Migrating and Tuning a CUDA-Based Stencil Computation to DPC++


Reverse Time Migration takes advantage of the finite-difference method (FD) to perform the numerical approximations for the acoustic wave equation. Stencil computation applied to this numerical method represents a computational bottleneck when developing RTM applications, and therefore needs to be optimized to guarantee timely results and efficiency when allocating resources for hydrocarbon exploration. This article presents a migration experience from a CUDA-based RTM source code to Data-Parallel C++ (DPC++) using Intel DPC++ Compatibility Tool (DPCT). Migrated source code is proven to be easier to maintain, as it unifies the algorithm execution flow for our application in a unique structure. Besides migration experience, we could easily explore memory management, achieving twice the performance and an arithmetic intensity 4,9x higher than the first version.


Clicia Pinto, Performance Engineer, SENAI CIMATEC

Clícia S. Pinto is a Senior Performance Engineer at the Supercomputing Center for Industrial Innovation of SENAI CIMATEC focusing on algorithm development for scientific computing and code optimization applied to industry and academia. She holds a Ph.D. in Computer Science from the Federal University of Bahia and her primary research interests include High-Performance Computing and Data-intensive Computing.