MeshCompiler.cc

/*===========================================================================*\
 *                                                                           *
 *                              OpenFlipper                                  *
 *           Copyright (c) 2001-2015, RWTH-Aachen University                 *
 *           Department of Computer Graphics and Multimedia                  *
 *                          All rights reserved.                             *
 *                            www.openflipper.org                            *
 *                                                                           *
 *---------------------------------------------------------------------------*
 * This file is part of OpenFlipper.                                         *
 *---------------------------------------------------------------------------*
 *                                                                           *
 * Redistribution and use in source and binary forms, with or without        *
 * modification, are permitted provided that the following conditions        *
 * are met:                                                                  *
 *                                                                           *
 * 1. Redistributions of source code must retain the above copyright notice, *
 *    this list of conditions and the following disclaimer.                  *
 *                                                                           *
 * 2. Redistributions in binary form must reproduce the above copyright      *
 *    notice, this list of conditions and the following disclaimer in the    *
 *    documentation and/or other materials provided with the distribution.   *
 *                                                                           *
 * 3. Neither the name of the copyright holder nor the names of its          *
 *    contributors may be used to endorse or promote products derived from   *
 *    this software without specific prior written permission.               *
 *                                                                           *
 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS       *
 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED *
 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A           *
 * PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER *
 * OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,  *
 * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,       *
 * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR        *
 * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF    *
 * LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING      *
 * NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS        *
 * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.              *
 *                                                                           *
\*===========================================================================*/




#include "MeshCompiler.hh"

#include <iostream>
#include <sstream>

#ifdef USE_OPENMP
#endif

#ifdef ACG_MC_USE_STL_HASH
#include <unordered_map> // requires c++0x
#else
#include <QHash>  // alternative to unordered_map
#endif // ACG_MC_USE_STL_HASH

#include <ACG/Geometry/GPUCacheOptimizer.hh>
#include <ACG/Geometry/Triangulator.hh>

namespace ACG{

/*
use case


mesh with float3 pos, float2 uv, float3 n:

VertexElement elements[] = {{..}, ..};
VertexDeclaration decl;

decl.addElement(..float3 position..)
decl.addElement(..float2 uv..)
decl.addElement(..float3 normal..)

MeshCompiler mesh(decl);

mesh.setVertices(..)
mesh.setNormals(..)
mesh.setTexCoords(..)


mesh.setNumFaces(333);
mesh.setTriangles(indexBuffer);
for each face i
  mesh.setFaceMaterial(i, matID)

mesh.compile(FLAG_OPTIMIZE | FLAG_COMPUTE_TANGENTS | FLAG_STATIC)
*/


/* TODO
- default declaration (declaration can not be changed afterwards)
   float3 pos, float2 uv, float3 normal

- option to set triangle input buffer

- compile function with options
   FLAG_OPTIMIZE         : optimize for better gpu cache usage
   FLAG_COMPUTE_TANGENTS : compute tangent vectors if tangent slots are available in buffer layout
   FLAG_STATIC           : no update needed in future
*/


MeshCompilerVertexCompare MeshCompiler::defaultVertexCompare;

void MeshCompiler::AdjacencyList::init( int n )
{
  delete [] start;
  delete [] buf;
  delete [] count;

  num = n;
  start = new int[n];
  count = new unsigned char[n];

  buf   = 0; // unknown buffer length

  // invalidate start indices
  memset(start, -1, n * sizeof(int));

  // reset count
  memset(count, 0, n * sizeof(unsigned char));  
}

int MeshCompiler::AdjacencyList::getAdj( const int i, const int k ) const
{
  assert(k < count[i]);
  assert(k > -1);

  int st = start[i];
  assert(st > -1);

  return buf[st + k];
}

int MeshCompiler::AdjacencyList::getCount( const int i ) const
{
  return count[i];
}

void MeshCompiler::AdjacencyList::clear()
{
  num = 0;
  delete [] start; start = 0;
  delete [] buf; buf = 0;
  delete [] count; count = 0;
  bufSize = 0;
}



void MeshCompiler::computeAdjacency(bool _forceRecompute)
{
  const int numVerts = input_[inputIDPos_].count;

  // ==============================================================
  // compute vertex -> face adjacency

  // count total number of adjacency entries
  // store adj entries in a single tightly packed buffer

  // check if user provided adjacency information
  if (_forceRecompute || (faceInput_->getVertexAdjCount(0) < 0 && adjacencyVert_.bufSize <= 0))
  {
    adjacencyVert_.init(numVerts);

    // count # adjacent faces per vertex
    for (int i = 0; i < numFaces_; ++i)
    {
      int nCorners = getFaceSize(i);

      for (int k = 0; k < nCorners; ++k)
      {
        //        int vertex = faceInput_->getSingleFaceAttr(i, k, inputIDPos_);
        int vertex = getInputIndex(i, k, inputIDPos_);

        adjacencyVert_.count[vertex]++;
      }
    }


    // count num of needed entries
    int nCounter = 0;

    for (int i = 0; i < numVerts; ++i)
    {
      adjacencyVert_.start[i] = nCounter; // save start indices

      nCounter += adjacencyVert_.count[i];

      adjacencyVert_.count[i] = 0; // count gets recomputed in next step
    }

    // alloc memory
    adjacencyVert_.buf = new int[nCounter];
    adjacencyVert_.bufSize = nCounter;

    // build adjacency list
    for (int i = 0; i < numFaces_; ++i)
    {
      int nCorners = getFaceSize(i);

      for (int k = 0; k < nCorners; ++k)
      {
        //        int vertex = faceInput_->getSingleFaceAttr(i, k, inputIDPos_);
        int vertex = getInputIndex(i, k, inputIDPos_);
        int adjIdx = adjacencyVert_.start[vertex] + adjacencyVert_.count[vertex]++;

        adjacencyVert_.buf[ adjIdx ] = i;
      }
    }

    // debug version:
    // dump computed and external adjacency for comparison
//     dbgdumpAdjList("dbg_adjacency_mc.txt");
// 
//     if (faceInput_->getVertexAdjCount(0) >= 0)
//     {
//       FILE* file = 0;
//       file = fopen("dbg_adjacency_ext.txt", "wt");
// 
//       if (file)
//       {
//         fprintf(file, "vertex-adjacency: \n");
//         for (int i = 0; i < input_[inputIDPos_].count; ++i)
//         {
//           // sorting the adjacency list for easy comparison of adjacency input
//           int count = faceInput_->getVertexAdjCount(i);
// 
//           std::vector<int> sortedList(count);
//           for (int k = 0; k < count; ++k)
//             sortedList[k] = faceInput_->getVertexAdjFace(i, k);
// 
//           std::sort(sortedList.begin(), sortedList.end());
// 
//           for (int k = 0; k < count; ++k)
//             fprintf(file, "adj[%d][%d] = %d\n", i, k, sortedList[k]);
//         }
// 
// 
//         fclose(file);
//       }
//     }
    
  }
    
}


void MeshCompiler::setVertices( size_t _num, const void* _data, size_t _stride, bool _internalCopy /*= false*/, GLuint _fmt, int _elementSize )
{
  setAttribVec(inputIDPos_, _num, _data, _stride, _internalCopy, _fmt, _elementSize);
}

void MeshCompiler::setNormals( size_t _num, const void* _data, size_t _stride, bool _internalCopy /*= false*/, GLuint _fmt, int _elementSize )
{
  setAttribVec(inputIDNorm_, _num, _data, _stride, _internalCopy, _fmt, _elementSize);
}

void MeshCompiler::setTexCoords( size_t _num, const void* _data, size_t _stride, bool _internalCopy /*= false*/, GLuint _fmt, int _elementSize )
{
  setAttribVec(inputIDTexC_, _num, _data, _stride, _internalCopy, _fmt, _elementSize);
}


void MeshCompiler::setAttribVec(int _attrIdx, size_t _num, const void* _data, size_t _stride, bool _internalCopy /*= false*/, GLuint _fmt, int _elementSize)
{
  // sets vertex data for each attribute individually
  // Example:
  // Format: float3 pos, float2 uv, float3 normal has 3 attribute vectors, that store all positions, texcoords etc. of the mesh
  // - positions: float3[n1]  n1: number of positions
  // - texcoords: float2[n2]  n2: number of texcoords
  // - normals: float3[n3]    n3: number of normals
  //  the vector sizes n1, n2, n3 do not have to be equal!

  if (_attrIdx < 0)
    return;

  assert(_attrIdx < (int)decl_.getNumElements());

  VertexElementInput* inbuf = input_ + _attrIdx;

  inbuf->count = _num;

  // size in bytes of one vertex element (eg. 12 for float3 element)
  int size = inbuf->attrSize = (int)VertexDeclaration::getElementSize(decl_.getElement(_attrIdx));


