Until which stage should we stop doing this hierarchy? How fine grained this method can be? Is this a greedy algorithm or DFS something like that?
Bananya
Looks like we can build a quadtree, leaves would be pixels and the inner nodes would be a block. I guess it's also possible to control how fine it should be by setting up the max depth of the quadtree.
keenan
@peanut Probably you should just not use a hierarchy like this one unless you have a very special application in mind (e.g., images with features of very different scales). In real hardware, there is generally just a one level "hierarchy": coarse tiles, and individual samples. The reason is both that there is some overhead to evaluating the hierarchy, and perhaps more importantly, it leads to more irregular computation that becomes hard to schedule efficiently. (That being said, never say never: you may encounter some algorithm or setting where this is really the killer approach…)
peanut
Thanks professor!
But isn't hierarchy the better approach for this line? For this line, do we have better choice than hierarchy in real practice?
L100magikarp
The point-in tests that we talked about today are applied per triangle. But the meshes that were shown at the beginning of the lecture can have many many triangles. It seems like the incremental information computed for one triangle would be useful for other triangles, eg neighboring edges. Do modern implementations keep track of this information across triangles? Or do they mostly rely on parallel computation?
xiaol3
I think it is a very elegant method and I can see that the amount of computation required can be large. I am curious about how other approach reduce the overhead.
barath
Does the 'setup' for each hardware architecture, say Turing 2080 and an Ampere 3080, for the same picture above, will 'n' value, number of splits per iteration or something like that, be decided based on the hardware, or is it based on the application? say, Do two different games decide on the number of split size per iteration ?
Arthas007
The hierarchy approach looks like a divide and conquer algorithm in 15-210, very fascinating.
Isaaz
How does this methods perform compared to methods based on scanline? Especially in software rasterizer.
In my understanding in software rasterizer scanline seems more efficient because it will never check for pixels outside of the triangle.
Until which stage should we stop doing this hierarchy? How fine grained this method can be? Is this a greedy algorithm or DFS something like that?
Looks like we can build a quadtree, leaves would be pixels and the inner nodes would be a block. I guess it's also possible to control how fine it should be by setting up the max depth of the quadtree.
@peanut Probably you should just not use a hierarchy like this one unless you have a very special application in mind (e.g., images with features of very different scales). In real hardware, there is generally just a one level "hierarchy": coarse tiles, and individual samples. The reason is both that there is some overhead to evaluating the hierarchy, and perhaps more importantly, it leads to more irregular computation that becomes hard to schedule efficiently. (That being said, never say never: you may encounter some algorithm or setting where this is really the killer approach…)
Thanks professor!
But isn't hierarchy the better approach for this line? For this line, do we have better choice than hierarchy in real practice?
The point-in tests that we talked about today are applied per triangle. But the meshes that were shown at the beginning of the lecture can have many many triangles. It seems like the incremental information computed for one triangle would be useful for other triangles, eg neighboring edges. Do modern implementations keep track of this information across triangles? Or do they mostly rely on parallel computation?
I think it is a very elegant method and I can see that the amount of computation required can be large. I am curious about how other approach reduce the overhead.
Does the 'setup' for each hardware architecture, say Turing 2080 and an Ampere 3080, for the same picture above, will 'n' value, number of splits per iteration or something like that, be decided based on the hardware, or is it based on the application? say, Do two different games decide on the number of split size per iteration ?
The hierarchy approach looks like a divide and conquer algorithm in 15-210, very fascinating.
How does this methods perform compared to methods based on scanline? Especially in software rasterizer. In my understanding in software rasterizer scanline seems more efficient because it will never check for pixels outside of the triangle.