So, work in progress.

[sharedmedia=gallery:images:8640]

I believe I've got filtering and brick building correct as of now (I've even checked the 3D texture and it seems correct to me. Anyways I'm using the sampling traversal as of now (which is incorrect, as it doesn't give me any advantage of using octree), but I had to switch back to it (for testing the filtering) due to having something wrong in the ray-octree one (which I'll need to update to cone-octree).

Whole SVO is built on GPU as of now, and the steps are following:

// Build a hierarchy (similar to how it is done in papers)
for (i = 0; i < levels; i++)
{
    FlagNodes();
    AllocNodes();
    InitNodes();
}

// Fill base level with data (as per your description, just done in 3D)
FillLeaves();
// Filter bricks (as we need neighbor information!)
FilterLeaves();

// Build interior nodes (interpolate from lower levels)
for (i = 0; i < levels; i++)
{
    BuildInterior(levels - i - 1);
}

Works like a charm, it is also very fast (I will do actual measurement shortly), and I am over-allocating memory a bit now (as I don't need log2(levels), but log2(levels)-1 depth of tree - as leaves (bricks) actually store one whole level in the 3D texture. I'm going to give figures for performance and memory out once I solve the traversal part (which I'm digging in now).

And I haven't even started optimization yet!

Thanks for advises in this thread. If everything is successful, then my next post will go into blog here on GameDev, with the figures and maybe some demo/code to show.

And no, I still don't have filtering right. I noticed that 2 items will cause problems, so the first one is filtering inside leaves:

Here is a pseudo-example:

This is our starting point - semi transparent nodes are visualization of part of the tree that is empty (therefore those nodes are not really there, and or course we can detect that there are empty nodes). Top left 4 nodes are full. Bottom left 4 nodes are also full (some with data that have alpha = 0).

I will demonstrate filtering along X-axis (as described in paper). So, the first step is:

Writing data to the right neighbor (of course as tree is sparse we can't really access the 'transparent' - non existing nodes). This will be the result:

Now, as the data in right-most voxels need to be the same as in the left-most voxels, we need to copy from right back to the left. Like:

And the problem is:

Obviously as we can't write into non-existent nodes (due to sparse nature of the tree), the values won't match (even when we assume that value was with alpha = 0 in the previous steps). All the data neighboring non-existent nodes will be invalid ending up in a border.

Of course the same problem is hit when computing interior nodes (mip mapped ones) - they will not match properly. Resulting in non-smooth ambient occlusion like this:

The problem with interpolation is, I can't do that beforehand. I have a set of elements from which I'm building a tree (those are voxels, generated for the scene). Now as I know the position of each voxel, I have no idea about its sorroundings (they can be in random order inserted into the set), all I know is their position and data (color, normal, whatever I wish to store). Adding voxels with interpolations would mean that I need to find neighbors in this set (which is actually easily done when octree building is finished).

This will be quite hard, as the tree can't be easily modified. EDIT: Thinking of this, yes technically it is possible to add nodes, removing them might be more tricky - but yes, also possible ... as actually each node represents sparse octree itself. I believe this might be a decent way to handle inserting dynamic objects too ... or maybe different levels of quality (whenever it is necessary). Making this easily usable even for case of 'ball in stadium'.

Each of the elements is inserted into the tree (the tree I generate is sparse), and doing any form of refinement will be hard, if not impossible once tree is generated (it uses parallel algorithm for that):

1. You start with root node (levels = 1)
2. Loop (for i = 0 to levels):
3. For each element - traverse into the 'levels', and flag which nodes are to be split
4. For each node in current level - if node is flagged, generate 8 child nodes
5. Allocate generated child nodes
6. Repeat the loop

Now, I can't really change a tree once it is finished. The point is, I realized that bricks are something different than what I thought they are. The value in the centre is the most important one, and 26 values sorrounding the center are duplicated (yes, it is "memory overkill" in this sense)! The correct version of previous images is (and should be):

The mipmapping part seemed tricky at first (but my algorithm is wrong), what I should actually do is, for child nodes, use their center values (and the values between the centres - which will be 3x3x3 total values! Not 4x4x4!) and average this into the parent node's center. Then put the boundaries to fill boundaries in the parent node.

Now, this will need additional step to perform the interpolations like I did for leaves, to guarantee that the border values in bricks are correct.