[Bf-committers] Does Cycles waste half it's possible performance or am i wrong?

[Tobias Oelgarte]

I uploaded a small illustration for my example and to illustrate the 
worst possible scenario for box filtering. I made the following 
assumptions for the worst case:

1. Despite normally random distribution all pixels are sampled at the 
right side, which would be even possible, but very seldom.
2. For the top left grid i made the assumption that every odd sample 
would return 1 and anything else 0, also possible and seldom. It also 
illustrates how the image would look in Cycles for this case.
3. For the bottom right grid i used area averaging, despite same odd 
sampling situation.

The implication of this sketch is, that the noise for 1 pass would not 
be reduced by Cycles. But if we would average the samples like in the 
bottom left illustration it would be divided by a factor of infinity, 
resulting in 0 for this special case and in factor 2 in average for a 
truly random distribution, since a sample would affect 2 pixels at a 
time in average. I i had chosen the grid intersections, to make it even 
worse, then this would be even a factor of 4, since four samples would 
contribute to a single pixel.

http://www.pasteall.org/pic/39642

Greetings Tobias