  // make an internal copy of the input if the user can not guarantee, that the data address is permanently valid
  if (_internalCopy)
  {
    delete [] inbuf->internalBuf;
    inbuf->internalBuf = new char[size * _num];
    inbuf->data = inbuf->internalBuf;

    inbuf->stride = size;

    if (_data)
    {
      // copy elementwise because of striding
      for (size_t i = 0; i < _num; ++i)
      {
        memcpy(inbuf->internalBuf + (size_t)(size * i),
          (const char*)_data + (size_t)(_stride * i),
          size);
      }
    }
  }
  else
  {
    // store data address only without making a copy
    inbuf->data = (const char*)_data;
    inbuf->stride = _stride ? _stride : size;

    delete [] inbuf->internalBuf;
    inbuf->internalBuf = 0;
  }


  inbuf->fmt = _fmt;
  inbuf->elementSize = _elementSize;
}

void MeshCompiler::WeldList::add(const int _face, const int _corner)
{
  const int stride = meshComp->getVertexDeclaration()->getVertexStride();

  // pointer address to vertex data
  char* vtx0 = workBuf + stride * list.size();
  char* vtx1 = workBuf;

  // query vertex data
  meshComp->getInputFaceVertexData(_face, _corner, vtx0);

  bool matched = false;

  // search for same vertex that is referenced already
  for (size_t i = 0; i < list.size() && !matched; ++i)
  {
    WeldListEntry* it = &list[i];
    // query referenced vertex data

    // compare vertices
    if (cmpFunc->equalVertex(vtx0, vtx1, meshComp->getVertexDeclaration()))
    {
      // found duplicate vertex
      // -> remap index data s.t. only one these two vertices is referenced

      WeldListEntry e;
      e.faceId = _face; e.cornerId = _corner;
      e.refFaceId = it->refFaceId; e.refCornerId = it->refCornerId;

      list.push_back(e);

      matched = true;
    }

    vtx1 += stride;
  }

  // unreferenced vertex
  if (!matched)
  {
    WeldListEntry e;
    e.faceId = _face; e.cornerId = _corner;
    e.refFaceId = _face; e.refCornerId = _corner;

    list.push_back(e);
  }

}

void MeshCompiler::weldVertices()
{
  const int numVerts = input_[inputIDPos_].count;

  // clear weld map
  vertexWeldMapFace_.resize(numIndices_, -1);
  vertexWeldMapCorner_.resize(numIndices_, -1);

  // alloc buffer to query vertex data
  int maxAdjCount = 0;
  for (int i = 0; i < numVerts; ++i)
  {
    const int n = getAdjVertexFaceCount(i);
    maxAdjCount = std::max(n, maxAdjCount);
  }

  // OPTIMIZATION: Now the vertex compare buffer works as a vertex cache, storing interpreted vertex data.
  //  Thus, each vertex has to be interpreted only once ( when it gets inserted into the welding list )
  char* vtxCompBuf = new char[decl_.getVertexStride() * (maxAdjCount + 1)];

  WeldList weldList;
  weldList.meshComp = this;
  weldList.cmpFunc = vertexCompare_;
  weldList.workBuf = vtxCompBuf;

  weldList.list.reserve(maxAdjCount);

  bool retry = false;
  
  for (int i = 0; i < numVerts; ++i)
  {
    // OPTIMIZATION: Moved constructor/destructor of WeldList out of for-loop
    // Create welding list for each vertex.
    weldList.list.clear();

    // Search for candidates in adjacent faces
    int numAdjFaces = getAdjVertexFaceCount(i);
    for (int k = 0; k < numAdjFaces; ++k)
    {
      const int adjFace = getAdjVertexFace(i, k);

      // find corner id of adj face
      int adjFaceSize = getFaceSize(adjFace);
      int adjCornerId = -1;
      for (int m = 0; m < adjFaceSize; ++m)
      {
        const int adjVertex = getInputIndex(adjFace, m, inputIDPos_);
        
        if (adjVertex == i)
        {
          adjCornerId = m;
          break;
        }
      }

      if (adjCornerId < 0)
      {
        // user provided adjacency is faulty
        // retry with internally compute adjacency
        retry = true;
        break;
      }
      else
      {
        // check for existing entry
        const int weldMapOffset = getInputFaceOffset(adjFace) + adjCornerId;

        if (vertexWeldMapFace_[weldMapOffset] >= 0)
          continue; // skip

        weldList.add(adjFace, adjCornerId);
      }
    }


    // apply local WeldList of a vertex to global weld map
    if (!retry)
    {
      for (size_t e = 0; e < weldList.list.size(); ++e)
      {
        const WeldListEntry* it = &weldList.list[e];
        const int weldMapOffset = getInputFaceOffset(it->faceId) + it->cornerId;

        if (vertexWeldMapFace_[weldMapOffset] >= 0)
          continue; // skip

        // store in map
        vertexWeldMapFace_[weldMapOffset] = it->refFaceId;
        vertexWeldMapCorner_[weldMapOffset] = it->refCornerId;
      }
    }
    else
      break;
  }




  // -------------------------------------------------------------
  // Alternative method that avoids iterating over adjacency list at cost of higher memory load
  // Could not measure any noticeable difference in run-time performance.
// 
//   std::vector< std::vector< std::pair<int,int> > > VertexColMap;
//   VertexColMap.resize(numVerts);
// 
//   for (int i = 0; i < numVerts; ++i)
//     VertexColMap[i].reserve( getAdjVertexFaceCount(i) );
// 
//   for (int i = 0; i < numFaces_; ++i)
//   {
//     for (int k = 0; k < getFaceSize(i); ++k)
//     {
//       int v = faceInput_->getSingleFaceAttr(i, k, inputIDPos_);
// 
//       VertexColMap[v].push_back( std::pair<int,int>(i,k) );
//     }
//   }
// 
//   for (int i = 0; i < numVerts; ++i)
//   {
//     // Create welding list for each vertex.
//     WeldList weldList;
//     weldList.meshComp = this;
//     weldList.cmpFunc = vertexCompare_;
//     weldList.workBuf = vtxCompBuf;
// 
//     for (int k = 0; k < VertexColMap[i].size(); ++k)
//       weldList.add(VertexColMap[i][k].first, VertexColMap[i][k].second);
// 
//     // apply local WeldList of a vertex to global weld map
// 
// //     for (std::list< WeldListEntry >::iterator it = weldList.list.begin();
// //       it != weldList.list.end();  ++ it)
//     for (size_t e = 0; e < weldList.list.size(); ++e)
//     {
//       const WeldListEntry* it = &weldList.list[e];
//       const int weldMapOffset = getInputFaceOffset(it->faceId) + it->cornerId;
// 
// //      if (vertexWeldMap_[weldMapOffset].first >= 0)
//       if (vertexWeldMapFace_[weldMapOffset] >= 0)
//         continue; // skip
// 
//       // store in map
// //      vertexWeldMap_[weldMapOffset] = std::pair<int, int> ( it->refFaceId, it->refCornerId );
//       vertexWeldMapFace_[weldMapOffset] = it->refFaceId;
//       vertexWeldMapCorner_[weldMapOffset] = it->refCornerId;
//     }
//   }
//   // -------------------------------------------------------------

  if (retry)
  {
    if (adjacencyVert_.bufSize <= 0)
    {
      // there is an issue with the external vertex-face adjacency list provided with the input interface

      // 1. compute internal adjacency list and use that instead
      computeAdjacency(true);

      // 2. rollback welding progress

      vertexWeldMapFace_.clear();
      vertexWeldMapCorner_.clear();

      // 3. try again with updated adjacency
      weldVertices();
    }
    else
    {
      // something went badly wrong..
      std::cerr << "MeshCompiler - faulty internal adjacency list" << std::endl;
    }
  }
  else
  {
    // fix incomplete welding map (isolated vertices)
    fixWeldMap();
  }

  delete [] vtxCompBuf;
}

void MeshCompiler::fixWeldMap()
{
  for (int i = 0; i < numFaces_; ++i)
  {
    const int fsize = getFaceSize(i);
    for (int k = 0; k < fsize; ++k)
    {
      const int weldMapOffset = getInputFaceOffset(i) + k;

      // if a (face,corner) pair is mapped to an invalid value, make it valid by mapping to itself
      // invalid value is caused by isolated vertices
      if (vertexWeldMapFace_[weldMapOffset] < 0)
      {
        vertexWeldMapFace_[weldMapOffset] = i;
        vertexWeldMapCorner_[weldMapOffset] = k;
      }
    }
  }
}

void MeshCompiler::findIsolatedVertices()
{
  const int nVerts = input_[inputIDPos_].count;

  numIsolatedVerts_ = 0;
  // For each vertex check if there exists a reference in the splitting list. We have found an isolated vertex if there is no reference.
  //  Checking the vertex-face adjacency count is also possible to detect isolated vertices.

