Approximate Puzzlepiece Compositing

TL;DR

The paper tackles the challenge of distributed volume rendering for large AMR and unstructured meshes with nonconvex, puzzle-piece-like partitions, where conventional sort-last compositing suffers from high communication and non-convexity issues. It introduces Approximate Puzzlepiece Compositing (APC), a two-stage distributed rendering approach that adapts Moment-Based Order-Independent Transparency (MBOIT) to compute per-pixel moments locally and then aggregate global moments with a single MPI_Allreduce, followed by local rendering using the global transmittance and a final image assembly via MPI_Reduce. The authors demonstrate that APC achieves scalable, high-quality rendering across synthetic and real CFD datasets, with minimal communication overhead and rendering quality close to single-node MBOIT, validating its suitability for in situ visualization on HPC systems. Although APC is approximate and may slightly overestimate transmittance, its order-independence and low communication cost provide strong practical benefits for visualizing complex, non-convex data distributions in large-scale simulations.

Abstract

The increasing demand for larger and higher fidelity simulations has made Adaptive Mesh Refinement (AMR) and unstructured mesh techniques essential to focus compute effort and memory cost on just the areas of interest in the simulation domain. The distribution of these meshes over the compute nodes is often determined by balancing compute, memory, and network costs, leading to distributions with jagged nonconvex boundaries that fit together much like puzzle pieces. It is expensive, and sometimes impossible, to re-partition the data posing a challenge for in situ and post hoc visualization as the data cannot be rendered using standard sort-last compositing techniques that require a convex and disjoint data partitioning. We present a new distributed volume rendering and compositing algorithm, Approximate Puzzlepiece Compositing, that enables fast and high-accuracy in-place rendering of AMR and unstructured meshes. Our approach builds on Moment-Based Ordered-Independent Transparency to achieve a scalable, order-independent compositing algorithm that requires little communication and does not impose requirements on the data partitioning. We evaluate the image quality and scalability of our approach on synthetic data and two large-scale unstructured meshes on HPC systems by comparing to state-of-the-art sort-last compositing techniques, highlighting our approach's minimal overhead at higher core counts. We demonstrate that Approximate Puzzlepiece Compositing provides a scalable, high-performance, and high-quality distributed rendering approach applicable to the complex data distributions encountered in large-scale CFD simulations.

Figures (5)

• Figure 1: The FUN3D data. Note the uneven and jagged data boundary of an individual subdomain's data in (a). An illustration of the volume boundaries of two subdomains is shown in (b).
• Figure 2: A comparison of approximate OIT techniques on a synthetic red and blue concentric circles example. This configuration produces sharp changes in color, challenging approximate OIT methods. (b,c) compare different MBOIT configurations with sort-last (green line): 4 power moments (red), 6 power moments (blue) and trigonometric moments (purple). We observe that APC achieves the same rendering quality as single-node MBOIT, providing a high-quality approximation. This is in contrast to the color artifacts of Hybrid Transparency (j) or occlusion errors from Weight-blended OIT (k). Furthermore, with 4 power moments and a bias (c), the APC image closely approximates the ground truth sort-last. The respective image similarity measurements are (h) vs (i): SSIM=0.99, MSE=38.18, PSNR=32.34 (h) vs (j) SSIM=0.87, MSE=618.76, PSNR=20.24 (h) vs (k) SSIM=0.98, MSE=21.55, PSNR=34.82
• Figure 3: A synthetic case where APC exhibits noticeable image quality loss. With thin, near-opaque red structures in the middle of a blue volume, we can see that MBOIT does not handle abrupt transmittance changes well such that (c) looks more blended, i.e., more purple, than (b).
• Figure 4: An illustration of the APC pipeline. Stage one computes a local moments vector per pixel on each rank, adding them up in the moments AllReduce step to form the global moments. The global moments are used in stage two on each rank to approximate transmittance when rendering their local volume to produce final subimages. Finally, all subimages are added using a Reduce onto the display rank.
• Figure 5: An example image produced from distributed rendering of a sandwiched volume over three ranks. Each rank has two interleaved slices of a cube. The occluded part has its opacity corrected with MBOIT's approximation, i.e., each rank here has its local objects rendered with the "correct alpha" as it would appear in the final image, which can thus be produced by a simple image space summation.