iFDK: A Scalable Framework for Instant High-Resolution Image Reconstruction

Computed Tomography (CT) is a widely used technology that requires compute-intense algorithms for image reconstruction. We propose a novel back-projection algorithm that reduces the projection computation cost to 1/6 of the standard algorithm. We also propose an efficient implementation that takes advantage of the heterogeneity of GPU-accelerated systems by overlapping the filtering and back-projection stages on CPUs and GPUs, respectively. Finally, we propose a distributed framework for high-resolution image reconstruction on state-of-the-art GPU-accelerated supercomputers. The framework relies on an elaborate interleave of MPI collective communication steps to achieve scalable communication. Evaluation on a single Tesla V100 GPU demonstrates that our backprojection kernel performs up to 1.6× faster than the standard FDK implementation. We also demonstrate the scalability and instantaneous CT capability of the distributed framework by using up to 2,048 V100 GPUs to solve 4K and 8K problems within 30 seconds and 2 minutes, respectively (including I/O).

1 Introduction

High-resolution Compute Tomography (CT) is a technology used in a wide variety of fields, e.g. medical diagnosis, non-invasive inspection [62], and reverse engineering [17], [50]. In the past decades, the size of a single three-dimensional (3D) volume generated by CT systems has increased from hundreds of megabytes (the typical sizes of a volume are 256³,512³ to several gigabytes (i.e. 2048³, 4096³) [7], [42], [66]. The increased demand for rapid tomography reconstruction and the associated high computational cost attracted heavy attention and efforts from the HPC community [8], [11], [19], [25], [28], [47], [54], [55], [66], [68], [76]. As illustrated in [48], the FDK¹

Feldkamp, Davis, and Kress [23] presented a convolution-backprojection formulation (known as FDK algorithm) for CT image reconstruction in 1984. FDK is also known as the Filtered Back Projection (FBP) algorithm.

algorithm is widely regarded as the primary method to reconstruct 3D images (or volumes) from projections, i.e. X-ray images. The FDK algorithm includes a filtering stage (also known as convolution) and a back-projection stage. The computational complexities of those two stages are and , respectively. Researchers are increasingly relying on the latest accelerators to improve the computational performance of FDK, e.g. Application Specific Integrated Circuits (ASIC) [72], Field-Programming Gate Array (FPGA) [16], [27], [64], [75], Digital Signal Processor (DSP) [37], Intel Xeon-Phi [53], Multi-core CPUs [68], and Graphics Processing Unit (GPU) [51], [73], [77], [78]. This paper focuses on GPU-accelerated supercomputers for two reasons. First, GPUs are dominantly used for tomographic image reconstruction [20], [28], [33], [55], [59], [74]. Second, GPU-accelerated supercomputers are increasingly gaining ground in top-tier HPC systems.