  for (int i = 0; i < nVerts; ++i)
  {
    if (splitter_->isIsolated(i))
      ++numIsolatedVerts_;
  }

  isolatedVertices_.clear();
  isolatedVertices_.reserve(numIsolatedVerts_);
  for (int i = 0; i < nVerts; ++i)
  {
    if (splitter_->isIsolated(i))
      isolatedVertices_.push_back(i);
  }
}

void MeshCompiler::splitVertices()
{
  /* algorithm overview

  we split by indices only,
  actual vertex data will not be taken into account.

  thus if the input contains two vertices with the same value,
  they still are treated as different vertices since they have different indices


  a shared vertex gets split whenever at least one attribute
  changes with respect to the face.
  i.e. one face wants to combine vertex i with attribute j
  while another face needs it with attribute k

  example:

  face 0 and 5 share vertex v3

  face 0 uses v3 combined with normal n0,
  but face 5 wants to have v3 with normal n1


  we look up if v3 has been split up already and search
  for a v3-n0 combination
  eventually this combination is added to the splitting list

  repeat for the v3-n1 combination

  */

  const int numPositions = input_[inputIDPos_].count;

  // estimate number of splits to avoid resizing too often
  int estimatedSplitCount = 0;
  int numDifferentInputCounts = 0;

  // simple heuristic:
  // if the number of elements of an arbitrary attribute is larger than the number of positions, add the difference to the number of splits
  //  the actual number of splits may be larger, in which case the array has to be resized.
  for (int i = 0; i < numAttributes_; ++i)
  {
    if (i != inputIDPos_)
    {
      if (input_[i].count > numPositions)
      {
        const int diff = input_[i].count - numPositions;

        if (diff > 0)
        {
          estimatedSplitCount = std::max(diff, estimatedSplitCount);
          ++numDifferentInputCounts;
        }
      }
    }
  }

  if (numDifferentInputCounts > 1)
  {
    // estimation probably too small, increase by 20 %
    estimatedSplitCount = int(float(estimatedSplitCount) * 1.2f);
  }

  assert(estimatedSplitCount >= 0);

  // worst case: each vertex can be used by only one face
  //  clamp estimation-count accordingly

  int maxSplitCount = 0;

  if (numIndices_ > 0)
  {
    if (numIndices_ > numPositions)
      maxSplitCount = numIndices_ - numPositions;
  }
  else
  {
    // numIndices_ is unknown
    int sumFaceSize = 0;

    if (constantFaceSize_)
      sumFaceSize = numFaces_ * maxFaceSize_;
    else
    {
      for (int i = 0; i < numFaces_; ++i)
        sumFaceSize += getFaceSize(i);
    }

    if (sumFaceSize > numPositions)
      maxSplitCount = sumFaceSize - numPositions;
  }

  estimatedSplitCount = std::min(estimatedSplitCount, maxSplitCount);

//  std::cout << "estimated split count: " << estimatedSplitCount << std::endl;

  // split vertices such that each index combination of a face corner (i_pos, i_uv, i_attr..) has a unique vertex id and the number of vertices is minimal
  delete splitter_;
  splitter_ = new VertexSplitter(decl_.getNumElements(),
                                 numPositions,
                                 numPositions + estimatedSplitCount,
                                 0.0f);

  faceBufSplit_.resize(numIndices_, -1);

  // count # vertices after splitting
  numDrawVerts_ = 0;

  for (int i = 0; i < numFaces_; ++i)
  {
    const int fsize = getFaceSize(i);
    for (int k = 0; k < fsize; ++k)
    {
      // indices of the face vertex into the attribute vectors
      int vertex[16];  // {i_pos, i_attr1, i_attr2 ..}


      // get indices of a face corner after welding
      getInputFaceVertex_Welded(i, k, vertex);

      // split vertices by index data only
      // value of position, normal etc. are not considered
      const int idx = splitter_->split(vertex);

      // handle index storage
      setInputIndexSplit(i, k, idx);
    }
  }


//  std::cout << "actual split count: " << (numDrawVerts_ - numPositions) << std::endl;



  // Fix splitting list if there are isolated vertices in between.
  // Isolated vertices currently occupy spots in in the interleaved vertex buffer.
  // -> Remove them from the vbo.
  findIsolatedVertices();

  if (numIsolatedVerts_ > 0)
  {
    // create table that stores how many isolated vertices have been encountered up to each vertex
    //  this is done in the index domain after splitting
    std::vector<int> IsoFix(splitter_->numVerts, 0);

    int fixIndex = 0;
    for (int i = 0; i < splitter_->numVerts; ++i)
    {
      if (splitter_->isIsolated(i))
        fixIndex--;

      IsoFix[i] = fixIndex;
    }

    // IsoFix[] array contains offsets <= 0 for each split vertex id.
    // It maps from such an id to the vbo index, which does not contain any isolates.
    // Isolates may be appended later to the vbo if the user wants that.

    numDrawVerts_ = 0;

    // apply index fixing table to current vertex ids
    for (int i = 0; i < numFaces_; ++i)
    {
      const int fsize = getFaceSize(i);
      for (int k = 0; k < fsize; ++k)
      {
        // get interleaved vertex id for (i, k) after splitting
        int idx = getInputIndexSplit(i, k);

        // idx is the split vertex id of (i, k)
        // IsoFix[idx] is the offset that has to be applied to that
        idx += IsoFix[idx];

        // store fixed vertex id
        setInputIndexSplit(i, k, idx);
      }
    }
  }
}


bool MeshCompiler_forceUnsharedFaceVertex_InnerValenceSorter( const std::pair<int, int>& a, const std::pair<int, int>& b )
{
  return a.second > b.second;
}

void MeshCompiler::forceUnsharedFaceVertex()
{
  // ==============================================
  //  face normal fix
  //  make sure that each triangle has at least one unique unshared vertex
  //  this vertex can store per face attributes when needed


  // sharedVertex[i] = 1 iff the vertex id of corner i is shared with any neighboring face
  // sharedVertex is computed on-the-fly for each face
  std::vector<int> sharedVertices;
  sharedVertices.resize(maxFaceSize_);

  // temporary copy of the vertex ids of a face
  std::vector<int> tmpFaceVerts; // used for face rotation-swap
  tmpFaceVerts.resize(maxFaceSize_);

  int numInitialVerts = numDrawVerts_;
  std::vector<int> VertexUsed(numDrawVerts_, -1); // marks vertices which are not shared with any neighboring face


  // process all n-polygons first

  /*
  new and better algorithm:   O(n * m^2) where n = numFaces, m = faceSize

  for each face:

    trisCovered = 0;

    while (trisCovered < faceSize)
    {
      compute inner valence of all corners in the remaining polygon

      add 'best' corner: - highest inner valence and unused by other tris

      for each triangle affected by this corner
        rotate triIndexBuffer entries of the tri
        remove tri from the remaining triangle list
        ++ trisCovered
    }
  */
  int triCounter = 0;


  for (int sortFaceID = 0; sortFaceID < numFaces_; ++sortFaceID)
  {
    // get original face id
    const int faceID = faceSortMap_.empty() ? sortFaceID : faceSortMap_[sortFaceID];

    const int faceSize = getFaceSize(faceID);

    if (faceSize > 3)
    {
      // vertexPriorities[priority] = pair(cornerID, valence)
      std::vector< std::pair<int, int> > vertexPriorities(faceSize);

      // linked ring list for all the triangles in the uncovered triangulation
      // ie. nextTri = remainingTris[currentTri];
      const int faceTris = faceSize - 2;

      std::vector<RingTriangle> remainingTris(faceTris);
      for (int i = 0; i < faceTris; ++i)
        remainingTris[i] = RingTriangle(i, &remainingTris[(i + faceTris - 1) % faceTris], &remainingTris[(i + 1) % faceTris]);


      RingTriangle* currentTri = &remainingTris[0];
      int numTrisCovered = 0;

      while (numTrisCovered < faceTris)
      {
        // compute valence of vertices within the remaining triangulation
        for (int k = 0; k < faceSize; ++k)
          vertexPriorities[k] = std::pair<int, int>(k, 0);

        RingTriangle* startTri = currentTri;
        int numRemainingTris = faceTris - numTrisCovered;
        for (int t = 0; t < numRemainingTris; ++t)
        {
          for (int k = 0; k < 3; ++k)
          {
            int cornerID = -1 - triIndexBuffer_[(triCounter + currentTri->id) * 3 + k];
            ++vertexPriorities[cornerID].second;
          }
          currentTri = currentTri->next;
        }
        assert(currentTri == startTri);

        // sort by valence
        std::sort(vertexPriorities.begin(), vertexPriorities.end(), MeshCompiler_forceUnsharedFaceVertex_InnerValenceSorter);

