Implementação do algoritmo TFM em uma unidade de processamento gráfico utilizando bibliotecas CUDA

Abstract

Ultrasonic Non-Destructive Testing are widely used in industry as a quality control tool, whose main objective is to identify and characterize corrosion and flaws that may compromise the vitality and usability of the analyzed material. One of the main algorithms used for processing data obtained by ultrasonic testing is the Total Focusing Method (TFM), which allows the identification, localization and dimensioning of defects through the image reconstruction of the internal structure of the tested object. Despite its advantages, TFM demands a high computational power, culminating in a high execution time. Thus, this work aimed to optimize this algorithm by parallelizing it on a NVIDIA GeForce 1050 Ti graphics processing unit. Three different versions of TFM were developed using the PyCUDA, CuPy and Numba CUDA libraries integrated into the Python language. Each developed version was executed using acquired data from a homogeneous aluminum block and a toothed acrylic specimen. Data acquisition for each specimen was done both by simulation, using the software CIVA, and physically, using the Panther™ M2M acquisition system. The results obtained showed that the parallel versions presented lower execution time, that is, a significant speed-up in relation to the reference serial version of the algorithm, especially for large data sets.