Portable vectorization with numerical accuracy control for multi-precision simulation codes

The performance of simulators has a direct impact both on the quality of the simulation results and on the study of a wide variety of scientific hypotheses.
Modern parallel computing resources are based on complex hardware architectures with several levels of parallelism involving SIMD computing units. This level of parallelism is crucial since it significantly increases the number of floating-point operations per second. In this thesis, we focus on this level of parallelism by trying to attain the optimal performance of the computation kernels used in our applications. 
The difficulties related to these computational units are multiple. Indeed, we are faced with different hardware architectures which are constantly evolving. Each architecture has different instruction sets where their optimal use depends on the size of the vector, the used type, and the required accuracy. For a given computation these instruction sets could be written in multiple ways, with different accuracies. Moreover, their usage is non-trivial for non-specialist developers. All these difficulties make their usage hard in multi-physique simulators, where their lifespan is longer than a target architecture.
Based on this observation, the objective of the proposed research work is to study and design abstraction techniques that provide an accurate SIMD intrinsic without accuracy lost for different hardware architecture such as Intel/AMD x86, ARM, and RISC-V.

Keywords: Abstraction, SIMD, code generation, variable precision, modern computer architecture

• Academic supervisor    Pr. Olivier AUMAGE, Laboratoire LaBRI
• Doctoral School    École Doctorale Mathématiques et Informatique (EDMI), Université de Bordeaux
• IFPEN supervisor    Dr. GUIGNON Thomas, Département Informatique scientifique, thomas.guignon@ifpen.fr
• PhD location    Laboratoire LaBRI, Bordeaux, France / IFPEN, Rueil Malmaison, France
• Duration and start date    3 years, starting in fourth quarter 2022
• Employer    IFPEN, Rueil Malmaison, France
• Academic requirements    University Master degree in computer science
• Language requirements    Fluency in French or English, willingness to learn French
• Other requirements    SIMD programming, code optimization techniques, computer architecture