Automated empirical optimization of high performance floating point kernels

Abstract

Using traditional methodologies and tools, the problem of keeping performance-critical kernels at high efficiency on hardware evolving at the incredible rates dictated by Moore's Law is almost intractable. On product lines where ISA compatibility is maintained through several generations of architecture, the growing gap between the machine that the software sees and the actual hardware exacerbates this problem considerably, as do the evolving software layers between the application in question and the ISA. To address this problem, we have utilized a relatively new technique, which we call AEOS (Automated Empirical Optimization of Software). In this paper, we describe the AEOS systems we have researched, implemented and tested. The first of these is ATLAS (Automatically Tuned Linear Algebra Software), which empirically optimizes key linear algebra kernels to arbitrary cache-based machines. Our latest research effort is instantiated in the iFKO (iterative Floating Point Kernel Optimizer) project, whose aim is to perform empirical optimization of relatively arbitrary kernels using a low-level iterative and empirical compilation framework.