I am in the process of implementing the geomipmapping algorithm, and I am now trying to solve the problems of the "cracks" (T-junctions or geometry gaps) who appears when two nearby patches have different LODs, it is advised in the paper to use "two triangle fans for this edge". A few questions to everyone who have already implemented this : 1) What kind of primitives do you use when rendering terrain ? I use triangle lists because triangle strips means making a lot of draw primitive quads for a single patchs (my patches are made of 32x32 quads), and triangle fans are not an option. 2) The paper advices not to render the edge of the higher level-patch (replacing it with the triangle fan). But in order not to render an edge, I have to split my DrawPrimitive calls for a patch, thus losing a lot of performance ! I don't get what I should do in this situation (of course rendering the edge's geometry twice is not an option) Thanks for your replies. EDIT: for those of you who don't know what geometry gaps are, have a look at (1) The cracks clearly visible on the terrain because of (2) the LOD variations [Edited by - bleyblue2 on October 28, 2005 4:58:22 AM]

Quote:1) What kind of primitives do you use when rendering terrain ? I use triangle lists because triangle strips means making a lot of draw primitive quads for a single patchs (my patches are made of 32x32 quads), and triangle fans are not an option.

My patches consist of one single indexed triangle strip. You'll have to hack in some extra vertices to make them all CW/CCW though. But it works for me, and it saves me a lot of memory and GPU bandwidth.

Quote:2) The paper advices not to render the edge of the higher level-patch (replacing it with the triangle fan). But in order not to render an edge, I have to split my DrawPrimitive calls for a patch, thus losing a lot of performance ! I don't get what I should do in this situation (of course rendering the edge's geometry twice is not an option)

Well, I don't know this paper, but I'll tell you what I did.
I rendered the patches based on their LOD but made sure that the difference in LOD between two neighbouring patches was never more or less than one level. By using a different list of vertex indices I made sure that the outer vertices (the edges) of the patch with the lowest LOD were all "divided" into two vertices so they would nicely fit next to vertices of the other patch.
This way I only needed a very simple system to avoid the cracks. I didn't need to fill the gaps with extra vertices (well, in a way I did, but they were already integrated in the patch) and I only needed to switch between list of indices without changing the vertices themselves.

I realize this is not the clearest of explanations, I really should provide some pictures with it. However, am not able to that right now. I can do it later, if you'd like/need that.

Thanks both for you replies - it's always a pleasure to have such quick answers.

Promit : I'm very interested in how you set up your triangle strips; maybe you could provide some code or an explanation ? As for the gaps "the paper" (I meant de Boer's paper, sorry not to have mentioned that, but I thought everyone implementing geomipmapping would have read it) advices not to add geometry, but rather to change the indices to omit some - still, I have to give you credit for your methode, esp. if you found it yourself, but I recommend you to read "Fast Terrain Rendering Using Geometrical MipMapping", Willem H. de Boer, where you might find some hints in how to improve your current implementation. Thanks a lot anyway !

Wanmaster : As stated before, I'm interested in the setup of the triangle strips, I expect you have to duplicate some vertices like Promit, but I recognize the cost might be worth it. Thanks you for re-explaining the paper's method, I understand it a lot more now. Still, feel free to post supplementary pictures, as it isn't the

Nevertheless, since you both seem to have successfully implemented a terrain rendering algorithm, I have to ask a few other questions :

In either way, you seem to be updating your indices for every patch who has changed LOD (or whose neighbor has) for each frame. Isn't the lock costly ? I don't think it's possible to do otherwise, though.
Wanmaster : when locking your Index Buffer, do you update all the indices or just the ones that have changed ? I'm speaking about the difference at the Lock() level, do you Lock the entire LOD-relative IB or do multiple locks for each index ? Do you keep the same IB from a patch to another (with the same LODs, of course) and modify it or do you create an IB for each "special" patch (i.e. patches whose cracks need to be fixed) ? Since there are also as many IBs as LODs, I suspect you might use the latter (creating all the IBs combinations at run time) as you say you switch between lists of indices but isn't it memory-expensive ? (since terrain is not the only thing I have to render, I'd like to keep it as lightweight as possible). Moreover the former (updating only parts of a single LOD-relative IB or using a single dynamic IB for each LOD, which would be modified whether the lower-level patch is on north, south, east...) seems pretty inefficient, but. To put it simple, which method are you using to create/modify your index buffers ?

As this is my last step toward a complete implementation of geomipmapping, I would appreciate any comments/advices from anyone.

Thanks in advance.

Promit : I have well understood the method you are using, thank you again ! (even though I will be doing like De Boer, because I do not have the time to measure the performance nor to implement every method). But maybe you could explain to me how you setup your triangle strips ? As for the IBs, by locking them I was implying modifications, and I have already enough to do with my fillrate, I want the terrain as lightweight as possible, etc. Could you explain what you mean by "caching" and how you code it though ? I agree on minimizing the LODs transitions, but only the user really decides...

JohnBolton : This skirt seems rather interesting, as it would be very simple to draw and code ! but I think the LOD changes would be too noticeable though - but maybe you could show me some screenshots/demos ? Thanks.

I just check to see if the neighboring chunk of terrain is farther from the viewer, and if it is at a lower LOD than the one I'm testing against. If it is, I hack in the extra verts/tris to stitch the terrain together. I am just GL_TRIANGLES'ing the triangles through the PIPE. Nothing fancy, but it runs great.

Wudan, you are using the same method as JohnBolton's, i.e. changing the terrain's geometry by adding triangles who don't follow the heightfield. It might be a good method, but I'm rather doing like De Boers and fill the gaps while allowing the terrain to follow the heightfield. Thanks anyway!

There is a very simple precalculation method, possibly mentioned before. The map should be divided into alternating big and small strips. The result should look like a rectangle grid on paper, with the lines being the small strips. First the detail levels for each big patch should be calulated. Then each small strip, that separates two big areas, should get 2 triangle lists for each detail level. One for continous detail and one for detail level change. During the rendering, the appropriate strip should be rendered.

Two examples:
normal: patch A (detail 1) - strip B (detail 1) - patch C (detail 1)
changing: patch A (detail 1) - strip B (detail 1to2) - patch C (detail 2)

The detail stacks:
patch A (1, 2, 3)
strip B (1, 1to2, 2, 2to3, 3)
patch C (1, 2, 3)

This way all detail levels can be calculated and stored before drawing, and no runtime updates are needed. If the detail changes more than one level in a single step, then the number of transitional strips per detail level would be equal to the maximal detail change allowed. This extra data is much less than the data required for storing the big patches, and precalculating and storing them will result in an improved performance. There is a very simple way to make the detail stacks, by first generating the whole terrain lods, then cutting out the big patches, and filling in the single triangle gaps with the strips. This algorithm only generates triangle strips or simple triangle lists.

Viktor