Correctness of Floating Point Programs - Exception Handling and Reproducibility
TimeMonday, November 12th2pm - 3pm
DescriptionWe consider two related aspects of analyzing and guaranteeing correctness of floating point programs: exception handling and reproducibility. Exception handling refers to reliable and consistent propagation of errors due to overflow, invalid operations (like sqrt(-1)), convergence failures, etc. Reproducibility refers to getting bitwise reproducible results from multiple runs of the same program, e.g., despite parallelism causing floating point sums to be evaluated in different order with different roundoff errors. We describe the efforts of two standards committees, the Basic Linear Algebra Subprograms (BLAS) Standard, and the IEEE 754 Floating Point Standard, to address these issues, and how these efforts should make it easier to accomplish these goals for higher level applications, such as linear algebra libraries.