Abstract
We present a new faster GPU computation and algorithm for CADMAS-SURF/3D by a CUDA parallelization techinique on a GPU accelerator "Tesla". In this paper, we examine the CUDA porting tool for GPU computation and suggest a macro code by C preprocessor for CADMAS-SURF/3D originally coded in FORTRAN to be adopted into CUDA C language. Furthermore, we optimize the data access sequence pattern from a profiling result to eliminate the calculation bottleneck. We inspect the legitimacy of the solution by examining the rounding error, and then examine the program tuning by a data sequence access pattern optimization.
