Floating-point error analysis and mixed-precision optimization are critical for ensuring correctness and reducing costs in High-Performance Computing. In this joint presentation, we introduce Verificarlo and Verrou, two interoperable tools dedicated to the verification, and optimization of numerical codes. Sharing specialized numerical backends through the ANR Interflop project, both frameworks evaluate code stability without modifying the source code.

Using perturbed floating-point paradigms like stochastic rounding, these tools pinpoint numerical bugs and probabilistically estimate accuracy. Additionally, their variable precision (VPREC) backend models lower precision formats to help explore mixed-precision strategies and identify code regions that tolerate reduced precision. Both also support delta-debugging workflows to automatically isolate the exact locations of numerical instabilities.

Their primary distinction lies in instrumentation: Verificarlo is an LLVM framework operating at the compiler level, while Verrou operates dynamically at runtime as a Valgrind tool. This gives developers the choice between compile-time and runtime analysis depending on their needs. We will illustrate these combined capabilities through practical case studies in fields like computational fluid dynamics (CFD), neuro-imaging, and quantum chemistry.