Iterative digital breast tomosynthesis (DBT) is a technology that mitigates many of the shortcomings associated with traditional mammography. Using multiple low-dose X-ray projections with an itera- tive maximum likelihood estimation method, DBT is able to create a high-quality, three-dimensional reconstruction of the breast. However, the usability of DBT depends largely on making the time for computation acceptable within a clinical setting. In this work we accelerate our DBT algorithm on multiple CUDA-enabled GPUs, reducing the execution time to under 20 seconds for eight iterations (the number usually required to obtain an acceptable quality reconstructed image). The algorithm studied in this work is representative of a general class of image reconstruction problems, as are the thread-mapping strategies, multi-GPU considerations, and optimizations employed in this work.

As general purpose computing on Graphics Processing Units (GPGPU) matures, more complicated scientific applications are being targeted to utilize the data-level parallelism available on a GPU. Implementing physically-based sim- ulation on data-parallel hardware requires preprocessing overhead which affects application performance. We discuss our implementation of physics-based data structures that provide significant performance improvements when used on data- parallel hardware. These data structures allow us to maintain a physics-based abstraction of the underlying data, reduce programmer effort and obtain 6x-8x speedup over previously implemented GPU kernels.

Loop vectorization, a key feature exploited to obtain high performance on Single Instruction Multiple Data (SIMD) vector architectures, is significantly hindered by irregular memory access patterns in the data stream. This paper describes data transformations that allow us to vectorize loops targeting massively multithreaded data parallel architectures. We present a mathematical model that captures loop-based memory access patterns and computes the most appropriate data transformations in order to enable vectorization. Our experimental results show that the proposed data transformations can significantly increase the number of loops that can be vectorized and enhance the data-level parallelism of applications. Our results also show that the overhead associated with our data transformations can be easily amortized as the size of the input data set increases. For the set of high performance benchmark kernels studied, we achieve consistent and significant performance improvements (up to 11.4X) by applying vectorization using our data transformation approach.

Given the rapid growth in computational requirements for medical image analysis, Graphics Processing Units (GPUs) have begun to be utilized to address these demands. But even though GPUs are well-suited to the underlying processing associated with medical image reconstruction, extracting the full benefits of moving to GPU platforms requires significant programming effort, and presents a fundamental barrier for more general adoption of GPU acceleration in a wider range of medical imaging applications. In this paper we describe our experience in accelerating a number of challenging medical imaging applications, and discuss how we utilize profile-guided analysis to reap the full benefits available on GPU platforms. Our work considers different GPU architectures, as well as how to fully exploit the benefits of using multiple GPUs.