An adaptation of the MWP-CWP model for a GPU architecture applied to 3-D stencil kernels

Finding the bottlenecks in the execution of a kernel in a GPU is essential to improve the performance of the implementation. Although there are several expertise techniques such as Assess, Parallelize, Optimize, Deploy (APOD), proposed by NVIDIA, the use of those techniques in computationally expensive algorithms such as Reverse Time Migration (RTM) is not an option. To solve this problem several models had been proposed for general and specific algorithms. In this paper, we present an adaptation of the Memory Warp Parallelism-Computation Warp Parallelism (MWP-CWP) model to estimate the execution time of 3D-stencil-based kernels used in the RTM algorithm implementation. The most important step in the adaptation of the model for the proposed kernel is the correct identification and extraction of the parameters required by the model.