User's Guide

Maximizing Floating-Point Performance

If performance is your primary concern, and you want your program to be relatively safe but do not mind if results are slightly different (generally more precise) from what they would be otherwise, optimize the program with the -O option and specify -qfloat=rsqrt:hssngl:fltint.

The following section describes the functions of these suboptions:

• The rsqrt suboption replaces division by a square root with multiplication by the reciprocal of the root, a faster operation that may not produce precisely the same result.

• The hssngl suboption is the opposite of rndsngl: it improves the performance of single-precision (REAL(4)) floating-point calculations by suppressing rounding operations that are required by the Fortran language, but are not necessary for correct program execution. The results of floating-point expressions are kept in double precision where the original program would round them to single-precision. These results are then used in later expressions instead of the rounded results.

  Rounding operations are still inserted when double-precision results are stored into single-precision memory locations, to detect single-precision floating-point overflows and underflows. However, if optimization removes such a store operation, hssngl also removes the corresponding rounding operation, possibly preventing the exception. (Depending on the characteristics of your program, you may or may not care whether the exception happens.)

  The hssngl suboption is safe for all types of programs because it always only increases the precision of floating-point calculations. Program results may differ because of the increased precision and because of avoidance of some exceptions.

• The fltint suboption speeds up float-to-integer conversions by reducing error checking for overflows. You should make sure that any floats that are converted to integers are not outside the range of the corresponding integer types.

In cases where speed is so important that you can make an informed decision to sacrifice correctness at boundary conditions, you can replace hssngl and fltint with the hsflt suboption: it does the same thing as fltint, and also suppresses rounding operations.

In suppressing rounding operations, hsflt works like hssngl, but it also suppresses rounding operations when double-precision values are assigned to single-precision memory locations. Single-precision overflow is not detected in such assignments, and the assigned value is not correctly rounded according to the current rounding mode.

Attention: When you use the hsflt suboption, observe these restrictions or your program may produce incorrect results without warning:

• Your program must never convert overly large floats to integer.
• Your program must never compute NaNs, INFs, or values outside the range of single precision.
• Your program must not depend on results to be correctly rounded to single precision, for example by comparing two single-precision values for equality.

Therefore, we recommend that you use this suboption only with extreme caution. It is for use by knowledgeable programmers in specific applications, such as graphics programs, where the computational characteristics are known. If you are at all unsure whether a program is suitable, or if the program produces unexpected results when you use this suboption, use hssngl instead.

Related Information:

Technical Details of the -qfloat=hsflt Option provides additional technical information about this suboption.