Silicon Graphics IRIS Workstation, User Manual

Page: 95 / 170

Version 1.0
IRIS WORKSTATION GUIDE 91 Appendix E:
• When a float is passed as an argument to a function, it must be
extended to double.
• An arithmetic operation performed on float ’s is done by first
extending the operands to double ’s. After the operation is
performed, the result is truncated to float .
The implementation of floating point in C on the IRIS Workstation defines the
C types double and float to be the same precision - that of single precision. This
decision was made on the following basis:
• Many of the operations performed on the IRIS Workstation provide
results destined for the Geometry Engine pipeline. Since the
precision of Geometry Engine format floating point is somewhat less
than that of IEEE floating point single precision, it is inefficient to
perform all floating point operations in double precision.
• There are many calculations in which double precision is not
needed.
• Since conversion between single and double precision is expensive
in IEEE format, passing float ’s between routines is inefficient if they
must be extended to double precision. This problem is amplified
when the receiving routine does not want the added precision and
converts the incoming value back to float .
All floating point numbers that are declared as either double or float will be
single precision on the IRIS Workstation. All operations on them will also be
single precision. This includes most calls to the standard set of math routines -
sin , cos , sqrt , etc. See Section 3 of the UNIX Programmer’s Manual for an exact list.
The special type combination long float can be used to specify double precision
where it is needed. This implementation differs from the C standard since the
types long float and double are not synonymous.
NOTE: to write portable code with double-precision floating point, use the
type
long float
instead of
double
.
This implementation has several nice properties:
• Current C programs that use float ’s and double ’s will compile without
alteration, although double will be single precision.
• C programs will execute correctly under other implementations,
since long float is usually synonymous with double , so long as double ’s
and long float ’s are not intermixed.
• Single-precision math routines ( exp , sin , etc.) can now be used.
These offer significant speed improvements over the double-
precision versions for those applications that do not require the extra
precision.