Performance improvements for iterative electron tomography reconstruction using graphics processing units (GPUs)

Abstract

Iterative reconstruction algorithms are becoming increasingly important in electron tomography of biological samples. These algorithms, however, impose major computational demands. Parallelization must be employed to maintain acceptable running times. Graphics Processing Units (GPUs) have been demonstrated to be highly cost-effective for carrying out these computations with a high degree of parallelism. In a recent paper by Xu et al.[1], a GPU implementation strategy was presented that obtains a speedup of an order of magnitude over a previously proposed GPU-based electron tomography implementation. In this technical note, we demonstrate that by making alternative design decisions in the GPU implementation, an additional speedup can be obtained, again of an order of magnitude. By carefully considering memory access locality when dividing the workload among blocks of threads, the GPU's cache is used more efficiently, making more effective use of the available memory bandwidth.Keywords: Electron Tomography, Reconstruction, GPU Recently, iterative algebraic methods, such as ART and SIRT, have gained popularity in the electron tomography community due to their flexibility with respect to the geometric parameters of the tilt series, and their ability to handle noisy projection data. The use of algebraic reconstruction methods imposes major computational demands. Depending on the number of iterations, reconstructing a large 3D volume with a sequential implementation can easily take days on a normal PC. This obstacle can be largely overcome by parallelizing the computations, in particular the projection and backprojection steps. Graphics Processing Units (GPUs) have recently emerged as powerful parallel processors for general-purpose computations. Their architecture allows operations to be performed on a large number of data elements simultaneously.Several algorithmic strategies have already been proposed for implementing algebraic methods for electron tomography on the GPU. In [2], it was demonstrated by Castaño-Diez et al. that at that time, a GPU implementation of the SIRT algorithm could achieve similar performance to a CPU implementation running on a medium sized cluster. Xu et al. recently proposed a different implementation strategy [1] that leads to a speedup of an order of magnitude compared to the results from [2]. They attribute this speedup to improvements in three categories: minimizing synchronization overhead, encouraging latency hiding, and exploiting RGBA channel parallelism. The first two design goals are interdependent and cannot be optimized separately. In this technical note, we argue that by exploiting data locality more effectively, the runtime of the projection and back projection operations can be substantially reduced, even though the required number of thread synchronization steps will increase. We demonstrate that a significant speedup can be gained in this manner. Exploiting data localityThe Graphics Processing Unit (GPU) is well suited for carrying out the computations involved in ...

Formation of bimetallic clusters in superfluid helium nanodroplets analysed by atomic resolution electron tomography

“…For the tomographic reconstructions, two different types of algorithm have been used. First, a reconstruction was done using a multiplicative SIRT algorithm with 200 iterations and a relaxation parameter NATURE COMMUNICATIONS | DOI: 10.1038/ncomms9779 ARTICLE l ¼ 0.3, implemented in MATLAB and based on projection and backprojection operators from the ASTRA toolbox using the Graphics Processing Unit 32 . With this reconstruction, high-resolution structural information (Fig.…”

Adaptive Grid Refinement for Discrete Tomography

“…The reconstruction is done on an underlying fine grid of 128 × 128 in order to avoid the inverse crime. We use the ASTRA toolbox to compute the forward and backward projections [15]. For the adaptive method the mapping matrices as discussed above are used to map to and from the QT grids to the …”

High-Level Algorithm Prototyping: An Example Extending the TVR-DART Algorithm

“…When working with tomographic inverse problems, one such task is computing the 2D/3D ray transform and its adjoint. For this ODL employs the ASTRA toolbox [11,13,14], which is a high-performance, GPU accelerated toolbox for tomographic reconstruction freely available from http://www.astra-toolbox.com. The toolbox supports many different data manifolds arising in tomography, including for example circular cone beam, laminography, tomosynthesis, and electron tomography, see [13] for details.…”