InterFLOP, Interoperable Tools for Computing, Debugging, Validation and Optimization of Floating-Point Programs
Only 127 seats left
Wednesday, June 30, 2021 3:00 PM to 4:00 PM · 1 hr. (Africa/Abidjan)
Information
Contributors:
Abstract:
The InterFLOP project aims at providing a modular and scalable platform to analyze and control the costs of FP behavior of today’s real programs facing new paradigms (bigger problems, new architectures, new representation formats). The results of existing tools often generate new questions such as: Is this small error guaranteed? May this numerical error occur and how to reduce it? Some more costly analyses bring pieces of answers to these questions, but also require more expertise. In InterFLOP, we propose new analyses and combinations of existing ones to address the challenge of providing a quick and precise numerical diagnosis requiring little user expertise. For that, InterFLOP will collect and combine information on numerical instabilities, catastrophic cancellations, unstable tests, build various statistical analyses of program executions at minimal overhead. InterFLOP capitalizes on the partners’ expertise of UPVD (efficient propagation of accuracy during execution with FP-ANR), of LIP6 (stochastic Cadna analysis with C, C++, Fortran front-ends), of EDF, UVSQ, Intel (Verificarlo, Verrou analysis with executable and llvm front-ends), of CEA LIST (guaranteed analysis by abstract interpretation with Fluctuat affine forms and with C, C++ front-ends), of ANEO (analysis of executable running on Windows/Linux servers). InterFLOP will 1)enlarge the class of possible applications by considering new front-ends and therefore new analyses 2)provide finer numerical analyses based on formalized composite analyses 3)verify the accuracy in the context of precision auto-tuning to make applications more efficient and robust 4)build statistical analyses tools and help the developer interpret the numerical behavior of program through graphical interpretation.Visit the Project Website
- Pablo De Oliveira Castro (UVSQ)
- François Fevotte (TriScale Innov)
- Stef Graillat (Sorbonne University)
- Fabienne Jezequel (Sorbonne University)
- Wilfried Kirschenmann (ANEO)
- Jean-Luc Lamotte (Sorbonne University)
- Bruno Lathuiliere (EDF)
- Yves Lhuilier (CEA)
- Eric Petit (Intel)
- Julien Signoles (CEA)
- Devan Sohier (UVSQ)
- Franck Vedrine (CEA)
- David Defour (University of Perpgnan)
Abstract:
The InterFLOP project aims at providing a modular and scalable platform to analyze and control the costs of FP behavior of today’s real programs facing new paradigms (bigger problems, new architectures, new representation formats). The results of existing tools often generate new questions such as: Is this small error guaranteed? May this numerical error occur and how to reduce it? Some more costly analyses bring pieces of answers to these questions, but also require more expertise. In InterFLOP, we propose new analyses and combinations of existing ones to address the challenge of providing a quick and precise numerical diagnosis requiring little user expertise. For that, InterFLOP will collect and combine information on numerical instabilities, catastrophic cancellations, unstable tests, build various statistical analyses of program executions at minimal overhead. InterFLOP capitalizes on the partners’ expertise of UPVD (efficient propagation of accuracy during execution with FP-ANR), of LIP6 (stochastic Cadna analysis with C, C++, Fortran front-ends), of EDF, UVSQ, Intel (Verificarlo, Verrou analysis with executable and llvm front-ends), of CEA LIST (guaranteed analysis by abstract interpretation with Fluctuat affine forms and with C, C++ front-ends), of ANEO (analysis of executable running on Windows/Linux servers). InterFLOP will 1)enlarge the class of possible applications by considering new front-ends and therefore new analyses 2)provide finer numerical analyses based on formalized composite analyses 3)verify the accuracy in the context of precision auto-tuning to make applications more efficient and robust 4)build statistical analyses tools and help the developer interpret the numerical behavior of program through graphical interpretation.Visit the Project Website
