Prototype renderer that can handle volumetric bricks ("gridlets") plus umeshes, such as they occur if one generates a dual AMR grid, stitches the boundary regions with unstructured elements, but clusters and then represents the interior of the duals with grid primitives.
This project contains sample code for the research paper presented at EuroVis 2023 (Computer Graphics Forum):
Stefan Zellmann, Qi Wu, Kwan-Liu Ma, Ingo Wald:
"Memory-Efficient GPU Volume Path Tracing of AMR Data Using the Dual Mesh"
Computer Graphics Forum, Vol. 42, Issue 3, 2023 (Proceedings of EuroVis 2023)
Citation:
@article {10.1111:cgf.14811,
journal = {Computer Graphics Forum},
title = {{Memory-Efficient GPU Volume Path Tracing of AMR Data Using the Dual Mesh}},
author = {Zellmann, Stefan and Wu, Qi and Ma, Kwan-Liu and Wald, Ingo},
year = {2023},
publisher = {The Eurographics Association and John Wiley & Sons Ltd.},
ISSN = {1467-8659},
DOI = {10.1111/cgf.14811}
}
Author version: https://pds.uni-koeln.de/sites/pds/szellma1/eurovis23-stitcher-author-version-compressed.pdf
High-res author version: https://pds.uni-koeln.de/sites/pds/szellma1/eurovis23-stitcher-author-version.pdf
Official version: https://diglib.eg.org/handle/10.1111/cgf14811
This sample code was tested on Ubuntu 22.04, with nvcc 11.4 and OptiX 7.4.
(Boundary) cells are generated with a separate tool (see
tools/amrMakeDualMesh. First use amrMakeDualMesh to
construct the dual mesh. The "real" unstructured cells end up in a .umesh
file, the remaining files in cube files that are subsequently clustered to
.grids files using the amrMakeGrids tool. These include the gridlets from
the paper.
The project includes an interactive renderer:
./exaStitchViewer boundaryCells.umesh -grids gridlets.grids -scalars scalars.bin
Example:
mkdir build
cd build
cmake .. -DOptiX_INSTALL_DIR="C:/ProgramData/NVIDIA Corporation/OptiX SDK 7.4.0" -DQt5Widgets_DIR="C:/Users/wilso/Documents/Softwares/win/Qt/5.15.2/msvc2019_64/lib/cmake/Qt5Widgets" -DBUILD_SHARED_LIBS=OFF
cmake --build . --config Release
i.e., don't build shared libs; otherwise similar to Linux.
The default file type interpreted by the viewer is umesh; the viewer will
build a sampling BVH over the unstructured elements contained inside and use
that to implement volume sampling with OptiX/RTX. Passing an umesh file is
mandatory for now, as we assume that we'll always need unstructured boundary
elements.
Note that it's perfectly valid for the umesh file to also contain all the
voxels/perfect cubes of the AMR hierarchy and represent them as hexes; that
way, the dual mesh is fully represented with unstructured elements, as one
would e.g. do in a traditional AMR stitching scenario.
Alternatively, the voxels may also come from a grids file (cmdline option
-grids), obtained by generating and then clustering the dual grid with the
amrMakeDual tool. We assume that the umesh then doesn't contain these
voxels/hexes, as they come from the grids file.
The umesh file is allowed to either include per-vertex scalar values
(float32), or cell IDs that point into the scalars file (cmdline option
-scalars). grids files are however always store scalar IDs, i.e., passing a
grids file without scalars will result in rendering being broken because the
scalars are missing
The GPU rendering code is primarily implemented in deviceCode.cu. Here a ray gen program is implemented, plus path tracers using Woodcock tracking, and a BVH traversal mode for majorants. During development we tested different majorant traversal modes and ultimately stuck with a DDA grid (see DDA.h and DDA.cu) (traversal is initiate in deviceCode.cu.
The OptiX intersection and closest hit programs used for sampling with zero-length rays live in separate samplers, implementations of which can be found in sampler. Samplers are associated with models (see model). When loading a model from file, the respective (GPU) sampler will automatically be constructed later on.
The sampler presented in the paper is the ExaStitch sampler. More examples
include the ExaBrick sampler (Wald et al. (2020) "Ray Tracing Structured AMR
Data Using ExaBricks", source code for this sampler is included), the BigMesh
sampler (source code not included, Wald et al. (2021) "A Memory Efficient
Encoding for Ray Tracing Large Unstructured Data"), and the QuickClusters
sampler (Morrical et al. 2022, "Quick Clusters: A GPU-Parallel Partitioning for
Efficient Path Tracing of Unstructured Volumetric Grids", source code
included). These other samplers are used for comparison in the paper, and the
source code was adapted from the original source code developed by the authors
and reused with friendly permission. For the BigMesh sampler we also used the
original implementation, yet cannot provide source code here due to licensing
issues.
Adding additional model/sampler pairs should be straightforward. Pointers how to invoke the various models/samplers can be found in the scripts directory. To generate "ExaBricks" files, the command line tool chain from the original repository must be used.
The CPU host code to invoke the GPU renderer is primarily implemented in OWLRenderer.h and OWLRenderer.cpp.
ExaStitch model files are generated using an offline toolchain (see the description above under "Usage").
The code includes a prototypical "anari" device; to be used with the ANARI-SDK. This code is under development, but will eventually allow us to integrate the data structure with VTK or VTK-m. Example code how to use the device and anari extension can be found under anari/example.md.
We use the following external dependencies (included in the submodules directory):
- OWL: the OptiX 7 wrapper library, providing easy access to the OptiX library, by Ingo Wald
- UMesh: library to represent unstructured element meshes on the host and file system, by Ingo Wald
- cuteeOwl: QT GUI providing transfer function editing, by Ingo Wald
- Visionaray: is used for a software BVH implementation; included here, but only relevant for building offline kd-trees with majorant (not discussed in the paper)
- Stefan Zellmann, Qi Wu, Kwan-Liu Ma, and Ingo Wald
- Nate Morrical (provided code for the "QuickClusters" sampler)
Apache 2, if not noted otherwise. See the LICENSE file.