FURTHER IMPROVEMENTS OF PARALLEL N-FINDR ALGORITHM USING NVIDIA GPUS

Based on our previous work about parallel implementation on N-FINDR in cluster environment, we give some supplements and further developed parallel N-FINDR algorithms by using NVidia graphics processing units (GPUs). Two characterizes are presented. First, extended volume calculation is further considered. Our proposed method is only involved in matrix-vector or vector-vector multiplication, which is suitable to be performed on GPUs. Second, due to the limitation of constant memory providing for GPUs, two patterns for parallel implementation of NFINDR are considered with the imagery in both reduced and original dimension. In experiments, our proposed algorithms greatly improved the performance of N-FINDR and obtained nearly real-time capability.