Choosing mathematical function implementations for speed and accuracy

Standard implementations of functions like sin and exp optimize for accuracy, not speed, because they are intended for general-purpose use. But just like many applications tolerate inaccuracy from cancellation, rounding error, and singularities, many application could also tolerate less-accurate function implementations. This raises an intriguing possibility: speeding up numerical code by using different function implementations. This paper thus introduces OpTuner, an automated tool for selecting the best implementation for each mathematical function call site. OpTuner uses error Taylor series and integer linear programming to compute optimal assignments of 297 function implementations to call sites and presents the user with a speed-accuracy Pareto curve. In a case study on the POV-Ray ray tracer, OpTuner speeds up a critical computation by 2.48x, leading to a whole program speedup of 1.09x with no change in the program output; human efforts result in slower code and lower-quality output. On a broader study of 36 standard benchmarks, OpTuner demonstrates speedups of 2.05x for negligible decreases in accuracy and of up to 5.37x for error-tolerant applications.