high performance computing on graphics processing units: hgpu.org

Ultrasonic imaging and reconstruction tools are com-monly used to detect, identify and measure defects in different mechanical parts. Due to the complexity of the underlying physics, and due to the evergrowing quantity of acquired data, computation time is becoming a limitation to the opti-mal inspection of a mechanical part. This article presents the performances of several implementations of a computational heavy algorithm, named Total Focusing Method, on both gra-phics processing units (GPU) and general purpose processors (GPP). The scope of this study is narrowed to planar parts tested for defects in immersion. Using algorithmic simplifications and architectural opti-mizations, the algorithm has been drastically accelerated re-sulting in memory-bound implementations. On GPU, high performances can be achieved by profiting from GPU long cache-lines and from hand managed memory. Whereas on GPP, computations cost are overrun by memory access resul-ting in less efficient performances compared to the computing capabilities available. The following study constitutes the first step toward ana-lyzing the target algorithm for diverse hardware in the non-destructive testing environment.