Rapid data processing for ultrafast x-ray computed tomography using scalable and modular CUDA based pipelines

Ultrafast X-ray tomography is an advanced imaging technique for the study of dynamic processes basing on the principles of electron beam scanning. A typical application case for this technique is e.g.the study of multiphase flows, that is, flows of mixtures of substances such as gas-liquid flows in pipelines or chemical reactors. At Helmholtz-Zentrum Dresden-Rossendorf (HZDR) a number of such tomography scanners are operated. Currently, there are two main points limiting their application in some fields. First, after each CT scan sequence the data of the radiation detector must be downloaded from the scanner to a data processing machine. Second, the current data processing is comparably time-consuming compared to the CT scan sequence interval. To enable online observations or use this technique to control actuators in real-time, a modular and scalable data processing tool has been developed, consisting of user-definable stages working independently together in a so called data processing pipeline, that keeps up with the CT scanner’s maximal frame rate of up to 8 kHz. The newly developed data processing stages are freely programmable and combinable. In order to achieve the highest processing performance all relevant data processing steps, which are required for a standard slice image reconstruction, were individually implemented in separate stages using Graphics Processing Units (GPUs) and NVIDIA’s CUDA programming language. Data processing performance tests on different high-end GPUs (Tesla K20c, GeForce GTX 1080, Tesla P100) showed excellent performance.

超快X射线断层成像（Ultrafast X-ray tomography）是一种基于电子束扫描原理的先进动态过程研究成像技术。该技术的典型应用案例包括多相流研究，即管道内或化学反应器中气液流等物质混合流体的流动研究。德累斯顿-罗森多夫亥姆霍兹中心（Helmholtz-Zentrum Dresden-Rossendorf, HZDR）目前运营有多台此类断层扫描设备。当前，两大瓶颈制约了该技术在部分领域的应用：其一，每次CT扫描序列完成后，辐射探测器采集的数据均需从扫描仪传输至数据处理主机；其二，当前的数据处理耗时相较CT扫描序列的间隔时长偏高。为实现在线观测或利用该技术对执行机构进行实时控制，研究团队开发了一款模块化且可扩展的数据处理工具：该工具由可自定义的独立处理单元组成，以所谓的数据处理流水线形式协同工作，可匹配CT扫描仪最高达8 kHz的帧率。全新开发的处理单元支持自由编程与灵活组合。为实现最优处理性能，标准断层图像重建所需的全部相关数据处理步骤均基于图形处理器（Graphics Processing Units, GPUs）及英伟达（NVIDIA）的CUDA编程语言，以独立单元形式逐一实现。针对多款高端GPU（Tesla K20c、GeForce GTX 1080、Tesla P100）开展的数据处理性能测试结果显示，该工具性能优异。