high performance computing on graphics processing units: hgpu.org

The rapid growth of the number of transistors on integrated circuits has enabled numerous advances in computational hardware. Computer graphics development benefited from these advances, reaching a stage where they deliver realistic and rich user experience through amazing graphics. GPUs are nowadays capable of processing massive amounts of data, by taking advantage of its intrinsic parallelism. To meet the processing demands of great amounts of data, primarily from the gaming industry, GPUs’ hardware soon evolved into the multi-core paradigm, and it is now common for any commodity GPU to incorporate several cores. GPUs exhibit hardware that can, potentially, deliver much higher GFLOPS than CPUs, but they are always tightly coupled with APIs designed for game development. With the emergence of platforms that enabled the replacement of otherwise fixed GPU functions in order to produce more complex graphical effects, new possibilities arose for general computing on graphical hardware, endowing the scientific community with new processing tools. This work describes the development of parallel, high performance algorithms for extraction of Digitally Reconstructed Radiographs (DRRs) out of 3D vertebrae models, using the Graphics Processing Unit (GPU). This problem has, inherently, huge amounts of data parallelism, which makes it suitable for acceleration under parallel architectures such as the GPU. The developed algorithms were implemented utilizing the NVIDIA CUDA platform, a parallel computing architecture that enables the programmer to take advantage of the GPU functions. This research is motivated by recent work aiming to attain a process to recover the shape of the spine of human scoliosis patients using two planar radiographs. The former encompasses solving 2D/3D non-rigid image registration problems, involving the creation of DRRs, hundreds of times per second. For this reason, it demands DRR extraction methods with higher throughputs than the ones a CPU can meet. Interesting results were drawn from the research done, which led to the development of ray casting and depth-peeling-based algorithms capable of extracting DRRs in one geometry pass. The work herein reported describes the context of DRR synthesis and the algorithms implemented under the CUDA platform, discussing the limitations imposed by the platform as well as the optimizations possible. It is further presented a set of benchmarks to the algorithms and performance comparison between the new and former solutions.