I have uploaded a demo of my realtime raytracer, Arauna: http://www.bik5.com/arauna_demo2.zip Instructions: 1. Unzip and run arauna_supersampling.exe to measure peak performance (~11M on my machine); 2. Open scene.txt and add '//' to the fifth line (speedtest); 3. Run arauna.exe to walk around the scene. Keys: A, S, D, W to walk forward/backward and strafe; cursor keys to rotate the camera; R and F to move down and up (reversed, sorry, forgot to fix). That's the basic stuff; there's more to try out using scene.txt, have a look at the file to see what's possible. The executables are optimized for architectural walkthroughs and perform bad for solo objects. Please post your timings! More info and ray tracing discussions: http://www.ompf.org/forum Greets Jacco Bikker EDIT: SSE2 required. Multiple cores supported.

Can you post rays per second instead of fps? It's more accurate.

Hi Jacco,

This is awesome! Performance on an Athlon 64 X2 4400+ @ 2.4 GHz: 15420K rays/s max.

The dual-core appears to add some shudder though, and the timings vary wildly. Would it be possible to add mouselook?

Cheers,

Nick

Edit: One more suggestion: with mipmapping the cache coherency is much higher and you could reach higher performance with little effort.

Nick: That's pretty good performance, close to what Conroe does. I'm still wondering though where Intel gets it's 100M figure from:

http://www.intel.com/technology/itj/2005/volume09issue02/art01_ray_tracing/p06_rt_performance.htm

I suppose it's a 10 triangle scene rendered at insane resolutions with tons of shadow rays or so, and no normalization of primary rays. The best I can do right now is 20M on a Conroe for a 50k or so scene, with one or two lights.

By the way, before the discussion starts: Ray tracing dynamic scenes is now also possible. Recently, papers where released on alternative data structures (e.g., BIH) that allow real-time rebuilds. Traversal for these scenes is slower than for kd-trees, but finally we can ray trace a fully dynamic scene of ~10k triangles @ 20fps for 512x384 pixels or so. Recent developments also seem to allow realtime construction of kd-trees; that would allow for even more complex scenes to be rendered in real-time.

Anyway, my demo at least proves that scenes of reasonable complexity (this one is 89000 triangles) can be ray traced; this allows for effects that are not easily doable using rasterizing hardware, e.g. reflection and refraction. If you edit scene.txt you can easily add a reflective sphere (mesh included in the demo). Diffuse reflections are also available, but very slow. Refraction is currently broken. Check here for a shot with refraction (while it was still working):



Jacco.

P4 3.4
Radeon X600
1Gb

arauna.exe 5825.05K Max
arauna_supersampling.exe 6617.45K Max

I don't see how you could do that. SSE2 operates on vectors of data, so one operation is applied to four data items. This is not the same as parallel code and cannot just be emulated using more registers.

Why does it require SSE2 in the first place? As far as I know SSE and MMX are perfect for processing vectors and colors.

Anyway, the extra registers offered by x64 could definitely create some new possibilities. By working on larger batches of data it's easier to hide instruction latencies.