Redesigning The Absorbing Boundary Algorithm for Asynchronous High Performance Acoustic Wave Propagation
TimeThursday, November 15th8:30am - 5pm
DescriptionExploiting high concurrency, relaxing the synchrony of existing algorithms, and increasing data reuse have immense effect in performance. We integrate the Multicore-optimized Wavefront Diamond (MWD) tiling approach by Malas et al. [SIAM SISC, 2015, ACM Trans. Parallel Comput. 2017], which takes into account the three aforementioned ingredients, into the industrial project codenamed ExaWave framework beside the traditional spatial blocking (SB) technique for stencil computations. However, the fine-grained asynchronous handling of the Convolution Perfectly Matched Layer (CPML) for absorbing boundary conditions turns out to be a challenging open research problem, due to severe inherent data dependencies constraints, which impedes MWD performance impact. We propose techniques of loop fusion to reduce memory traffic and sliding windows to cut down the engendered extra flops, in order to consolidate CPML integration with the overall asynchronous MWD technique. The experimental results on Intel's latest processors show the effectiveness of the proposed techniques.