Project deliverable

D3.5 – Report on the reduction of numerical precision for computation and I/O

  • 27 Sep 2023
  • .

This deliverable adresses the critical concern of floating-point accuracy in numerical simulations and computation-intensive codes, emphasizing the need for early detection and resolution of numerical bugs. It introduces Verificarlo-CI, a continuous integration workflow that monitors and optimizes numerical accuracy during code development within QMCkl. It also details the design of optimized versions of Sherman-Morrison-Woodbury (SMWB) kernels and show they achieve a good tradeoff between accuracy and performance. The report also explores the benefits of mixed precision in the CHAMP code. Overall, it underscores the importance of early numerical accuracy assessment and offers practical solutions for tuning code performance and accuracy.