My name is Holger Grün. I do work within the European GPU ISV relations team of AMD. If you want to know how to drastically reduce the number of texture operations for conventional shdaow map filtering you should read my article. With the methods described in the article one can evaluate shadow map filters defined by a separable weights matrix in only (N/2)x(N/2) texture operations. A filter that does consider 8x8 visibility samples would usually be carried out using 49 PCF operations but now only 16 PCF operations are necessary. The article further describes techniques for fast shadow filtering using non-separable weight matrices and presents two usecases for these techniques.
Hey Holger, I'd love to know what do you think about this article. Am I missing something?
ReplyDeleteHi Ignacio,
ReplyDeletethanks for your comment and for pointing me to your article. I don't disagree with what you write there.
The GDC presentation shouldn't have called the filters used for the GDC screenshots 'Gaussian', as the filters used were not separable. This was just wrong!
Otherwise the method presented allows for more freedom for defining weights matrices for shadow filtering and has real use cases - some of these are shown in the GPUPro article.
Cheers,
Holger
Ignacio,
ReplyDeleteto add to my earlier post, I think your article is very interesting and your findings can be used to produce very good approximations of many of the filters covered by the techniques described in my article. Nevertheless my article shows exact solutions for certain filter kernels that have not been commonly used in shadow mapping and that were not known to a broad audience of graphics programmers.
I hope to see more article like yours answering my article to facilitate the usage of more interesting conventional shadow filtering techniques
Cheers,
Holger