RAPTOR: Practical Numerical Profiling of Scientific Applications

TL;DR

RAPTOR is a practical profiling tool that helps scientists identify which parts of their scientific code can safely use lower-precision computations, improving performance without sacrificing accuracy.

Contribution

It introduces a novel LLVM-based tool that transparently replaces high-precision computations with low-precision ones for numerical profiling in scientific applications.

Findings

Successfully profiled four real-world multi-physics applications.

Identified code regions suitable for precision lowering.

Demonstrated ease of use and effectiveness of RAPTOR.

Abstract

The proliferation of low-precision units in modern high-performance architectures increasingly burdens domain scientists. Historically, the choice in HPC was easy: can we get away with 32 bit floating-point operations and lower bandwidth requirements, or is FP64 necessary? Driven by Artificial Intelligence, vendors introduce novel low-precision units for vector and tensor operations, and FP64 capabilities stagnate or are reduced. This forces scientists to re-evaluate their codes, but a trivial search-and-replace approach to go from FP64 to FP16 will not suffice. We introduce RAPTOR: a numerical profiling tool to guide scientists in their search for code regions where precision lowering is feasible. Using LLVM, we transparently replace high-precision computations using low-precision units, or emulate a user-defined precision. RAPTOR is a novel, feature-rich approach -- with focus on ease of use -- to change, profile, and reason about numerical requirements and instabilities, which we demonstrate with four real-world multi-physics Flash-X applications.