Accelerating CT Reconstruction for Synchrotron Experiments using Approximated computing

by Harish Kumar Harihara Subramanian

Tuesday Sep 20, 2022, 11:00 AM → 11:45 AM Europe/Berlin

https://kit-lecture.zoom.us/j/63760248497?pwd=dVJzbkdCeHlkWEo0NVBVTkNWQ2hyZz09 (Zoom)

Joint seminar with IPS

Lecture of Harish Kumar Harihara Subramanian: Accelerating CT Reconstruction for Synchrotron Experiments using Approximated computing

 

Computed Tomography (CT) is a non-destructive technique to observe the internal structure of materials and smaller organisms. There is an increasing drive to use CT also to study dynamic processes in materials and to perform in vivo imaging of living organism. There are challenges to 1) enable real-time reconstruction to monitor and adjust experimental conditions on the fly 2) perform high-quality reconstruction from under-sampled data, hence reducing required radiation dose. In this work, we aimed to improve performance of both analytical and iterative reconstruction methods. Faster analytical reconstruction enables real-time tomography and speed-up of iterative techniques allows us to bring them to a practical use on large real-world datasets. For these ends, we designed faster back-projection and forward-projection algorithms. The proposed methods rely on additional levels of parallelism available in the texture engine of nowadays GPUs and improved utilization of the texture cache via optimized access patterns and mixed-precision computations. This allows us to reconstruct multiple slices in parallel and results in 5x speed-up of both backward- and forward-projectors as compared to the state-of-the-art methods. The new projection primitives are integrated into the UFO framework to enable fast real-time reconstruction. We also integrated Ufo framework as a backend to provide projection primitives into the CCPi iterative reconstruction framework. As a result, we achieve up to 5x speed-up in both analytical and iterative reconstruction. We will discuss the optimization techniques and evaluate trade-offs between the performance and reconstruction quality resulting from mixed-precision computing.

 

Lecture Language: English

-----

Vortrag von Harish Kumar Harihara Subramanian zum Thema:

Vortragssprache: Englisch