high performance computing on graphics processing units: hgpu.org

Dose Volume Histogram(DVH) is necessary for evaluating radiotherapy planning. With the increase of patient CT slices and the development of intensity-modulated radiation therapy(IMRT) technology, statistical process of DVH requires a large number of cubic interpolation calculation, and the sequential single threaded DVH code on the CPU can not meet the real-time requirement. The paper presents an efficient parallel algorithm of DVH on the Compute Unified Device Architecture (CUDA) platform. The parallel algorithm firstly divides the patient’s tumor targets and the key organs covered by radiation computed tomography(CT) images into independent parts and samples them respectively, then loads their sampled points into the global memory of GPU by CUDA streams and the dose matrix to the texture memory on which the cubic interpolation on dose matrix is frequently computed. The statistics results of DVH are put into the high-speed shared memory. In order to avoid bank conflict on shared memory, the paper have presented a vote algorithm of the share memory, which could greatly reduce the conflict. Experimental results have shown that an average speedup of 7.49 was achieved for the CUDA-based DVH code, while a further improvement with the bank conflict avoidance approach achieved approximate speedup up to 8.0.