Eurographics Symposium on Parallel Graphics and Visualization (2015), pp. 1–9 C. Dachsbacher, P. Navrátil (Editors) SIMD Parallel Ray Tracing of Homogeneous Polyhedral Grids Brad Rathke 1,2 Ingo Wald 1 Kenneth Chiu 2 Carson Brownlee 3 1 Technical Computing Group, Intel Corp. 2 Binghamton University, State University of New York 3 Texas Advanced Computing Center, University of Texas Abstract Efficient visualization of unstructured data is vital for domain scientists, yet is often impeded by techniques which rely on intermediate representations that consume time and memory, require resampling data, or inefficient im- plementations of direct ray tracing methods. Previous work to accelerate rendering of unstructured grids have focused on the use of GPUs that are not available in many large-scale computing systems. In this paper, we present a new technique for directly visualizing unstructured grids using a software ray tracer built as a module for the OSPRay ray tracing framework from Intel. Our method is capable of implicit isosurface rendering and di- rect volume ray casting homogeneous grids of hexahedra, tetrahedra, and multi-level datasets at interactive frame rates on compute nodes without a GPU using an open-source, production-level ray tracing framework that scales with variable SIMD widths and thread counts. 1. Introduction Rapid advancements in computing power and the sophis- tication of simulations have allowed scientists to simulate physical phenomena at increasingly large scales. Such sim- ulations are challenging to effectively visualize at full res- olution. Many techniques to optimize massive data render- ing have focused on polygonal data or regular grids, how- ever simulations often generate unstructured data in the form of tetrahedral, hexahedral, or other polyhedra which pro- duce many challenges for interactive rendering. For isosur- facing, visualization tools such as VisIt [CBW ∗ 12] and Para- View [Hen04] use marching cubes [LC87] to extract a sur- face explicitly. This intermediate representation requires ad- ditional time and memory over the existing dataset and needs to be regenerated every time the isovalue is modified result- ing in unnecessary computational overhead during data ex- ploration. Rasterizing the resulting triangles also presents several issues for performance and usability. Many of the result- ing triangles may be occluded or of sub-pixel size in large datasets. Transparency presents another issue for traditional rasterization techniques due to the rendering order depen- dent nature of rasterizing transparent objects. Culling and Intel, Xeon, and Xeon Phi are trademarks of the Intel Corporation in the U.S. and other countries. Other product names and brands may be claimed as property of others. Figure 1: Our method allows for interactively rendering unstructured data sets (in this example, the “Jets” of 12M tetrahedral cells), and supports both ray tracing of the imp- licit isosurfaces as well as direct volume ray casting. Left: Three semi-transparent (implicit) isosurfaces, with proper transparency handled by the ray tracer. Right: The same vol- ume with direct volume ray casting, using a transfer function chosen to highlight the same features shown in the isosurface rendering. depth peeling methods can alleviate problems with rasteri- zation of transparent objects, but add complexity to the ren- dering algorithms. Directly ray tracing the unstructured grid, however, requires no intermediate representation, can accu- rately render transparency, and implicitly handles occluded or sub-pixel regions. In the case of volume rendering, special-case solutions can be used that do not easily integrate with the rest of the visualization tool, use lower fidelity splatting, or resample submitted to Eurographics Symposium on Parallel Graphics and Visualization (2015)