I've been studying Nvidia's examples from the SDK, in particular the Island11 project and I've found something curious about a piece of HLSL code which corrects the reflections up and down depending on the state of the wave's height.
Naturally, after examining the brief paragraph of code:
// calculating correction that shifts reflection up/down according to water wave Y position float4 projected_waveheight = mul(float4(input.positionWS.x,input.positionWS.y,input.positionWS.z,1),g_ModelViewProjectionMatrix); float waveheight_correction=-0.5*projected_waveheight.y/projected_waveheight.w; projected_waveheight = mul(float4(input.positionWS.x,-0.8,input.positionWS.z,1),g_ModelViewProjectionMatrix); waveheight_correction+=0.5*projected_waveheight.y/projected_waveheight.w; reflection_disturbance.y=max(-0.15,waveheight_correction+reflection_disturbance.y);
My first guess was that it compensates for the planar reflection when it is subjected to vertical perturbation (the waves), shifting the reflected geometry to a point where is nothing and the water is just rendered as if there is nothing there or just the sky:
Now, that's the sky reflecting where we should see the terrain's green/grey/yellowish reflection lerped with the water's baseline. My problem is now that I cannot really pinpoint what is the logic behind it. Projecting the actual world space position of a point of the wave/water geometry and then multiplying by -.5f, only to take another projection of the same point, this time with its y coordinate changed to -0.8 (why -0.8?).
Clues in the code seem to indicate it was derived with trial and error because there is redundancy. For example, the author takes the negative half of the projected y coordinate (after the w divide):
And then does the same for the second point (only positive, to get a difference of some sort, I presume) and combines them:
By removing the divide by 2, I see no difference in quality improvement (if someone cares to correct me, please do). The crux of it seems to be the difference in the projected y, why is that? This redundancy and the seemingly arbitrary selection of -.8f and -0.15f lead me to conclude that this might be a combination of heuristics/guess work. Is there a logical underpinning to this or is it just a desperate hack?
Here is an exaggeration of the initial problem which the code fragment fixes, observe on the lowest tessellation level. Hopefully, it might spark an idea I'm missing. The -.8f might be a reference height from which to deduce how much to disturb the texture coordinate sampling the planarly reflected geometry render and -.15f might be the lower bound, a security measure.