Paralelização do algoritmo parametrizações progressivas para arquiteturas multicore

Abstract

Platforms with multicore processors have become common in the past few years, disseminating parallel computing. With the popularization of GPUs, it has become increasingly common to find computers with a graphic card, enabling heterogeneous processing When demanded. By using methods of communication and synchronization between the processors of the heterogeneous platform, it is possible to achieve performance gains previously feasible only in clusters and grids. Each processor within the heterogeneous platform behaves in a way, therefore, balancing between different types of processing loads is essential. One problem that can be solved with this type of computation is the parameterization of triangular meshes. The parameterization process aims to create a bijection between two surfaces with an equivalence of points to solve several problems of the computer graphics. The most recent parameterization algorithm is the algorithm of progressive parameterization. This algorithm has as a principle the rapid convergence for the ideal parameterization result. This work aims to explore the parallel potential of the algorithm in heterogeneous platforms using CPU and GPU. The implementation of the parallel algorithm in the CPU was performed using Intel TBB and on the GPU using the libraries cuSolver, ArrayFire, MAGMA, ViennaCL and CUSP. With the tests performed, it was verified the behavior of the algorithm when running on the GPU is not efficient with the existing solutions. A new proposal with parallel optimization on the CPU was used and When configured to run with 16 threads achieved a speedup of 6 in some steps of the algorithm When compared with the sequential version.