This paper proposes a cache-aware optimization method to accelerate out-of-core cone beam computed tomography reconstruction on a graphics processing unit (GPU) device. Our proposed method extends a previous method by increasing the cache hit rate so as to speed up the reconstruction of high-resolution volumes that exceed the capacity of device memory. More specifically, our approach accelerates the well-known Feldkamp-Davis-Kress algorithm by utilizing the following three strategies: (1) a loop organization strategy that identifies the best tradeoff point between the cache hit rate and the number of off-chip memory accesses; (2) a data structure that exploits high locality within a layered texture; and (3) a fully pipelined strategy for hiding file input/output (I/O) time with GPU execution and data transfer times. We implement our proposed method on NVIDIA's latest Maxwell architecture and provide tuning guidelines for adjusting the execution parameters, which include the granularity and shape of thread blocks as well as the granularity of I/O data to be streamed through the pipeline, which maximizes reconstruction performance. Our experimental results show that it took less than three minutes to reconstruct a 20483-voxel volume from 1200 20482-pixel projection images on a single GPU; this translates to a speedup of approximately 1.47 as compared to the previous method.

Fast computation of singular value decomposition (SVD) is of great interest in various machine learning tasks. Recently, SVD methods based on randomized linear algebra have shown significant speedup in this regime. For processing large-scale data, computing systems with accelerators like GPUs have become the mainstream approach. In those systems, access to the input data dominates the overall process time; therefore, it is needed to design an out-of-core algorithm to dispatch the computation into accelerators. This paper proposes an accurate two-pass randomized SVD, named block randomized SVD (BRSVD), designed for matrices with a slow-decay singular spectrum that is often observed in image data. BRSVD fully utilizes the power of modern computing system architectures and efficiently processes large-scale data in a parallel and out-of-core fashion. Our experiments show that BRSVD effectively moves the performance bottleneck from data transfer to computation, so that outperforms existing randomized SVD methods in terms of speed with retaining similar accuracy.