1: \begin{abstract}

2: Numerical exceptions, which may be caused by overflow, operations like

3: division by 0 or sqrt($-1$), or convergence failures, are unavoidable in

4: many cases, in particular when software is used on unforeseen and

5: difficult inputs. As more aspects of society become automated, e.g.,

6: self-driving cars, health monitors, and cyber-physical systems more

7: generally, it is becoming increasingly important to design software that

8: is resilient to exceptions, and that responds to them in a consistent

9: way. Consistency is needed to allow users to build higher-level software

10: that is also resilient and consistent (and so on recursively). In this

11: paper we explore the design space of consistent exception handling for

12: the widely used BLAS and LAPACK linear algebra libraries, pointing out a

13: variety of instances of inconsistent exception handling in the current

14: versions, and propose a new design that balances consistency,

15: complexity, ease of use, and performance. Some compromises are needed,

16: because there are preexisting inconsistencies that are outside our

17: control, including in or between existing vendor BLAS implementations,

18: different programming languages, and even compilers for the same

19: programming language. And user requests from our surveys are quite diverse.

20: We also propose our design as a possible model for other

21: numerical software, and welcome comments on our design choices.

22: \end{abstract}

23: