Abstract: Forward rendering pipelines shade fragments in triangle-submission order. Consequently, non-visible fragments are often wastefully shaded before being subsequently occluded--a phenomenon known as over-shading. A popular way to avoid over-shading is to only compute surface attributes for each fragment during a forward pass and store them in a buffer. Lighting is performed in a subsequent pass, consuming the attributes buffer. This strategy is commonly known as deferred shading.

We identify two notable deficits. First, the buffer for the geometric surface attributes (g-buffer) is large, often 20+ bytes per visibility sample. The bandwidth required to read and write large g-buffers can be prohibitive on mobile or integrated GPUs. Second, the separation of shading work and visibility is incomplete. Surface attributes are eagerly computed in the forward pass on occluded fragments, wasting texture bandwidth and compute resources.

To address these problems, we propose to replace the g-buffer with a simple visibility buffer that only stores a triangle index and instance ID per sample, encoded in as few as four bytes. This significantly reduces storage and bandwidth requirements. Generating a visibility buffer is cheaper than generating a g-buffer and does not require texture reads or any surface-material-specific computation. The deferred shading pass accesses triangle data with this index to compute barycentric coordinates and interpolated vertex data. By minimizing the memory footprint of the visibility solution, we reduce the working set of the deferred rendering pipeline. This results in improved performance on bandwidth-limited GPU platforms, especially for high-resolution workloads.