Escolha do Ladrilhamento para um Simulador de Ondas Acústicas em GPUs por meio de Aprendizado de Máquina

The simulation of acousitc wave propagation is the kernel for important industrial applications like the Full-Waveform Inversion (FWI) and Reverse-Time Migration (RTM). The kernel solves partial differential equations (PDEs) based on the finite differences method, which can be significantly accelerated with the support of GPUs. One of the main challenges for accelerating this stencil computations on GPUs is to reduce the overhead of memory accesses, and tiling is an important optimization which can accelerate wave propagation kernels. However, deciding the tile sizes for these computations is not a straightforward question, which usually depend upon many architectural and application parameters. In the present work, we employ six machine learning methods for providing recommendations for the sizes of tiles to use. Our best strategy has achieved a improvement coefficient of 1.17 and 1.11 on two GPUs with Turing and Volta architectures.