AVSM: Adaptive Volumetric Shadow Maps

Adaptive volumetric shadow maps (AVSM) is a new approach for real-time shadows that supports highquality shadowing and self-shadowing from dynamic volumetric media for both forward and deferred rendering. Our sample focuses on smoke, but this technique could also be applied to hair and other similar, translucent media. When rendering such effects, self-shadowing is a highly valued feature that gives the proper visual cues needed for defining the shape and structure of the media. In order to perform self-shadowing in volumetric media, it is required to be able to accumulate the partial occlusion value at any point in the volumetric object. This feature usually requires one of two methods: depth slicing or deep shadow maps. Depth slicing often suffers from noticeable slicing artifacts, has limited depth ranges, and/or is limited to one type of media. Deep shadow maps do not suffer from these problems, but are primarily used in off-line rendering systems as they have not been efficiently mapped to current realtime graphics hardware – often due to the unbounded nature of their data storage requirements. In adaptive volumetric shadow mapping, the algorithm generates an adaptively sampled representation of the volumetric transmittance in a shadow-map-like data structure. Each texel of this shadow map stores a compact approximation to the transmittance curve along the corresponding light ray. The main innovation of AVSM is a new streaming, lossy compression algorithm that is capable of building a constant-storage, variable-error representation of visibility that can be used in later shadow lookups. In this sample, a traditional DirectX implementation is provided. We also include a second implementation that uses a new feature of Intel graphics processors. AVSM Algorithm overview The heart of the AVSM algorithm is the generation of a unique style of shadow map that encodes the fraction of visible light at specific distances from the light source. The faction of visible light (transmittance) values are in the interval [0,1]. This algorithm consists of two major steps: a capture stage, and a compression stage.