FSSpMDM - Accelerating Small Sparse Matrix Multiplications by Run-Time Code Generation

FSSpMDM - Accelerating Small Sparse Matrix Multiplications by Run-Time Code Generation

Explore a seminar on accelerating small sparse matrix multiplications through run-time code generation in this FEM@LLNL presentation. Delve into the strategies employed by the Fixed Size Sparse Matrix-Dense Matrix (FSSpMDM) routine in libxsmm for generating performant operator kernels for prismatic and hexahedral elements. Examine code generation techniques for both x86-64 and AARCH64 instruction sets, and compare results on recent Intel and Apple CPUs against the GiMMiK C code generation library. Gain insights into optimizing high-order finite element method solvers and enhancing performance in small matrix multiplications crucial for modern computational applications.