Compress to Impress: Redesigning Matrix Computations for Computational Astronomy and Seismic Imaging Applications#
Hatem Ltaief (KAUST)
Abstract#
The talk presents an HPC implementation of Tile Low-Rank Matrix-Vector Multiplication (TLR-MVM) on various hardware systems, including the NEC SX-Aurora Tsubasa vector engine. TLR-MVM is one of the most time-consuming computational kernels for seismic wave-equation-based processing and ground-based computational astronomy applications. TLR-MVM exploits data sparsity of the respective operators and relies on an efficient data layout to saturate memory bandwidth of the underlying hardware architectures. We report preliminary results and show the performance superiority of TLR-MVM against state-of-the-art dense implementations.