Independent Study: Week 4

Slow progress this week because of early midterms and projects in other courses ramping up. I'm still well on track with the tentative schedule I put into the original proposal however, which is a good thing. I plan to make a lot of progress on the material / object systems in the next few days in preparation for the photon mapping which I hope to start implementing this weekend. For now, here's a video that shows that I can successfully integrate my quaternion camera that I developed for a previous project into this project with moderate ease:

I think what I would like to do for photon mapping is get a very basic implementation done initially that uses something like a 1-dimensional (or perhaps 2-dimensional) sorted array for storing the photon map. That way I can get an idea of what kind of performance I have base-line, and then work on optimization. Here's the original projected schedule that I included in my proposal:

Week 1: Final decisions on architectures and hardware to be used are made.
Weeks 2-3: Basic OpenCL / CUDA Ray-Caster implemented (trivial, no bouncing)
Weeks 4-6: OpenCL / CUDA photon-map computation is implemented
Weeks 7-8: Benchmarking and Profiling, bottlenecks are identified and addressed
Weeks 9-10: Finishing touches. Final project completed.
Week 11: Formal Writeup / Final Paper completed.



It's currently week 4, so as I said I'm still well on track. I've also left myself an additional 2 weeks to get this part done, but more importantly is the 2 weeks I've allocated for optimization following that. At the latest, that is when I hope to implement an acceleration structure such as a kd-tree so that I can get some high frame rates. One of the cool things about this approach is that while performance will be decreased by the photon mapping itself, it's overall limited by the ray-tracer's resolution during final gathering. That means that  my "real-time" photon mapper may not be real-time at 1280x720, but dropping it down to 640x480 or lower could still yield decent frame rates. That's sort of cheating however, and I'd rather just get an awesome implementation that runs well :).

Also, caustics are (temporarily) off the table on this project until I get a better idea of the performance boundaries I'm dealing with. I'd love to implement the technique if I have time and I can keep the speed up.




Update: 4/7/2012


...for loops...


Tried adding multiple lights to my scene, a single light that I illuminate with phong shading gets me around 200 FPS at 1280x720 resolution. I up that to two hard-coded lights: 160 FPS...not bad at all. I keep the same two lights, but write the code with a for-loop so that I can handle an arbitrary number of lights: 100 FPS! Just the loop iteration is significantly reducing my frame-rates. I'm not really shocked, but it still sucks. I tried manually unwinding the loop in code and the speed actually went down, so I'm guessing OpenCL already takes care of that to some extent during compilation. Oh well, seeing as everything about this project is supposed to be about indirect lighting, I shouldn't technically need more than one light in the scene anyway. I've resorted to a preprocessor flag, USE_MULTIPLE_LIGHTS that when defined makes the code use a for loop, otherwise it just uses the first light in the scene. I might come back to this later on and see if I can find a faster way to get these going but I just don't see how at the moment...

Independent Study: Week 3

Tried Cuda, hated it. Way too much work to do the same thing as OpenCL.

I'm sticking to the mentality that I don't know OpenCL well enough to be concerned about the framerates I'm getting. But I've learned a LOT in the last few days about how it works.

I've now got dynamic scene support, so I can add planes and spheres at will (excuse the overlap, I haven't implemented depth yet):

Update: (3/26/2012 @ 8:05 PM)

I got accurate depth and intersections done, the following is rendering surface normals at the points of intersection:

Update: (3/26/2012 @ 11:37 PM)

Phong shading integrated successfully:

Independent Study: Week 2

Red Dot:

Still a little quirky, lots of hard-coded stuff (namely the geometry in the world), but it's working!

Annnnnnd I'm probably going to have to start over with CUDA because of that performance thing I was complaining about. But it's not a sure thing yet.

Independent Study: Week 1

Quick update. I've got OpenCL interfacing with OpenGL and drawing a real nice gradient:

I'd be happier if the framerate was higher however. 800 FPS might sound great, but I know that's going to plummet when I actually start doing stuff in the OpenCL kernel code. Right now it's doing some floating point arithmetic and that's it. Also a similar procedural gradient as a shader in OpenGL would probably be hitting 3000 FPS on my test system so I'm not sure what's up there either. Definitely going to have to work on it.