overlapping 3d particles not blending nicely when at the same z

Question

I'm working on a 3d particle emitter. It's going pretty good but I'm still having a problem with triangle sorting.

As you can see in the photos, there are some particles that are not blending properly presumably because of sorting. I do sort my quads before I draw them back to front. And they are all being drawn in a single glDrawElements(GL_TRIANGLES... call.

Blending is on & depth buffer is on. Depth buffer has to be on because of the other 3d objects in the scene that are sometimes in front and sometimes behind the particles.

When they're moving in a z space they're sorting better and you don't see it as much:

I'm using a sort with the following compare:

int compare (const void *a, const void *b) {
   Particle *ia = (Particle*)a;
   Particle *ib = (Particle*)b;

   float t = ib->distanceToCamera - ia->distanceToCamera;
   if (t<0) {
       return -1;
   }
   return (t>0);

}

But I think there are times when some particles have the same z and then they're not blending nicely.

One thought was to somehow force them to not be the same z. But with them moving on their own I'm not sure about that.

QUESTION - if overlapping triangles are close in z space but not equal, and still drawn back to front within the draw command, could they still overlap like this?

ANSWER - I set a break point and printed out all my structures to find that there are a ton of quads in the same z. In groups.

Using the advice from the checked answer I was able to fix it. By making sure to render the solids first and the particles last I don't need to worry about writing to the depth buffer while rendering the particles. In my game they're all like gases, smoke and air. So they all need to blend with each other and on to things deeper in the scene than them. When I start to render the particles I set:

glDepthFunc(GL_LEQUAL);
glDepthMask(GL_FALSE);

And when I'm done rendering them I return it to:

glDepthFunc(GL_LESS);
glDepthMask(GL_TRUE);

And look at the result:

What I believe was happening was that as particle triangles were drawn if they were on the same plane they were writing to the depth buffer and then the next one on that z if larger would get masked. By turning off depth mask it doesn't write to the mask as it's drawing the particles so they're not masking each other any more. Since I draw them last they don't need to mask other things.

UPDATE: Another benefit I just figured out. Now I don't have to sort the particles at all. The sort was only to make them draw properly with the depth test on. And that's no longer needed. So I have that savings as well.

Answer 1

The general rule of thumb when drawing alpha polys is:

1 - Draw all solid polys first.

2 - Sort back to front if you can. The main reason for this is to ensure that the final colour produced by the blending equation is consistent frame to frame. I often don't bother with this step unless it is something provided by the engine and I can justify the extra cost with my own eyes.

I don't think that this is the meat of the solution to your problem in this particular case as it looks like you have polys that are intersecting each other and in this case there is really no perfect way of sorting them bar sorting the individual fragments in a pixel shader or using some other order independent transparency technique (but that is usually massive overkill).

3 - Set the ztest renderstate to less equal or less

4 - Don't write to the z buffer by setting the zwriteenable renderstate to false

I suspect it is this last step that you are missing. You mention needing to have the z buffer turned on so that the alpha polys don't sort incorrectly with the solid but it isn't quite that simple.

You are able to control reading from and writing to the z buffer separately from each other, and in this case you want to read i.e. do the z test so that your polys sort correctly with the existing contents of the z buffer; but you don't want to write. Not writing to the z buffer will result in the next alpha poly essentially ignoring the previous one and you should see the seams disappear.

The downside of this approach is that the colour blending equation now produces slightly incorrect results for some fragments depending on the particular situation but there isn't a straightforward solution to this and generally the results are good enough.