Skip to Main content Skip to Navigation

Visualisations interactives haute-performance de données volumiques massives : une approche out-of-core multi-résolution basée GPUs

Abstract : The needs for volume data visualization are common in several scientic elds, in par- ticular in bio-medical imaging. Indeed, several types of frequently used acquisition devices generate scalar and vectorial elds represented as a 3D regular grid. It is important to be able to visualize interactively these volumes, in order to extract information or to vali- date experimental results. However, the increase in acquisition accuracy of these modern devices induces an exponential growth of the amount of data. To deal with this problem, visualization algorithms must be adapted, both, to the volume of data they handle and its steady growth. The use of GPU accelerator cards is particularly well suited to the nature of volume data and the associated visualization algorithms. High-performance computing en- vironments are now turning to solutions that uses a large number of such cards. These are, by their massively parallel nature, good candidates to oer high-performance visualization solutions. However, the amount of memory in GPUs is very limited, and is much less impor- tant than the size of the raw data of the volumes to be handled. One solution is to design out-of-core algorithms, where the computing unit is dissociated from the data storage unit. In this thesis work, we propose a complete pipeline for interactive visualization on the GPU of very large volumes of data exceeding the physical capacities of the GPU and the CPU memory independently of the machine used for the rendering. For this purpose, we study an out-of-core management model, based on a memory virtualization principle, particularly well adapted to very large volumes. We propose an approach that includes a virtual addressing structure, fully managed on the GPU. We are also interested in the compatibility of this model for dierent types of volume data visualization applications. We propose a rst application that uses a virtual microscope principle to provide autoste- reoscopic 3D visualization of ultra-high resolution image stacks ; a second one that oers interactive direct volume rendering with a ray-guided approach, showing the usability of our out-of-core management model in hybrid, multi-GPUs, multi-CPUs high-performance computing environments.
Complete list of metadatas

Cited literature [124 references]  Display  Hide  Download
Contributor : Jonathan Sarton <>
Submitted on : Tuesday, March 12, 2019 - 12:01:26 PM
Last modification on : Wednesday, July 17, 2019 - 12:18:03 PM
Long-term archiving on: : Thursday, June 13, 2019 - 2:50:41 PM


Files produced by the author(s)


  • HAL Id : tel-02064918, version 1



Jonathan Sarton. Visualisations interactives haute-performance de données volumiques massives : une approche out-of-core multi-résolution basée GPUs. Calcul parallèle, distribué et partagé [cs.DC]. Université de Reims Champagne Ardenne URCA, 2018. Français. ⟨tel-02064918⟩



Record views


Files downloads