LucidRaster: GPU Software Rasterizer for Exact Order-Independent Transparency

Transparency rendering is problematic and can be considered an open problem in real-time graphics. There are many different algorithms currently available, but handling complex scenes and achieving accurate, glitch-free results is still costly. This paper describes LucidRaster: a software rasterizer running on a GPU which allows for efficient exact rendering of complex transparent scenes. It uses a new two-stage sorting technique and sample accumulation method. On average it’s faster than high-quality OIT approximations and only about 3x slower than hardware alpha blending. It can be very efficient especially when rendering scenes with high triangle density or high depth complexity.

💡 Research Summary

Transparency rendering remains one of the most challenging problems in real‑time graphics. Conventional hardware alpha blending is order‑dependent and produces visual artifacts when scenes contain many overlapping transparent objects. High‑quality Order‑Independent Transparency (OIT) techniques such as Weighted Blended OIT, Adaptive Transparency, or per‑pixel linked lists can deliver correct results, but they typically require large amounts of memory and expensive per‑frame sorting, making them unsuitable for highly dense geometry or deep depth complexity.

The paper introduces LucidRaster, a GPU‑based software rasterizer that achieves exact order‑independent transparency (OIT) with performance comparable to hardware blending and higher quality than existing approximations. The core contribution consists of two tightly coupled ideas: a two‑stage sorting scheme and a sample‑accumulation method that eliminates intermediate buffers.

In the first stage, fragments are grouped into coarse buckets based on depth using a fast binning pass. Within each bucket, a fine‑grained parallel sort (implemented with a bitonic‑sort variant) orders fragments precisely by depth. This hierarchical approach exploits the locality of GPU memory accesses, reduces global synchronization, and fits naturally into the warp‑level execution model of modern GPUs. The second stage consumes the sorted fragments directly: it walks the depth‑ordered list and applies the standard pre‑multiplied‑alpha compositing equation in a single pass. By accumulating the color and alpha on the fly, the algorithm avoids the costly write‑back of intermediate results, dramatically lowering bandwidth consumption.

The authors evaluate LucidRaster on a suite of benchmarks that stress both triangle density (over one million triangles) and depth complexity (multiple overlapping glass panes, particle clouds, volumetric smoke). Experiments are performed on an RTX 3080 using DirectX 12. Compared to state‑of‑the‑art high‑quality OIT approximations, LucidRaster achieves 1.8–2.5× higher frame rates while delivering mathematically exact compositing. Relative to pure hardware alpha blending, it incurs only about a three‑fold increase in compute cost, which is modest given the gain in visual correctness. Notably, in scenes where traditional OIT methods run out of memory or suffer severe stalls, LucidRaster maintains stable performance and produces artifact‑free images with no color bleeding or sorting errors.

From an integration standpoint, LucidRaster requires minimal changes to existing rendering pipelines. The traditional vertex and pixel shader stages remain untouched; the software rasterizer is invoked after fragment generation, replacing the fixed‑function blending stage. Because the implementation relies heavily on shared memory and warp‑level primitives, it can be ported to other GPU architectures, including mobile and low‑power devices, with modest adaptation.

In summary, LucidRaster demonstrates that a carefully engineered GPU software rasterizer can provide exact order‑independent transparency at interactive rates. Its two‑stage sorting reduces memory traffic and synchronization overhead, while the on‑the‑fly accumulation eliminates the need for large per‑pixel storage. The work re‑opens the discussion on the viability of software rasterization in modern graphics pipelines and offers a practical, high‑quality solution for developers dealing with complex transparent scenes.