pouët.net

kravi, oh I see. Yep, what I do for primary rays is sphere tracing then, thx! Thx also for the suggestion of listening to windows events, I have 100 unused bytes, I will add it I think (even though inmy vista it seems to work all fine). Some answers in return: there is not any table for the scalar field, it's pure procedural and it's evaluated anytime it's needed (like during the raymarching or shading, or collision detection if it was a demo...). The monster is not done by sweeping a sphere. Tentacles are perlin noise deformed lines, and the "body" a sphere blended with those.
bonzaj, yippi, I will send it to you tonight or tomorrow.
thx to the others for the comments, it encourages me a lot to continue learning/improving :)

and big shit, I did it again, two times already - the self thumb up :(

calexico :: it uses indeed as many cores as you have (up to 128, tested on 16 cores). scheduling is based on breaking the screen in tiles.

psycholns :: nop, both glsl and hlsl compilers produce (apparently) similar code, since both compile in 20 seconds (more than 10 thousand instructions generated) and run at the very same speed (2 fps). also, both seem to unroll all the loops and expand all the functions calls (checked on the hlsl bytecode).

stan :: compo rules dissalowed any kind of progress bar. i personally think it's a good idea to add one to these productions though if one has the bytes for it.

M:ET :: oh, sceners never give up, i think what we see today is just the very beggining of the game.

kraviz :: i sampled the function at each step, yes; what i understood is that you were thinking i precomputed the function and stored the values in a 3d table/array or something, what is not the case. for the amount of function evaluations, i cannot say it now, i will have a look. tentacles, yep, distance to a line (segment). the ao is fake, i don't really sample the hemisphere as in a montecarlo pathtracer, but really evaluate the function in 3d points near the evaluation point as you say (10 samples), then do some magic to combine the distances into an occlusion distance. the trick for the soft shadows are similar (6 samples). lastly, there is no shortcuts for the function). we can discuss more by mail perhapse.

kraviz, I fixed the white window problem on your vista. The download link is the same. You might see the white window during some time but eventually the image will appear on the screen.

I quickly "formulanimated" a bit the scene and rendered a very short preview video of it. It's 4 seconds only, I don't think I will do more this time, better jump to other projects I guess.

So, after porting it to GLSL and HLSL, it ws now time for CUDA. Results are dissapointing. Even if the CUDA framework fits in 1k, the cuda compiler is expanding again all the function calls and unrolling loops :( So in the end it suffers the same problems and the two other GPU variants. Also, speed is not faster than GLSL, so... (CUDA might help in Ixaleno however, I should try). Jsyk !