Can you update the link to the master thesis?
The old one seems broken, but this one seems to work fine:

https://pdfs.semanticscholar.org/1f23/57fc5a6102b745eb37ba1ab03a9c382de1f2.pdf

I’m glad you liked my site. Sadly I can’t update it for 2 reasons: 1) I don’t do much graphics progamming any more. Moved to robotics, and was, for a time, trying to write a CAD package for designing PCBs (www.liquidpcb.org). 2) I don’t have access to my own site any more, so I can’t add articles or fix errors. :(

Now you want to compute the lighting of these texels near the wall. However, the center of some of these texels lies outside of the floor polygon (behind the wall). The solution that I propose is to clip the texel quad against the floor polygon and instead of using the center of the quad, use the center of the clipped quad. That’s what’s usually known as centroid sampling and it guarantees that the sampling point is outside of the wall.

Note that on the GPU the centroid is computed approximately numerically by taking a fixed number of samples. In this case, since we don’t have overlaps in the parameterization, we can compute the centroid analytically.

you guys provably read it already, but it’s very informative and educational. really interesting..

http://www.astrolog.org/labyrnth/algrithm.htm

have fun making those trees behave and hemicubing stuff!

b/c it sounds as if it’s something revolutionary or something that is noth normal in games or that a lot of games don’t have.

sorry, i’m not a technical person so all that was like reading japanese or something… i’m bearly in my first year of computer science in high school so all the 3d space thing is unknown to me.

the pictures look very pretty, that castle like tower looks very cool and fantasy yet real like. do i even make sense? i remember a few years ago in the Braid blog that Jonathan was talking about “iconic landmarks” and he talked about how iconic things in games should be tied to interactivity (since that is what the meadium is supposed to be about, actual gameplay) with places like 2 fort and de-dust that encourtage sertain types of gameplays and then spaces that just look cool, like that “hell?” level in Painkiller.

i remembered this b/c the concept art of that tower and the windmill look super iconic and immediately recognisable. i wander how interactivity will be in those spaces…

good luck finding all the answers and algorithms and doing all the math, really best of wishes. and be sure to report back to tell us how you did it!

@Daniel: Irradiance Caching is the topic of the next post. The main reason why I implemented the integration on the CPU was that the computation of the irradiance gradients was complex enough to make it worth prototyping it on the CPU.

I thought about using NVIRT, since I’m somewhat familiar with it, but there are some disadvantages:
– Having to write and maintain surface shaders for all our materials.
– Dealing with texture limitations in NVIRT. We currently use a lot more textures than NVIRT supports.
– KD-Tree construction times used to be pretty slow, but maybe that has changed recently. The tracer would not be suitable for quick prototyping if you have to rebuild the hierarchy and wait a long time every time you move an object, and it might be too slow if every independent object uses a separate acceleration structure.

On the other side, it would also have other advantages… I would love to hear people’s experiences in this regard.

I’m a little bit ‘grossed out’ by the choice to use hemicubes on the GPU due to the inflexibility and CPU scene traversal bottlenecks instead of ray tracing which is trivially parallelizable and much more flexible (importance sampling), but it’s really useful to see all the stuff that’s involved with that approach. The async transfer in DX9 is especially interesting, no doubt I’ll find a use for that somewhere.

Since you are reading the data back to the CPU, is there a reason you can’t use irradiance caching to reduce the number of final gathers and speed up the process? Irradiance caching is also really useful for smoothing the results too because you can just interpolate in a second pass with larger thresholds than you used in the first pass when deciding where to place samples.