ven- do rs in defining the standard has resulted in a design tJ1at does much to ensure implementors can offer high performance VSIPL libraries o n a vari- ety of a•·chitectures. James Lebak jlebak@ l l.mit. edu Mark Richards mar k. ri c ha rds@gtri. ga tee h .e el u Randy Janka ra nda l l .ja n ka @g tri .ga tech. e d u CO-C H A IR S, VSIPL FORUM WHAT'S ON YOUR MIND? Embedded Systems Programming gladly welcomes feedback from readers. Please send any question or comments to Editor in Chief Michael Barr at mbarr(

Articles in this issue

Archives of this issue

view archives of EETimes - Embedded Systems September 2000 Vol13_10