MATLAB codes for scaling a matrix given in fp64 or fp32 into fp16, and using the scaled matrix to solve system of linear equations using GMRES based iterative refinement. A modified version of GMRES-based iterative refinement has been used in these codes. The orignal codes for GMRES-based iterative refinement can be downloaded from https://github.com/eccarson.
- N. J. Higham, S. Pranesh, M. Zounon. Squeezing a Matrix Into Half Precision, with an Application to Solving Linear Systems
- MAIN.m Generates all the tables in the manuscript into the folder results/.
- The codes have been developed and tested with MATLAB 2018a.
- This code requires Cleve Laboratory to perform half precision computations and Advanpix Multiprecision Computing Toolbox for extended precision computations. A free trial of Advanpix is available for download from https://www.advanpix.com/.
- Srikara Pranesh
- Nicholas. J. Higham