Tutorial on Floating-Point Analysis Tools

Navigating Mixed-Precision: Tools and Techniques for Floating-Point Error Profiling

SC26, Chicago

November 16, 2026

Time: 1:30pm - 5:00pm CST (half-day tutorial)
Location: McCormick Place, Chicago, Room: W184b-c

Description

Floating-point arithmetic is the bedrock of scientific computing, yet it frequently introduces subtle errors and non-deterministic behavior in numerical software. As the HPC community shifts toward mixed-precision execution to unlock performance and energy-efficiency gains, developers require robust methodologies to rigorously analyze floating-point error accumulation.

This tutorial provides a comprehensive introduction to two state-of-the-art tools designed to guide precision-tuning decisions in both CPU and GPU-accelerated applications. The first tool is FPChecker, which leverages the Clang/LLVM compiler infrastructure to perform static and dynamic analysis of C/C++ applications. The second is NixNaN, a high-performance profiler built on the NVIDIA NVBit dynamic instrumentation framework, engineered to detect floating-point exceptional values arising during CUDA kernel execution without requiring source modification or recompilation.

Throughout this session, participants will learn:

  1. to assess the floating-point dynamic range of applications,
  2. visualize the propagation of rounding errors, and
  3. isolate performance-critical numerical issues, such as overflows, and NaNs.

The tutorial emphasizes a hands-on approach: attendees will engage with accessible examples (e.g., simple linear solvers) configured to run locally on their laptops. In the advanced session, we transition to real-world impact, where participants will apply these tools to larger numerical libraries (e.g., Hypre).

Presenter

AWS Instances

https://fpchecker.org/usernames

Presentation Slides

The tutorial slides are here: slides

Repositories: