Jump to content

  • Log In with Google      Sign In   
  • Create Account


Algorithm for generating indices


Old topic!
Guest, the last post of this topic is over 60 days old and at this point you may not reply in this topic. If you wish to continue this conversation start a new topic.

  • You cannot reply to this topic
10 replies to this topic

#1 wh1sp3rik   Members   -  Reputation: 248

Like
0Likes
Like

Posted 17 January 2013 - 03:03 PM

Hello,

 

I coded a function, which takes as a input a vertex buffer. Output is a vertex buffer with removed duplications and index buffer.

It's working well, but the main problem of this function is, it's very slow. It contains three for loop and its complexity is roughly O( n^2 * log n ).

 

I just go though vertex array and adding index numbers, but when duplication is found, i delete it, i have to search, what index this duplication had and used it. Of course, i have to search whole vertex array again and again and search for duplication. ( i am comparing not only vertices, but normals, uvs too, they have to be same then i can mark it as duplication )

 

I would like to ask, if you know any technique, which is faster. I am searching the internet and i still did not find aynthing :(

 

thank you very much.


DirectX 11, C++


Sponsor:

#2 Álvaro   Crossbones+   -  Reputation: 12471

Like
3Likes
Like

Posted 17 January 2013 - 04:26 PM

You can do this in linear time. Use a hash to store the vertices you have seen so far and check if a vertex is new or not. Use something like the remove_if algorithm to keep only one copy of each vertex. Build a map from old indices to new indices as you go along, then replace all the indices.

#3 wh1sp3rik   Members   -  Reputation: 248

Like
0Likes
Like

Posted 18 January 2013 - 03:12 AM

Linear time, wau.

 

When i am building that hash and storing what i have seen so far, it's getting bigger and bigger and i have to search it to find a vertex if it's there or not, it doesn't look like linear. Perhaps i have to think about it more smile.png it will be nice to see some pseudocode.

 

thank you

 

edit: i just saw some small examples using hashes and i have a better view now

edit2: there can be another problem, most of verticles don't have to be exact same, but there can be small difference 0.001, so these will be "warped" into one.


Edited by wh1sp3rik, 18 January 2013 - 03:31 AM.

DirectX 11, C++


#4 Felix Ungman   Members   -  Reputation: 951

Like
1Likes
Like

Posted 18 January 2013 - 03:51 AM

It seems to me that if indexing would be a significant optimization, then you preferably would want some kind of shared-node structure where you don't need to post-process vertices. But that might not always be easy to do, depending of where the vertices are originated.


openwar  - the real-time tactical war-game platform


#5 wh1sp3rik   Members   -  Reputation: 248

Like
1Likes
Like

Posted 18 January 2013 - 04:02 AM

This is for my exporter to my model format. I would like to make indices and sort them, but my current solution .. is working, but it's not fast :)

 

I checked the hashing. it looks like, the main problem is, duplicates doesn't mean same values but values which are VERY close to be same ( for example two vertex very very very close to each other ).


DirectX 11, C++


#6 Chris_J_H   Members   -  Reputation: 256

Like
2Likes
Like

Posted 18 January 2013 - 04:02 AM

Suggestion: How about hashing by rounded floating point values.

Edited by Chris_J_H, 18 January 2013 - 04:08 AM.


#7 wh1sp3rik   Members   -  Reputation: 248

Like
0Likes
Like

Posted 18 January 2013 - 04:41 AM

Chris: Yep, that's good idea :)

 

I have no experience in hashing. Let's say, it's a function, where input gives you an index for an array.

How do i hash a struct, where are vertex positions, normals, uvs and colours? should i hash a part of memory, where struct is stored ?


DirectX 11, C++


#8 Álvaro   Crossbones+   -  Reputation: 12471

Like
0Likes
Like

Posted 18 January 2013 - 06:13 AM

You can change my algorithm into something that uses a k-d tree to look for the closes vertex and now I believe the algorithm is O(n*log(n)).

Merging vertices that are close is a messy problem, and one that I would be careful to specify much more precisely than you have in this thread. Rounding will work in many instances, because 9.631 and 9.633 will round to the same box; but 9.999 and 10.001 might round differently, even though they are just as close as the other two.

You have to think about the non-transitivity of the "is close to" relationship. If my model has a million equally-spaced vertices joining (0,0,0) and (10,10,10), do you meld them all into a single vertex, or partition them into groups?

#9 eppo   Crossbones+   -  Reputation: 2341

Like
1Likes
Like

Posted 18 January 2013 - 06:22 AM

Another method:

 

- Assign every vertex an index

- Create an array of vertex-count number of pointers to indices and set all pointers to null.

- Loop over all vertices in all triangles and create indices whenever a vertex' element in the index-array is still null -> assign new indices to the array



#10 wh1sp3rik   Members   -  Reputation: 248

Like
0Likes
Like

Posted 18 January 2013 - 07:33 AM

I have headache from it, lol.

 

It seems .. i will just drop a "duplicating vertex, when their distance is very small" feature ... graphic designer should be careful, where he puts verticles hehe.

Then .. i can use hash table and a great function called "IsVertexAlreadyInTable". That should help.


DirectX 11, C++


#11 wh1sp3rik   Members   -  Reputation: 248

Like
0Likes
Like

Posted 18 January 2013 - 07:48 AM

ok, i think, i know what to do :)

 

thank you all for your ideas.


DirectX 11, C++





Old topic!
Guest, the last post of this topic is over 60 days old and at this point you may not reply in this topic. If you wish to continue this conversation start a new topic.



PARTNERS