Awesome

Forward and Deferred Hashed Shading Paper

About

This repository houses the Forward and Deferred Hashed Shading paper and corresponding source files produced as part of my master thesis in computer science. The paper introduces the Hashed Shading light assignment algorithm used to render a large number of lights efficiently. This algorithm is compared with the popular light assignment algorithms: Tiled Shading and Clustered Shading.

Abstract

(as written in the paper)

This paper introduces the Hashed Shading algorithm, a light assignment algorithm for forward and deferred shading. It uses a subdivision independent from the view frustum in order to reuse data structures between frames. The scene is subdivided into cubes of a specified size and represented using a linkless octree to store the data efficiently in memory and allow for a fast retrieval of relevant lights during shading. The performance of Hashed Shading is compared to both Tiled and Clustered Shading. Forward Hashed Shading reduces the number of light calculations and the execution time by a factor of two compared to Forward Tiled Shading. It achieves a similar reduction in the number of light calculations as Clustered Shading. Furthermore the Hashed Shading algorithm scales slightly better in regards to resolution than Tiled and Clustered Shading due to the camera-independent subdivision. Currently Hashed Shading does not support dynamic lights, and requires significant memory to store the linkless octree. Solutions for these issues are proposed but have not been implemented.

Algorithm

The Hashed Shading algorithm uses an octree to subdivide the scene space. In each leaf node, the intersecting finite lights are stored. This octree can be queried when frames are subsequently rendered, with the scene positions of pixels. Each query returns the set of lights which intersect with the leaf node the pixel falls within. Thus reducing the set of lights that needs to be evaluated.

This octree data structure is independent from the view frustum, and thus the creation of the initial octree can executed as a pre-process.

In order to efficiently access the octree on the GPU, the linkless octree implementation is used. This approach utilises spatial hash maps which are saved in textures, to represent each layer of the octree. Because each layer is represented by a texture, it can be efficiently accessed on the GPU.

Technique comparisons

Hashed Shading is compared with Tiled Shading, Clustered Shading and runs without any light assignment algorithms. The results of these experiments are visualised in the following videos. On the left the actual rendered frame is displayed, on the right the number of light calculations per pixel is visualised, where a white pixel corresponds with the maximum number of light calculations, and black with no light calculations. These experiments were executed for three demo scenes, which can be found in the data repository.