Non-Linearly Quantized Moment Shadow Maps

Christoph Peters
In proceedings of Proceedings of the 9th Conference on High-Performance Graphics, Los Angeles, CA, ACM, July 2017
 

Abstract

Moment shadow maps enable direct filtering to accomplish proper antialiasing of dynamic hard shadows. For each texel, the moment shadow map stores four powers of the depth in either 64 or 128 bits. After filtering, this information enables a heuristic reconstruction. However, the rounding errors introduced at 64 bits per texel necessitate a bias that strengthens light leaking artifacts noticeably. In this paper, we propose a non-linear transform which maps the four moments to four quantities describing the depth distribution more directly. These quantities can then be quantized to a total of 32 or 64 bits. At 64 bits, the results are virtually indistinguishable from moment shadow mapping at 128 bits per texel. Even at 32 bits, there is hardly any additional light leaking but banding artifacts may occur. At the same time, the computational overhead for the reconstruction is reduced. As a prerequisite for the use of these quantization schemes, we propose a compute shader that applies a resolve for a multisampled shadow map and a 9² two-pass Gaussian filter in shared memory. The quantized moments are written back to device memory only once at the very end. This approach makes our technique roughly as fast as variance shadow mapping with 32 bits per texel. Since hardware-accelerated bilinear filtering is incompatible with non-linear quantization, we employ blue noise dithering as inexpensive alternative to manual bilinear filtering.

Keywords: 32-bit quantization, compute shader, fast Gaussian blur, moment shadow mapping, non-linear quantization, on-chip filtering, real-time rendering, shadows

An executable demo is available on the author's blog.

Images

Download Paper

Download Paper

Additional Material

Bibtex

@INPROCEEDINGS{peters-2017-quantized-msm,
     author = {Peters, Christoph},
      title = {Non-Linearly Quantized Moment Shadow Maps},
  booktitle = {Proceedings of the 9th Conference on High-Performance Graphics},
     series = {HPG '17},
       year = {2017},
      month = jul,
  publisher = {ACM},
   location = {Los Angeles, CA},
   keywords = {32-bit quantization, compute shader, fast Gaussian blur, moment shadow mapping, non-linear
               quantization, on-chip filtering, real-time rendering, shadows},
   abstract = {Moment shadow maps enable direct filtering to accomplish proper antialiasing of dynamic hard
               shadows. For each texel, the moment shadow map stores four powers of the depth in either 64 or 128
               bits. After filtering, this information enables a heuristic reconstruction. However, the rounding
               errors introduced at 64 bits per texel necessitate a bias that strengthens light leaking artifacts
               noticeably. In this paper, we propose a non-linear transform which maps the four moments to four
               quantities describing the depth distribution more directly. These quantities can then be quantized
               to a total of 32 or 64 bits. At 64 bits, the results are virtually indistinguishable from moment
               shadow mapping at 128 bits per texel. Even at 32 bits, there is hardly any additional light leaking
               but banding artifacts may occur. At the same time, the computational overhead for the reconstruction
               is reduced. As a prerequisite for the use of these quantization schemes, we propose a compute shader
               that applies a resolve for a multisampled shadow map and a 9² two-pass Gaussian filter in shared
               memory. The quantized moments are written back to device memory only once at the very end. This
               approach makes our technique roughly as fast as variance shadow mapping with 32 bits per texel.
               Since hardware-accelerated bilinear filtering is incompatible with non-linear quantization, we
               employ blue noise dithering as inexpensive alternative to manual bilinear filtering.},
       isbn = {978-1-4503-5101-0/17/07},
        url = {https://doi.org/10.1145/3105762.3105775},
        doi = {10.1145/3105762.3105775}
}