.#
.#
.#
.ch "Introduction"
Users of Prime computers have been aware for some time of a number of
shortcomings in the floating point arithmetic firmware.
In addition, there have been a number of inaccuracies found in
the standard math libraries which have gone uncorrected for years
({1}, {2}).  Unlike other major computer firms, Prime
has not published any documents dealing with the algorithms or
error analysis of their math routines.
.pp
In the winter of 1982 I undertook the coding of a new math library
to support the Georgia Tech SWT Pascal compiler, and the Georgia Tech
C compiler.  The results of tests on that library and the standard Prime
libraries have revealed a number of interesting facts.  Additionally,
further experimentation with the floating point mechanisms has revealed
some bugs in the way arithmetic is performed, in some cases.
.pp
First, this guide describes the architecture of the floating point mechanism,
including some error analysis and description of quirks in the
hardware.  This includes a description of incompatibilities between
the 400/550 cpu and the 750/850 cpu floating point register structure.
Next is a description of the SWT Math library. Last is a discussion
of some preliminary error analysis of the SWT library and the Prime
standard library functions.  The appendices contain information on
auxiliary programs supplied with the library which will aid
users in writing their own routines, and checking existing routines
and floating point firmware.
.#
.#
.mh Acknowledgements
I would like to thank Roy Mongiovi for his help in debugging
some of the SWT Math routines, and Peter Wan for his help in
preparing this guide.  I would also like to thank Ann Vitale, Ron Kurtzer, and
especially Emory Stevens of the Atlanta Prime office for their
co-operation and aid in the testing of these routines.
.pp
Research contributing to the development of this report was conducted
while the author was receiving a National Science Foundation Graduate
Fellowship, support which is gratefully acknowledged.