high performance computing on graphics processing units: hgpu.org

Tomographic image reconstruction, such as the reconstruction of CT projection values, of tomosynthesis data, PET or SPECT events, is computational very demanding. The most time-consuming step is the backprojection which is often limited by the memory bandwidth. Recently, a novel general purpose architecture optimized for distributed computing became available: the Cell Broadband Engine (CBE). Its eight synergistic processing elements (SPEs) currently allow for a theoretical performance of 192 GFlops (3 GHz, 8 units, 4 floats per vector, 2 instructions, multiply and add, per clock). To maximize image reconstruction speed we modified our parallel-beam backprojection algorithm that is highly optimized for standard PCs, and optimized the code for the cell processor. Data mining techniques and double buffering of source data were extensively used to optimally utilize both the memory bandwidth and the available local store of each SPE. The pixel-driven backprojection code uses floating point arithmetic and either linear interpolation (LI) or nearest neighbor (NN) interpolation between neighboring detector channels. Performance was measured using simulated data with 512 parallel beam projections per half rotation and 1024 detector elements. The data were backprojected into an image of 512 by 512 pixels using our PC-based approach and the new cell-based algorithm. Both the PC and the CBE were clocked at 3 GHz. Images obtained were found to be identical with both approaches. A throughput of 11 fps (LI) and 15 fps (NN) was measured on the PC whereas the CBE achieved 126 fps (LI) and 165 fps (NN). Thereby, the cell greatly outperforms today’s top-notch backprojections based on graphical processing units (GPU). Using both CBEs of our dual cell-based blade (Mercury Computer Systems) one can backproject 252 images per second with LI and and 330 images per second with NN.