Software Aspects of IEEE Floating-Point Computations for Numerical Applications in High Energy Physics

About

Floating-point computations are at the heart of much of the computing done in high energy physics. The correctness, speed and accuracy of these computations are of paramount importance. The lack of any of these characteristics can mean the difference between new, exciting physics and an embarrassing correction.

This talk will examine practical aspects of IEEE 754-2008 floating-point arithmetic as encountered in HEP applications. After describing the basic features of IEEE floating-point arithmetic, the presentation will cover:

common hardware implementations (SSE, x87)

techniques for improving the accuracy of summation, multiplication and data interchange

compiler options for gcc and icc affecting floating-point operations

hazards to be avoided

About the speaker

Jeffrey M Arnold is a Senior Software Engineer in the Intel Compiler and Languages group at Intel Corporation. He has been part of the Digital->Compaq->Intel compiler organization for nearly 20 years; part of that time, he worked on both low- and high-level math libraries. Prior to that, he was in the VMS Engineering organization at Digital Equipment Corporation. In the late 1980s, Jeff spent 2½ years at CERN as part of the CERN/Digital Joint Project. In 2008, he returned to CERN to spent 10 weeks working with CERN/openlab. Since that time, he has returned to CERN multiple times to teach at openlab workshops and consult with various LHC experiments. Jeff received his Ph.D. in physics from Case Western Reserve University.

.

0:00:00 Slide 1

0:03:17 Slide 2

0:06:33 Slide 3