Mixed Precision Tuning with Salsa

Precision tuning consists of finding the least floating-point formats enabling a program to compute some results with an accuracy requirement. In mixed precision, this problem has a huge combinatory since any value may have its own format. Precision tuning has given rise to the development of several tools that aim at guarantying a desired precision on the outputs of programs doing floating-point computations, by minimizing the initial, over-estimated, precision of the inputs and intermediary results. In this article, we present an extension of our tool for numerical accuracy, Salsa, which performs precision tuning. Originally, Salsa is a program transformation tool based on static analysis and which improves the accuracy of floating-point computations. We have extended Salsa with a precision tuning static analysis. We present experimental results showing the efficiency of this new feature as well as the additional gains that we obtain by performing Salsa’s program transformation before the precision tuning analysis. We experiment our tool on a set of programs coming from various domains like embedded systems and numerical analysis