EUPEX webinar – Vectorization of FEM kernels​ (ESPRESO FEM)

The EUPEX researchers share their results and lessons learnt!

Webinar

In collaboration with

11am to 12pm CET

We are organising monthly webinars about the outcomes of EUPEX (full programme here). Our fifth webinar will be dedicated to one of the many applications studied and enhanced under EUPEX: ESPRESO FEM

Registration>>

Abstract: Finite Element Methods (FEM) are widely used for numerical simulations, relying on local matrix kernels to transform physical parameters into systems of linear equations. Optimizing these kernels is crucial for fully leveraging modern hardware architectures, particularly through vectorization (using SIMD instructions). However, despite the quality of current compilers, vectorization is still challenging even for suitable applications, as the quality of vectorization is highly dependent on tiny details in the source code.

In this webinar, we describe a straightforward code modification utilizing:

  • Cross-element vectorization for the simple transformation of the original scalar code to a vectorized one
  •  Meta-programming leveraging compiler capabilities to build sub-kernels tailored for a particular set of parameters
  • Run-time selection of kernels depending on a user setting or a current computation phase

We will demonstrate how these techniques, combined with appropriate data structures, enable the development of highly efficient kernels with near-optimal performance across a wide range of modern hardware architectures all with minimal effort.

Speaker: Ondrej Meca, IT4I 

Ondřej Meca holds a PhD degree in Computer Science from VSB—Technical University of Ostrava, Czech Republic. He is currently a member of the Infrastructure Research Lab at IT4Innovations National Supercomputing Center. His research interests include verifying parallel algorithms, developing pre/post-processing algorithms for large-scale engineering problems, and developing highly scalable linear solvers.

Registration>>

Date

March

26

Location

Zoom

Share this article

Contact us