        // find a good corner
        int goodCorner = -1;
        int goodVertexID = -1;
        int bestValence = -1;
        for (int k = 0; k < faceSize && vertexPriorities[k].second; ++k)
        {
          int cornerID = vertexPriorities[k].first;
          int vertexID = getInputIndexSplit(faceID, cornerID);

          int valence = vertexPriorities[k].second;

          if (vertexID >= numInitialVerts || (VertexUsed[vertexID] == faceID))
          {
            // best case, this vertex is already owned by the polygon
            // stop the search
            goodCorner = cornerID;
            goodVertexID = vertexID;
            bestValence = valence;
            break;
          }
          else if (VertexUsed[vertexID] < 0 && bestValence < valence)
          {
            goodCorner = cornerID; // best for now, but continue the search
            goodVertexID = vertexID;
            bestValence = valence;
          }
        }


        // maybe add a new vertex
        if (goodCorner < 0)
        {
          // have to add a new vertex
          // use the one with highest inner valence

          goodCorner = vertexPriorities[0].first; // polygon corner

          // add new vertex at the end of the buffer
          goodVertexID = numDrawVerts_;
          setInputIndexSplit(faceID, goodCorner, goodVertexID);
        }
        else
        {
          // mark the polygon as owner of the vertex
          VertexUsed[goodVertexID] = faceID;
        }

        // process tris
        for (int t = 0; t < numRemainingTris; ++t)
        {
          // check if the triangle references the good corner by testing the 3 vertices of the triangulation
          for (int k = 0; k < 3; ++k)
          {
            int cornerID = -1 - triIndexBuffer_[(triCounter + currentTri->id) * 3 + k];

            if (cornerID == goodCorner)
            {
              // rotate the triangle such that the first corner of the triangle references the good corner
              int rotCount = 3 - k;

              // make a temp copy of current triangle
              int tmpTriVerts[3] =
              {
                triIndexBuffer_[(triCounter + currentTri->id) * 3],
                triIndexBuffer_[(triCounter + currentTri->id) * 3 + 1],
                triIndexBuffer_[(triCounter + currentTri->id) * 3 + 2],
              };

              // apply rotation
              for (int i = 0; i < 3; ++i)
                triIndexBuffer_[(triCounter + currentTri->id) * 3 + (i + rotCount) % 3] = tmpTriVerts[i];


              ++numTrisCovered;

              // remove triangle from the ring list
              currentTri->prev->next = currentTri->next;
              currentTri->next->prev = currentTri->prev;
              break;
            }
          }

          currentTri = currentTri->next;
        }
      }

    }

    triCounter += faceSize - 2;
  }

  // process all triangles now
  numInitialVerts = VertexUsed.size();
  triCounter = 0;

  for (int sortFaceID = 0; sortFaceID < numFaces_; ++sortFaceID)
  {
    int faceID = faceSortMap_.empty() ? sortFaceID : faceSortMap_[sortFaceID];
    const int numCorners = getFaceSize(faceID);
    
    if (numCorners == 3)
    {
      // reset shared list
      memset(&sharedVertices[0], 0, sizeof(int) * maxFaceSize_);
      int numShared = 0;

      // find shared list (corners of this face, that are shared with the neighbors)
      for (int v0 = 0; v0 < numCorners && numShared < numCorners; ++v0)
      {
        if (sharedVertices[v0])
          continue;

        const int vertexID0 = getInputIndexSplit(faceID, v0);

        // EDIT:
        // vertexID0 >= numInitialVerts || (...) seemed wrong
        if (vertexID0 < numInitialVerts && (VertexUsed[vertexID0] >= 0 && VertexUsed[vertexID0] != faceID))
        {
          sharedVertices[v0] = true;
          ++numShared;
        }
      }

      int rotCount = 0;

      if (numShared == numCorners)
      {
        // worst-case: all vertices shared with neighbors

        // add split vertex to end of vertex buffer, which is used exclusively by the current face
        // current vertex count is stored in numDrawVerts_

        setInputIndexSplit(faceID, 0, numDrawVerts_);
      }
      else if (sharedVertices[0])
      {
        // validation code
        int x = 0;
        for (int i = 0; i < numCorners; ++i)
          x += sharedVertices[i];
        assert(x < numCorners);

        // we have to make sure that an unshared vertex is the first referenced face vertex
        // this is currently not the case, so rotate the face indices until this is true

        // make copy of current face splitVertexID
        for (int i = 0; i < numCorners; ++i)
          tmpFaceVerts[i] = getInputIndexSplit(faceID, i);

        // rotation order: i -> i+1
        // find # rotations needed
        rotCount = 1;
        for (; rotCount < numCorners; ++rotCount)
        {
          if (!sharedVertices[rotCount % numCorners])
          {
            if (tmpFaceVerts[rotCount] < numInitialVerts)
              VertexUsed[tmpFaceVerts[rotCount]] = faceID;
            break;
          }
        }

        assert(rotCount < numCorners);

        // rotate:  i -> i+rotCount
        rotCount = numCorners - rotCount;

        for (int i = 0; i < numCorners; ++i)
        {
//        setInputIndexSplit(faceID, i, tmpFaceVerts[(i + numCorners - rotCount) % numCorners]);

          triIndexBuffer_[triCounter * 3 + i] = tmpFaceVerts[(i + numCorners - rotCount) % numCorners];
        }
      }
      else
      {
        // best-case: unshared vertex at corner 0
        const int idx = getInputIndexSplit(faceID, 0);
        if (idx < numInitialVerts)
          VertexUsed[idx] = faceID;
      }
    }

    triCounter += numCorners - 2;
  }

//  std::cout << "force unshared num added: " << (numDrawVerts_ - numInitialVerts) << std::endl;
}

void MeshCompiler::getInputFaceVertex( const int _face, const int _corner, int* _out ) const
{
  for (unsigned int k = 0; k < decl_.getNumElements(); ++k)
    _out[k] = getInputIndex(_face, _corner, k);
}

void MeshCompiler::getInputFaceVertex_Welded( const int i, const int j, int* _out ) const
{
  int face = i;
  int corner = j;

  // apply welding map if available
  if (!vertexWeldMapFace_.empty())
  {
    const int offset = getInputFaceOffset(i);

    face = vertexWeldMapFace_[offset + j];
    corner = vertexWeldMapCorner_[offset + j];
  }

  for (unsigned int k = 0; k < decl_.getNumElements(); ++k)
    _out[k] = getInputIndex(face, corner, k);
}

void MeshCompiler::getInputFaceVertexData( const int _faceId, const int _corner, void* _out ) const
{
  for (int i = 0; i < numAttributes_; ++i)
  {
    const VertexElement* el = decl_.getElement(i);

    const int idx = getInputIndex(_faceId, _corner, i);

    input_[i].getElementData(idx, (char*)_out + (size_t)el->pointer_, el);
  }
}




MeshCompiler::MeshCompiler(const VertexDeclaration& _decl)
: decl_(_decl)
{
  faceInput_ = 0;
  deleteFaceInputeData_ = false;

  splitter_ = 0;
  numSubsets_ = 0;
  numIndices_ = 0;
  numTris_ = 0;

  numFaces_ = 0;
  curFaceInputPos_ = 0;

  numDrawVerts_ = 0;
  numIsolatedVerts_ = 0;

  maxFaceSize_ = 0;
  constantFaceSize_ = false;

  provokingVertex_ = -1;
  provokingVertexSetByUser_ = false;

  // search for convenient attribute indices
  numAttributes_ = decl_.getNumElements();
  inputIDNorm_ = inputIDPos_ = inputIDTexC_ = -1;

  for (int i = 0; i < (int)decl_.getNumElements(); ++i)
  {
    const VertexElement* e = decl_.getElement(i);

    switch (e->usage_)
    {
    case VERTEX_USAGE_POSITION: inputIDPos_ = i; break;
    case VERTEX_USAGE_NORMAL: inputIDNorm_ = i; break;
    case VERTEX_USAGE_TEXCOORD: inputIDTexC_ = i; break;
    default: break;
    }
  }


  vertexCompare_ = &defaultVertexCompare;

}

MeshCompiler::~MeshCompiler()
{
  if (deleteFaceInputeData_)
    delete faceInput_;

  delete splitter_;
}


int MeshCompiler::getInputIndexOffset( const int _face, const int _corner ) const
{
  assert(_face >= 0);
  assert(_face < numFaces_);

  // baseIdx: offset to first index of the (face, corner) pair
  const int baseIdx = int(faceStart_.empty() ? maxFaceSize_ * _face : faceStart_[_face]);
  return baseIdx + _corner;
}


void MeshCompiler::setInputIndexSplit( const int _face, const int _corner, const int _val )
{
  const int offset = getInputIndexOffset(_face, _corner);

  // keep track of number of vertices after splitting process
  if ( static_cast<size_t>(_val) >= numDrawVerts_)
    numDrawVerts_ = _val + 1;

  faceBufSplit_[offset] = _val;
}

int MeshCompiler::getInputIndexSplit( const int _face, const int _corner ) const
{
  const int offset = getInputIndexOffset(_face, _corner);

  return faceBufSplit_[offset];
}




MeshCompilerDefaultFaceInput::MeshCompilerDefaultFaceInput(int _numFaces, int _numIndices)
: numFaces_(_numFaces), numIndices_(_numIndices)
{
  faceOffset_.resize(numFaces_, -1);
  faceSize_.resize(numFaces_, 0);
  faceData_[0].reserve(_numIndices);
}


void MeshCompiler::setNumFaces( const int _numFaces, const int _numIndices )
{
  if (faceInput_)
    return;

  MeshCompilerDefaultFaceInput* internalInput = new MeshCompilerDefaultFaceInput(_numFaces, _numIndices);

  numFaces_ = _numFaces;

  faceInput_ = internalInput;
  deleteFaceInputeData_ = true;

}


void MeshCompiler::setFaceAttrib( int _i, int _numEdges, int* _v, int _attrID )
{
  if (!_v || _attrID < 0) return;

  if (!faceInput_)
    faceInput_ = new MeshCompilerDefaultFaceInput(0, 0);

  MeshCompilerDefaultFaceInput* input = dynamic_cast<MeshCompilerDefaultFaceInput*>(faceInput_);
  if (input)
  {
    input->setFaceData(_i, _numEdges, _v, _attrID);
  }
}

void MeshCompiler::setFaceAttrib( int _i, int _v0, int _v1, int _v2, int _attrID )
{
  int tmp[3] = {_v0, _v1, _v2};
  setFaceAttrib(_i, 3, tmp, _attrID);
}



MeshCompiler::VertexSplitter::VertexSplitter(int _numAttribs,
                                         int _numVerts,
                                         int _numWorstCase,
                                         float _estBufferIncrease)
: numAttribs(_numAttribs), numVerts(_numVerts), numBaseVerts(_numVerts)
{
  if (_numWorstCase <= 0)
    _numWorstCase = int(float(_numVerts) * (_estBufferIncrease + 1.0f));

  const int maxCount = (_numAttribs + 1) * (_numWorstCase + 1);

  // alloc split list and invalidate
  splits.resize(maxCount, -1);

  dbg_numResizes = 0;
  dbg_numSplits = 0;
}



MeshCompiler::VertexSplitter::~VertexSplitter()
{
}


int MeshCompiler::VertexSplitter::split(int* vertex)
{
  int pos = vertex[0];
  int next = getNext(pos);

  if (next < 0)
  {
    // 1st time reference

    // store attributes
    setAttribs(pos, vertex);

    // mark as referenced (next = this)
    setNext(pos, pos);
  }
  else
  {
    // already referenced

    int bSearchSplit = 1;

    // search vertex in split list
    while (pos >= 0 && bSearchSplit)
    {
      // is vertex already in split list?
      if (!memcmp(vertex, getAttribs(pos), numAttribs * sizeof(int)))
      {
        // found! reuse index
        return pos;
      }
      else
      {
        next = getNext(pos);
        
        if (next < 0) break;    // end of list
        if (next == pos) break; // avoid loop

        pos = next;             // go to next entry
      }
    }

    // combination not found -> add new vertex

    int newID = numVerts++;

    setNext(pos, newID);
    setAttribs(newID, vertex);

    pos = newID;

    ++dbg_numSplits;
  }

  return pos;
}



int MeshCompiler::VertexSplitter::getNext(const int id)
{
  assert(id >= 0);

  const int entryIdx = id * (1 + numAttribs);

  // need more space?
  if (entryIdx >= (int)splits.size())
  {
    splits.resize(entryIdx + numAttribs * 100, -1);
    ++dbg_numResizes;
  }

  return splits[entryIdx];
}

void MeshCompiler::VertexSplitter::setNext(const int id, const int next)
{
  assert(id >= 0);
  
  const int entryIdx = id * (1 + numAttribs);

  // need more space?
  if (entryIdx >= (int)splits.size())
  {
    splits.resize(entryIdx + numAttribs * 100, -1);
    ++dbg_numResizes;
  }

  splits[entryIdx] = next;
}

int* MeshCompiler::VertexSplitter::getAttribs(const int id)
{
  assert(id >= 0);

  const int entryIdx = id * (1 + numAttribs) + 1;

  // need more space?
  if (entryIdx + numAttribs >= (int)splits.size())
  {
    splits.resize(entryIdx + numAttribs * 100, -1);
    ++dbg_numResizes;
  }

  return &splits[0] + entryIdx;
}

void MeshCompiler::VertexSplitter::setAttribs( const int id, int* attr )
{
  memcpy(getAttribs(id), attr, numAttribs * sizeof(int));
}

bool MeshCompiler::VertexSplitter::isIsolated( const int vertexPosID )
{
  return (vertexPosID < numBaseVerts) && (getNext(vertexPosID) < 0);
}



MeshCompiler::VertexElementInput::VertexElementInput()
: internalBuf(0), data(0),
  count(0), stride(0), attrSize(0),
  fmt(0), elementSize(-1)
{
}

MeshCompiler::VertexElementInput::~VertexElementInput()
{
  delete [] internalBuf;
}








void MeshCompiler::triangulate()
{
  // count no. of triangles

  int numTris = 0;

  for (int i = 0; i < numFaces_; ++i)
    numTris += getFaceSize(i) - 2;

  numTris_ = numTris;

  triIndexBuffer_.resize(numTris * 3);
  triToSortFaceMap_.resize(numTris);

  // - find mapping (triToFaceMap):  triangle id -> sorted face id
  // - build triangle index buffer:  triIndexBuffer_

  // NOTE: triIndexBuffer_ contains local indices for each face, that is indices in the range [0, .. faceSize-1]
  //  these are encoded as a negative value starting at -1:  (-1 - localID)
  //  this change is necessary to implement the forceUnsharedVertices() function for complex polygons
  //  the negative values are resolved later in the function resolveTriangulation()

  int triCounter = 0;
  int indexCounter = 0;

  for (int sortFaceID = 0; sortFaceID < numFaces_; ++sortFaceID)
  {
    // get original face id
    const int faceID = faceSortMap_.empty() ? sortFaceID : faceSortMap_[sortFaceID];

    const int faceSize = getFaceSize(faceID);

    if (faceSize < 4)
    {
      // save face index mapping
      triToSortFaceMap_[triCounter++] = sortFaceID;

      for (int k = 0; k < 3; ++k)
        triIndexBuffer_[indexCounter++] = -1 - k; // getInputIndexSplit(faceID, k);
    }
    else
    {
      // use ACG::Triangulator to process complex polygons
      std::vector<Vec3f> poly(faceSize);
      for (int k = 0; k < faceSize; ++k)
      {
        VertexElement posElement;
        posElement.type_ = GL_FLOAT;
        posElement.numElements_ = 3;
        posElement.usage_ = VERTEX_USAGE_POSITION;
        posElement.pointer_ = 0;
        posElement.shaderInputName_ = 0;
        posElement.divisor_ = 0;
        posElement.vbo_ = 0;
        int posID = getInputIndexSplit(faceID, k);
        input_[inputIDPos_].getElementData(posID, &poly[k], &posElement);
      }
      Triangulator tris(poly);

      if (tris.convex())
      {
        // best case: convert polygon into triangle fan
        // NOTE: all triangles must use the first face-vertex here!
        triToSortFaceMap_[triCounter++] = sortFaceID;
        for (int k = 0; k < 3; ++k)
          triIndexBuffer_[indexCounter++] = -1 - k;

        for (int k = 3; k < faceSize; ++k)
        {
          // added tri belongs to current face
          triToSortFaceMap_[triCounter++] = sortFaceID;

          triIndexBuffer_[indexCounter++] = -1; // getInputIndexSplit(faceID, 0);
          triIndexBuffer_[indexCounter++] = -1 - (k - 1); //  getInputIndexSplit(faceID, k - 1);
          triIndexBuffer_[indexCounter++] = -1 - k; // getInputIndexSplit(faceID, k);
        }
      }
      else
      {
        // concave polygon
        // enforcing an unshared vertex gets ugly now

        for (size_t i = 0; i < tris.numTriangles(); ++i)
        {
          triToSortFaceMap_[triCounter++] = sortFaceID;
          for (int k = 0; k < 3; ++k)
          {
            int cornerID = tris.index(i * 3 + k);

            triIndexBuffer_[indexCounter++] = -1 - cornerID; // getInputIndexSplit(faceID, cornerID);
          }
        }
      }
    }
  }
 
  // ---------------
  // fill out missing subset info:

  for (int i = 0; i < numSubsets_; ++i)
  {
    subsets_[i].startIndex = 0xffffffff;
    subsets_[i].numTris = 0;
  }

  // triangle count per subset

  for (int i = 0; i < numTris; ++i)
  {
    const int faceID = mapTriToInputFace(i);

    const int faceGroup = getFaceGroup(faceID);
    Subset* sub = &subsets_[findGroupSubset(faceGroup)];

    ++sub->numTris;
  }

  // start index

  for (int i = 0; i < numSubsets_; ++i)
  {
    if (i > 0)
      subsets_[i].startIndex = subsets_[i-1].startIndex + subsets_[i-1].numTris * 3;
    else
      subsets_[i].startIndex = 0;
  }

}


void MeshCompiler::resolveTriangulation()
{
  // rotate tris such that the unshared face vertex is at the wanted provoking position of each triangle

  if (provokingVertex_ >= 0)
  {
    for (int i  = 0; i < numTris_; ++i)
    {
      for (int k = 0; k < 3 - provokingVertex_; ++k)
      {
        const int tmp =  triIndexBuffer_[i*3];
        triIndexBuffer_[i*3] = triIndexBuffer_[i*3 + 1];
        triIndexBuffer_[i*3 + 1] = triIndexBuffer_[i*3 + 2];
        triIndexBuffer_[i*3 + 2] = tmp;
      }
    }
  }

  // resolve triangulation to indices
  for (int drawTriID = 0; drawTriID < numTris_; ++drawTriID)
  {
    if (triIndexBuffer_[drawTriID * 3] < 0)
    {
      // triIndexBuffer stores the corner ids of the triangulations as:
      //  triIndexBuffer[idx] = -cornerID - 1

      // get original face id
      const int sortFaceID = triToSortFaceMap_[drawTriID];
      const int faceID = faceSortMap_.empty() ? sortFaceID : faceSortMap_[sortFaceID];

      for (int k = 0; k < 3; ++k)
      {
        int negCornerID = triIndexBuffer_[drawTriID * 3 + k];
        int cornerID = -1 - negCornerID;
        triIndexBuffer_[drawTriID * 3 + k] = getInputIndexSplit(faceID, cornerID);
      }
    }
  }
}

void MeshCompiler::sortFacesByGroup()
{
  // sort faces based on their group id
  // faces within the same group can be rendered in one batch

  numSubsets_ = 0;

  std::map<int, unsigned int> GroupIDs; // map[groupID] = first face

  // initialize GroupIDs map
  for (int face = 0; face < numFaces_; ++face)
  {
    const int texID = getFaceGroup(face);

    if (GroupIDs.find(texID) == GroupIDs.end())
      GroupIDs[texID] = face;
  }

  // alloc subset array
  numSubsets_ = int(GroupIDs.size());
  subsets_.resize(numSubsets_);

  if (numSubsets_ > 1)
    faceSortMap_.resize(numFaces_, -1);

  // initialize subsets and face sorting map

  std::map<int, unsigned int>::const_iterator it = GroupIDs.begin();
  unsigned int numSortedFaces = 0;

  for (unsigned int i = 0; it != GroupIDs.end(); ++i, ++it)
  {
    // subset id = group id
    subsets_[i].id = it->first;

    // store id mapping (optimization)
    subsetIDMap_[it->first] = i;

    // rearrange by subset chunks, face offset = # processed faces
    subsets_[i].numFaces = 0;
    subsets_[i].startFace = numSortedFaces;

    // triangle information is computed later
    subsets_[i].startIndex = 0;
    subsets_[i].numTris = 0;

    if (numSubsets_ == 1)
    {
      subsets_[i].numFaces = numFaces_;
    }
    else
    {
      // - find # faces in this subset
      // - create face sorting map:  map[sortFaceID] = faceID
      for (int k = it->second; k < numFaces_; ++k)
      {
        const int texID = getFaceGroup(k);

        if (texID == subsets_[i].id)
        {
          subsets_[i].numFaces++;

          faceSortMap_[numSortedFaces++] = k;
        }
      }
    }
  }

}


void MeshCompiler::optimize()
{
  indices_.resize(numTris_ * 3);
  triOptMap_.resize(numTris_, -1);

  for (int i = 0; i < numSubsets_; ++i)
  {
    Subset* pSubset = &subsets_[i];

    const int vcacheSize = 24;
    GPUCacheOptimizerTipsify copt(vcacheSize, pSubset->numTris, numDrawVerts_, 4, &triIndexBuffer_[0] + pSubset->startIndex);
    copt.WriteIndexBuffer(4, &indices_[pSubset->startIndex]);


    // apply changes to trimap
    const unsigned int StartTri = pSubset->startIndex/3;
    for (unsigned int k = 0; k < pSubset->numTris; ++k)
    {
      const unsigned int SrcTri = copt.GetTriangleMap()[k];
      triOptMap_[k + StartTri] = SrcTri + StartTri;
    }

  }

  // call to GPUCacheOptimizer::OptimizeVertices!

  unsigned int* vertexOptMap = new unsigned int[numDrawVerts_];

  GPUCacheOptimizer::OptimizeVertices(numTris_, numDrawVerts_, 4, indices_.data(), vertexOptMap);

  // apply vertexOptMap to index buffer

  for (int i = 0; i < numTris_ * 3; ++i)
    indices_[i] = vertexOptMap[indices_[i]];


  // apply opt-map to current vertex-map

  for (int i = 0; i < numFaces_; ++i)
  {
    const int fsize = getFaceSize(i);

    for (int k = 0;  k < fsize; ++k)
    {
      int oldVertex = getInputIndexSplit(i, k);

      int newVertex = vertexOptMap[oldVertex];

      setInputIndexSplit(i, k, newVertex);
    }
  }

  delete [] vertexOptMap;
}



void MeshCompiler::build(bool _weldVertices, bool _optimizeVCache, bool _needPerFaceAttribute, bool _keepIsolatedVertices)
{
  // track memory report for profiling/debugging
  const bool dbg_MemProfiling = false;

  // array allocation/copy data/validation checks etc.
  prepareData();

  /*
  1. step
  Split vertices s.t. we can create an interleaved vertex buffer.
  Neighboring faces may share a vertex with different normals,
  which gets resolved here.

  output:
   faceBufSplit_

  faceBufSplit_ reroutes the original index buffer to their unique
  vertex indices. So the orig. indices (faceBuf_) are still needed
  to know how to compose each vertex and faceBufSplit_ gives them
  a new index.
 
  faceBufSplit_ should not be used directly,
  use get/setInputIndexSplit for the mapping between interleaved indices and face vertices.
  */

  if (_weldVertices)
  {
    if (dbg_MemProfiling)
      std::cout << "computing adjacency.., memusage = " << (getMemoryUsage() /(1024 * 1024)) << std::endl;

    computeAdjacency();

    if (dbg_MemProfiling)
      std::cout << "vertex welding.., memusage = " << (getMemoryUsage() /(1024 * 1024)) << std::endl;

    weldVertices();

    // delete adjacency list (high memory cost)
    adjacencyVert_.clear();
  }
 
  if (dbg_MemProfiling)
    std::cout << "vertex splitting.., memusage = " << (getMemoryUsage() /(1024 * 1024)) << std::endl;

  splitVertices();

  if (dbg_MemProfiling)
    std::cout << "splitting done.., memusage = " << (getMemoryUsage() /(1024 * 1024)) << std::endl;

  // delete splitting and welding map ( high mem cost and is no longer needed after splitting )
  {
    delete splitter_; splitter_ = 0;

    {
      std::vector< int > emptyVec;
      vertexWeldMapFace_.swap(emptyVec);
    } {
      std::vector< int > emptyVec;
      vertexWeldMapCorner_.swap(emptyVec);
    }
  }

  /*
  2. step

  Sort faces by materials, grouping faces with the same material together.

  output:
   subsets_           (tri group offsets for each mat-id)
   triIndexBuffer_  (updated)
   triToSortFaceMap_      (updated)
  */
  if (dbg_MemProfiling)
    std::cout << "sorting by mat.., memusage = " << (getMemoryUsage() /(1024 * 1024)) << std::endl;
  sortFacesByGroup();




  /*
  3. step

  Triangulate any n-polys and unfold triangles to a new index buffer.

  output:
   numTris_
   triIndexBuffer_  (tri index buffer)
   triToSortFaceMap_      (map triangle id -> face id)

  From this point on we have to keep track of triToSortFaceMap_
  to compose vertices. 
  
  Assembling the vertex for corner "k" of triangle "tri" works as followed.
    int dstID = triIndexBuffer_[tri * 3 + k];
    int srcFace = triToSortFaceMap_[tri];

    int vertexAttribs[16];
    getInputFaceVertex(srcFace, k, vertexAttribs);

    vertexBuffer[dstID].pos = inputVerts_[vertexAttribs[POS]];
    vertexBuffer[dstID].normal = inputVerts_[vertexAttribs[NORMAL]];
    ...
  */
  if (dbg_MemProfiling)
    std::cout << "triangulate.., memusage = " << (getMemoryUsage() /(1024 * 1024)) << std::endl;
  triangulate();



  if (_needPerFaceAttribute)
  {
    if (dbg_MemProfiling)
      std::cout << "force unshared vertex.., memusage = " << (getMemoryUsage() / (1024 * 1024)) << std::endl;

    // The provoking vertex of each face shall not be referenced by any other face.
    // This vertex can then be used to store per-face data

    // default provoking position 2
    if (provokingVertex_ < 0)
      provokingVertex_ = 2;

    provokingVertex_ = provokingVertex_ % 3;

    // Adjacency info needed here
    forceUnsharedFaceVertex();
  }

  resolveTriangulation();

  /*
  4. step

  Optimize indices for efficient gpu vcache usage.
  This is done within the boundaries of each subset.
  Also, this operates on triangles instead of single indices,
  which means only complete triangles are reordered and
  will not be rotated afterwards.

  Additionally reorders vertex indices in increasing order.

  output:
   triIndexBuffer_  (updated)
   triOptMap
  */
  if (_optimizeVCache)
  {
    if (dbg_MemProfiling)
      std::cout << "optimizing.., memusage = " << (getMemoryUsage() /(1024 * 1024)) << std::endl;

    optimize();
  }
  else if (!triIndexBuffer_.empty())
    triIndexBuffer_.swap(indices_);

  if (dbg_MemProfiling)
    std::cout << "creating maps.., memusage = " << (getMemoryUsage() /(1024 * 1024)) << std::endl;

  // delete temporary tri index buffer (use indices_ instead)
  {
    std::vector< int > emptyVec;
    triIndexBuffer_.swap(emptyVec);
  }

  createVertexMap(_keepIsolatedVertices);
  createFaceMap();

  // delete intermediate mappings
  {
    std::vector< int > emptyVec;
    triToSortFaceMap_.swap(emptyVec);
  } {
    std::vector< int > emptyVec;
    triOptMap_.swap(emptyVec);
  } {
    std::vector< int > emptyVec;
    faceSortMap_.swap(emptyVec);
  }

  if (dbg_MemProfiling)
    std::cout << "finished.., memusage = " << (getMemoryUsage() /(1024 * 1024)) << std::endl;


  // debugging
//   dbgdumpInputObj("../../../dbg_meshcompinput.obj");
//   dbgdump("../../../dbg_meshcomp.txt");
//   dbgdumpObj("../../../dbg_meshcomp.obj");
//   dbgVerify("../../../dbg_maps.txt");
}






void MeshCompiler::setAttrib( int _attrIdx, int _v, const void* _data )
{
  VertexElementInput* inbuf = input_ + _attrIdx;

  if (!inbuf->internalBuf)
  {
    std::cerr << "setAttrib failed: please allocate internal input buffer before using setAttrib" << std::endl;
    return;
  }

  assert(inbuf->count > _v);

  memcpy(inbuf->internalBuf + (size_t)(_v * inbuf->stride), _data, inbuf->attrSize);
}

int MeshCompiler::getNumInputAttributes( int _attrIdx ) const
{
  assert (_attrIdx >= 0);
  assert(_attrIdx < int(decl_.getNumElements()));

  return input_[_attrIdx].count;
}


int MeshCompiler::mapTriToInputFace( const int _tri ) const
{
  assert(_tri >= 0);
  assert(_tri < numTris_);

  int sortFaceID = triToSortFaceMap_[_tri];
  assert(sortFaceID >= 0);
  assert(sortFaceID < numFaces_);

  int faceID = faceSortMap_.empty() ? sortFaceID : faceSortMap_[sortFaceID];
  assert(faceID >= 0);
  assert(faceID < numFaces_);

  return faceID;
}

int MeshCompiler::getFaceGroup( int _faceID ) const
{
  if (faceGroupIDs_.empty())
    return -1;

  return (int)faceGroupIDs_[_faceID];
}

int MeshCompiler::findGroupSubset( int _groupID )
{
  return subsetIDMap_[_groupID];
}

const MeshCompiler::Subset* MeshCompiler::getSubset( int _i ) const
{
  return &subsets_[_i];
}

std::string MeshCompiler::vertexToString( const void* v ) const
{
  std::stringstream str;

  for (int i = 0; i < (int)decl_.getNumElements(); ++i)
  {
    const VertexElement* el = decl_.getElement(i);

    str << el->shaderInputName_ << " [";

    const void* data = static_cast<const char*>(v) + (size_t)el->pointer_;

    switch ( el->type_ )
    {
    case GL_DOUBLE:
      {
        const double* d0 = static_cast<const double*>(data);

        for (int k = 0; k < (int)el->numElements_; ++k)
           str << d0[k] << ", ";
      } break;


    case GL_FLOAT:
      {
        const float* f0 = static_cast<const float*>(data);

        for (int k = 0; k < (int)el->numElements_; ++k)
          str << f0[k] << ", ";
      } break;

    case GL_INT:
    case GL_UNSIGNED_INT:
      {
        const int* i0 = static_cast<const int*>(data);

        for (int k = 0; k < (int)el->numElements_; ++k)
          str << i0[k] << ", ";
      } break;

    case GL_SHORT:
    case GL_UNSIGNED_SHORT:
      {
        const short* i0 = static_cast<const short*>(data);

        for (int k = 0; k < (int)el->numElements_; ++k)
          str << i0[k] << ", ";
      } break;

    case GL_BYTE:
    case GL_UNSIGNED_BYTE:
      {
        const char* i0 = static_cast<const char*>(data);

        for (int k = 0; k < (int)el->numElements_; ++k)
          str << ((int)i0[k]) << ", ";
      } break;

    default: std::cerr << "MeshCompiler: vertexToString() unknown type: " << el->type_ << std::endl;
    }


    str << "] ";

  }

  return str.str();
}

bool MeshCompiler::dbgVerify(const char* _filename) const
{
  int numTotalErrors = 0;

  std::ofstream file;

  if (_filename)
    file.open(_filename);

  if (file || !_filename)
  {

    // ---------------------------------------------------------------------------------------

    int numErrors = 0;

    // check draw_tri <-> face   (triangulation)
    if (file.is_open())
      file <<  "checking draw_tri <-> face mapping ..\n";

    for (int face = 0; face < getNumFaces(); ++face)
    {
      int numTrisOfFace = 0;
      mapToDrawTriID(face, 0, &numTrisOfFace);

      for (int k = 0; k < numTrisOfFace; ++k)
      {
        // triangle id of the k'th triangle of face
        int tri = mapToDrawTriID(face, k, 0);

        // the corresponding face of that triangle (inverse map)
        int dbg_face = mapToOriginalFaceID(tri);
        if (face != dbg_face)
        {
          if (file.is_open())
            file << "error: face " << face << " -> (numTris: " << numTrisOfFace << ",  tri " << tri << " -> face " << dbg_face << ")\n";
          ++numErrors;
        }
      }
    }

    if (file.is_open())
      file << numErrors << " errors found\n\n";
    numTotalErrors += numErrors;

    // ---------------------------------------------------------------------------------------

    // check input (face, corner) -> vertex id
    if (file.is_open())
      file <<  "checking (face, corner) -> vbo by comparing vertex data ..\n";
    numErrors = 0;

    char* vtxCmpData = new char[decl_.getVertexStride() * 2];

    for (int face = 0; face < getNumFaces(); ++face)
    {
      int fsize = getFaceSize(face);

      for (int k = 0; k < fsize; ++k)
      {
        char* v0 = vtxCmpData;
        char* v1 = vtxCmpData + decl_.getVertexStride();

        memset(v0, 0, decl_.getVertexStride());
        memset(v1, 0, decl_.getVertexStride());

        // get input vertex
        getInputFaceVertexData(face, k, v0);

        // get output vertex
        int vertex = mapToDrawVertexID(face, k);
        getVertex(vertex, v1);

        // allow slightly larger errors
        MeshCompilerVertexCompare verifyCmp(1e-3, 1e-3f);

        if (!verifyCmp.equalVertex(v0, v1, &decl_))
        {
          std::string vertexData0 = vertexToString(v0);
          std::string vertexData1 = vertexToString(v1);

          if (file)
            file << "error: (face " << face << ", corner " << k << ") -> vertex " << vertex << ": " << vertexData0.c_str() << " != "<< vertexData1.c_str() << "\n";

          ++numErrors;
        }
      }
    }

    if (file.is_open())
      file << numErrors << " errors found\n\n";
    numTotalErrors += numErrors;

    numErrors = 0;

    // ---------------------------------------------------------------------------------------

    // check vertex id -> input (face, corner)
    if (file.is_open())
      file <<  "checking vbo -> (face, corner) by comparing vertex data ..\n";

    for (int vertex = 0; vertex < getNumVertices(); ++vertex)
    {
      // map from vbo vertex id back to (face, corner)
      int face = 0, corner = 0;
      int posID = mapToOriginalVertexID(vertex, face, corner);

      char* v0 = vtxCmpData;
      char* v1 = vtxCmpData + decl_.getVertexStride();

      memset(v0, 0, decl_.getVertexStride());
      memset(v1, 0, decl_.getVertexStride());

