Micro-Optimization of Floating-Point Operations
Abstract
This paper describes micro-optimization, a technique for reducing the operation count and time required to perform floating-point calculations. Micro optimization involves breaking floating-point operations into their constituent micro-operations and optimizing the resulting code. Exposing micro-operations to the compiler creates many opportunities for optimization. Redundant normalization operations can be eliminated or combined. Also, scheduling micro- operations separately results allows dependent operations to be partially overlapped. A prototype expression compiler has been written to evaluate a number of micro-optimizations. On a set of benchmark expressions operation count is reduced by 33% and execution time is reduced by 40%.
Document Details
- Document Type
- Technical Report
- Publication Date
- Aug 01, 1988
- Accession Number
- ADA202001
Entities
People
- Bill Dally
Organizations
- Massachusetts Institute of Technology