6140

Hybrid GPU-CPU Adaptive Precision Ray-Triangle Intersection Tests for Robust High-Performance GPU Dosimetry Computations

Lancelot Perrotte, Bruno Bodin, Laurent Chodorge
CEA-LIST, 18 Route du Panorama, 92265 Fontenay-aux-Roses, France
International Conference on Mathematics and Computational Methods Applied to Nuclear Science and Engineering, 2011

@article{perrotte2011hybrid,

   title={HYBRID GPU-CPU ADAPTIVE PRECISION RAY-TRIANGLE INTERSECTION TESTS FOR ROBUST HIGH-PERFORMANCE GPU DOSIMETRY COMPUTATIONS},

   author={Perrotte, L. and Bodin, B. and Chodorge, L.},

   booktitle={International Conference on Mathematics and Computational Methods Applied to Nuclear Science and Engineering},

   year={2011}

}

Download Download (PDF)   View View   Source Source   

1229

views

Before an intervention on a nuclear site, it is essential to study different scenarios to identify the less dangerous one for the operator. Therefore, it is mandatory to dispose of an efficient dosimetry simulation code with accurate results. One classical method in radiation protection is the straight-line attenuation method with build-up factors. In the case of 3D industrial scenes composed of meshes, the computation cost resides in the fast computation of all of the intersections between the rays and the triangles of the scene. Efficient GPU algorithms have already been proposed, that enable dosimetry calculation for a huge scene (800000 rays, 800000 triangles) in a fraction of second. But these algorithms are not robust: because of the roundings caused by floating-point arithmetic, the numerical results of the ray-triangle intersection tests can differ from the expected mathematical results. In worst case scenario, this can lead to a computed dose rate dramatically inferior to the real dose rate to which the operator is exposed. In this paper, we present a hybrid GPU-CPU algorithm to manage adaptive precision floating-point arithmetic. This algorithm allows robust ray-triangle intersection tests, with very small loss of performance (less than 5 % overhead), and without any need for scene-dependent tuning.
No votes yet.
Please wait...

* * *

* * *

HGPU group © 2010-2017 hgpu.org

All rights belong to the respective authors

Contact us: