Data Driven Calculations Histories to Minimize IEEE-755 Floating-point Computational Error
The widely implemented and used IEEE-754 Floating-point specification defines a method by which floating-point values may be represented in fixed-width storage. This fixed-width storage does not allow the exact value of all rational values to be stored. While this is an accepted limitation of using...
| Main Author: | |
|---|---|
| Format: | Others |
| Published: |
NSUWorks
2004
|
| Subjects: | |
| Online Access: | http://nsuworks.nova.edu/gscis_etd/830 http://nsuworks.nova.edu/cgi/viewcontent.cgi?article=1829&context=gscis_etd |
| Summary: | The widely implemented and used IEEE-754 Floating-point specification defines a method by which floating-point values may be represented in fixed-width storage. This fixed-width storage does not allow the exact value of all rational values to be stored. While this is an accepted limitation of using the IEEE-754 specification, this problem is compounded when non-exact values are used to compute other values. Attempts to manage this problem have been limited to software implementations that require special programming at the source code level. While this approach works, the problem coder must be aware of the software and explicitly write high-level code specifically referencing it. The entirety of a calculation is not available to the special software so optimum results cannot always be obtained when the range of operand values is large. This dissertation proposes and implements an architecture that uses integer algorithms to minimize precision loss in complex floating-point calculations. This is done using runtime calculation operand values at a simulated hardware level. These calculations are coded in a high-level language such that the coder is not knowledgeable about the details of how the calculation is performed. |
|---|