Thursday, October 07, 2004
Correctly Rounding Math Library for Java
Correctly Rounding Math Library for Java
"The objective of the project was to produce a library of twelve fast math function subroutines, written in the (standard) C programming language. Its main feature is the rounding mode of the functions. Each of the routines produces results that are always exactly equal to the accurate, infinite precision values of the corresponding functions, rounded to the nearest/even machine number. The importance of this is that all routines written for the same function, with this standard, are bit by bit compatible, even if the details of the algorithms differ. This enables the user to achieve, in a most natural way, the Java objective of 'develop once and run everywhere'. The routines were produced for the IEEE 754 standard 'double' format, assuming round to the nearest rounding mode for the 'double' data type. The library is portable (provided the arithmetic and 'double' format are right!) and was tested on IBM RS/6000, IBM PC (Pentium), IBM S/390 with the binary floating point feature, and SUN SPARC computers."
"The objective of the project was to produce a library of twelve fast math function subroutines, written in the (standard) C programming language. Its main feature is the rounding mode of the functions. Each of the routines produces results that are always exactly equal to the accurate, infinite precision values of the corresponding functions, rounded to the nearest/even machine number. The importance of this is that all routines written for the same function, with this standard, are bit by bit compatible, even if the details of the algorithms differ. This enables the user to achieve, in a most natural way, the Java objective of 'develop once and run everywhere'. The routines were produced for the IEEE 754 standard 'double' format, assuming round to the nearest rounding mode for the 'double' data type. The library is portable (provided the arithmetic and 'double' format are right!) and was tested on IBM RS/6000, IBM PC (Pentium), IBM S/390 with the binary floating point feature, and SUN SPARC computers."
0 Comments:
