ETER: Elastic Tessellation for Real-Time Pixel-Accurate Rendering of Large-Scale NURBS Models

ACM Transactions on Graphics (Proc. SIGGRAPH), 42(4), 2023

Teaser: Our elastic tessellation framework, ETER, enables the rendering of NURBS models with pixel accuracy while maintaining real-time frame rates. By utilizing on-the-fly computation of pixel-accurate tessellation factors and our novel crack-filling algorithm, we are able to achieve sufficient mesh density and high rendering quality without cracks at arbitrary zoom levels, as shown in the green and yellow boxes. The performance of our ETER is demonstrated in the histogram on the right, which shows the FPS of rendering according to various scales of four NURBS models (upper right, the numbers #P below the models indicate the numbers of the Bézier patches respectively). The horizontal axis represents the duplication count (#D) and the triangle count (#T, in millions) of the models respectively. The results show that ETER is able to render millions of Bézier patches with 0.25 billion triangles in real-time on an RTX 3090 Ti GPU.

Non-uniform rational B-splines curves and surfaces (NURBS) are the industry standard for modeling. Efficient rendering of large-scale NURBS Models without visual artifacts is highly desirable and regarded as a hard problem.

Existing commercial CAD systems deal with this problem by pre-tessellating NURBS models and discretizing them into LOD (level-of-detail) triangle meshes in an offline manner. However, the pre-tessellation nature of those solutions leads to three defects:
(1) the resolution of the tessellated triangle meshes is fixed once generated, for which unpleasing faceted shading effects and rough polygonal silhouette might be perceived at certain viewing angles.
(2) cracks between adjacent patches/LOD levels are difficult to deal with and most likely will be revealed at close-up views if not handled well.
(3) CAD users are obliged to configure a smaller tolerance value to limit object-space sampling error for favorable rendering quality, resulting in substantially increasing file size and memory usage, putting heavy burden on I/O bandwidth and affecting rendering performance eventually.

Instead of tessellating in a pre-processing step, we turn to pursue an on the fly adaptive tessellation framework for real-time rendering of large-scale NURBS models. Up to the present, no NURBS rendering framework has achieved pixel accuracy, crack-free, and real-time frame rate at the same time when handling large-scale NURBS models.

We present a novel adaptive tessellation framework, named ETER, for achieving pixel-accurate, crack-free, and real-time rendering at the same time for large-scale NURBS models. We develop a highly parallel adaptive tessellation algorithm, a CUDA-based and Tensor Cores accelerated NURBS evaluation algorithm, and an efficient crack-filling algorithm. To evaluate ETER's capability,we implement a prototype rendering system that integrates ETER with hybrid rasterization and generates the output into the visibility buffer.

To the best of our knowledge, this work is the first in the literature to achieve pixel-accurate, crack-free, and real-time rendering at the same time for large-scale NURBS models. We are also the first to utilize Tensor Cores, Mesh Shaders, and hybrid software/hardware rasterization in NURBS model rendering.

Fig. 2: Overview of our rendering pipeline. Boxes with background in orange represent our proposed algorithms.

Fig. 4: (a) Using independent tessellation factor estimation and uniform sampling grid result in cracks, as shown in white pixels. (b) Directly applying conservative rasterization fills the cracks, but also results in severe jagged effects. (c) Our algorithm not only fills the cracks, but also avoids conservative rasterization induced jagged effects.

Fig. 5: The breakdown of the frame time for each stage of the models. It shows the percentage of the total frame time allocated to each stage, including tessellation factor estimation (TPE), evaluation (Eval), rasterization (RAS), trimming test (TRI), crack filling (CRK), shading (SHD), and others.

Back to Homepage