      // compare vbo data at [vertex] with the input data at (face, corner)

      // get output vertex
      getVertex(vertex, v0);

      // get input vertex
      if (face >= 0)
        getInputFaceVertexData(face, corner, v1);
      else
      {
        // isolated vertex
        for (unsigned int i = 0; i < decl_.getNumElements(); ++i)
        {
          const VertexElement* el = decl_.getElement(i);
          input_[i].getElementData(posID, (char*)v1 + (size_t)el->pointer_, el);
        }
      }

      if (!vertexCompare_->equalVertex(v0, v1, &decl_))
      {
        std::string vertexData0 = vertexToString(v0);
        std::string vertexData1 = vertexToString(v1);

        if (file)
          file << "error: vertex " << vertex << " -> (face " << face << ", corner " << corner << "): " << vertexData0.c_str() << " != " << vertexData1.c_str() << "\n";
        ++numErrors;
      }
    }

    if (file.is_open())
      file << numErrors << " errors found\n\n";
    numTotalErrors += numErrors;

    numErrors = 0;

    delete [] vtxCmpData;


    // -----------------------------------------------------------

    // check unshared vertex
    if (provokingVertex_ >= 0)
    {
      if (file.is_open())
        file <<  "checking unshared per face vertices ..\n";

      // count number of references for each vertex
      std::vector< std::map<int, int> > VertexRefs;
      VertexRefs.resize(getNumVertices());

      for (int face = 0; face < numFaces_; ++face)
      {
        int nTris = 0;
        mapToDrawTriID(face, 0, &nTris);

        for (int k = 0; k < nTris; ++k)
        {
          int tri = mapToDrawTriID(face, k, 0);

          int faceVertex = getIndex(tri * 3 + provokingVertex_);

          VertexRefs[faceVertex][face] = 1;
        }

      }

      // the first vertex of each face should not be referenced by more than one face

      for (int i = 0; i < getNumVertices(); ++i)
      {
        if (VertexRefs[i].size() > 1)
        {
          if (file)
          {
            file << "error : vertex " << i << " is referenced by " << VertexRefs[i].size() << " faces: ";

            for (std::map<int, int>::iterator it = VertexRefs[i].begin(); it != VertexRefs[i].end(); ++it)
              file << it->first << ", ";

            file << "\n";
          }
          ++numErrors;
        }
      }

      if (file.is_open())
        file << numErrors << " errors found\n\n";
      numTotalErrors += numErrors;

      numErrors = 0;


      // -----------------------------------------------------------
      // check face group sorting

      if (file.is_open())
        file << "checking face group sorting ..\n";

      for (int i = 0; i < getNumSubsets(); ++i)
      {
        const ACG::MeshCompiler::Subset* sub = getSubset(i);

        for (int k = 0; k < (int)sub->numTris; ++k)
        {
          int faceID = mapToOriginalFaceID(sub->startIndex/3 + k);
          int grID = getFaceGroup(faceID);

          // input face group should match the subsets

          if (grID != sub->id)
          {
            if (file.is_open())
              file <<  "error: face " << faceID << " with group-id " << grID << " was mapped to subset-group " << sub->id << "\n";

            ++numErrors;
          }

        }
      }

      if (file.is_open())
        file << numErrors << " errors found\n\n";
      numTotalErrors += numErrors;

      numErrors = 0;
    }

    // -----------------------------------------------------------
    // check triangulation

    if (file.is_open())
      file << "checking triangulation ..\n";

    for (int i = 0; i < numFaces_; ++i)
    {
      int faceSize = getFaceSize(i);

      std::vector<int> facePositions(faceSize, -1);

      for (int k = 0; k < faceSize; ++k)
        facePositions[k] = getInputIndex(i, k, inputIDPos_);

      int numFaceTris = 0;
      mapToDrawTriID(i, 0, &numFaceTris);

      for (int t = 0; t < numFaceTris; ++t)
      {
        int triID = mapToDrawTriID(i, t);

        int triPosOccurrence[3] = {-1, -1, -1};

        for (int k = 0; k < 3; ++k)
        {
          int vertexID = getIndex(triID * 3 + k);

          int originalFace = -1, originalCorner = -1;
          int posID = mapToOriginalVertexID(vertexID, originalFace, originalCorner);

          // check if the triangle positions make a subset of the polygon positions
          for (int m = 0; m < faceSize; ++m)
          {
            if (posID == facePositions[m])
            {
              triPosOccurrence[k] = m;
              break;
            }
          }

          if (triPosOccurrence[k] < 0)
          {
            if (file.is_open())
              file <<  "error: vertex at triangulated face " << i << " at triangle " << t << " at corner " << k << "is not even part of the original face!\n";

            ++numErrors;
          }
        }

        // check face winding of triangulation
        int numInversions = 0;

        for (int k = 0; k < 3; ++k)
        {
          int p1 = triPosOccurrence[k];
          int p2 = triPosOccurrence[(k + 1) % 3];

          if (p1 > p2)
            ++numInversions;
        }

        if (numInversions > 1)
        {
          if (file.is_open())
            file <<  "error: triangulation of face " << i << " at triangle " << t << " has flipped winding order!\n";

          ++numErrors;
        }
      }
    }

    if (file.is_open())
      file << numErrors << " errors found\n\n";
    numTotalErrors += numErrors;

    if (file.is_open())
      file.close();
  }

  return numTotalErrors == 0;
}

void MeshCompiler::dbgdump(const char* _filename) const
{
  // dump all the internal arrays to text

  std::ofstream file;
  file.open(_filename);

  if (file.is_open())
  {
    for (int i = 0; i < numAttributes_; ++i)
    {
      const VertexElementInput* inp = input_ + i;

      file << "attribute[" << i << "]: internalbuf " << ((const void*)inp->internalBuf) << ", data " << ((const void*)inp->data) << ", count " << inp->count << ", stride " << inp->stride << ", attrSize " << inp->attrSize << "\n";
    }

    file << "\n\n";
    file << "faces " << numFaces_ << "\nindices " << numIndices_ << "\n";

    if (!vertexWeldMapFace_.empty())
    {
      for (int i = 0; i < numFaces_; ++i)
      {
        for (int k = 0; k < getFaceSize(i); ++k)
        {
          int face = vertexWeldMapFace_[getInputFaceOffset(i) + k];
          int corner = vertexWeldMapCorner_[getInputFaceOffset(i) + k];
          file << "vertexWeldMap_[" << i <<", "<<k<<"] = ["<<face<<", "<<corner<<"]\n";
        }
      }
    }

    for (size_t i = 0; i < faceBufSplit_.size(); ++i)
      file << "faceBufSplit_["<<i<<"] = "<<faceBufSplit_[i]<<"\n";

    file << "\n\n";

    for (size_t i = 0; i < faceGroupIDs_.size(); ++i)
      file << "faceGroupIDs_["<<i<<"] = "<<faceGroupIDs_[i]<<"\n";

    file << "\n\n";

    for (size_t i = 0; i < faceSortMap_.size(); ++i)
      file << "faceSortMap_["<<i<<"] = "<<faceSortMap_[i]<<"\n";

    file << "\n\n";

    for (size_t i = 0; i < triIndexBuffer_.size()/3; ++i)
      file << "tri["<<i<<"] = "<<triIndexBuffer_[i*3+1]<<" "<<triIndexBuffer_[i*3+1]<<" "<<triIndexBuffer_[i*3+2]<<"\n";

    file << "\n\n";

    for (size_t i = 0; i < subsets_.size(); ++i)
    {
      const Subset* sub = &subsets_[i];
      file <<"subset["<<i<<"]: id "<<sub->id<<", startIndex "<<sub->startIndex<<", numTris "<<sub->numTris<<", numFaces "<<sub->numFaces<<", startFace "<<sub->startFace<<"\n";
    }

    file << "\n\n";

    for (std::map<int, int>::const_iterator it = subsetIDMap_.begin();
      it != subsetIDMap_.end(); ++it)
      file << "subsetIDMap["<<it->first<<"] = "<<it->second<<"\n";

    file << "\n\n";

    for (int i = 0; i < numFaces_; ++i)
    {
      for (int k = 0; k < getFaceSize(i); ++k)
        file <<"mapToDrawVertexID["<<i<<", "<<k<<"] = "<<mapToDrawVertexID(i, k)<<"\n";
    }

    file << "\n\n";

    for (int i = 0; i < getNumVertices(); ++i)
    {
      int f, c;
      mapToOriginalVertexID(i, f, c);
      file <<"mapToOriginalVertexID["<<i<<"] = ["<<f<<", "<<c<<"]\n";
    }

    file << "\n\nadjacencyVert\n";
    adjacencyVert_.dbgdump(file);

    file << "\n\n";

    file << "\n\n";


    for (int i = 0; i < numFaces_; ++i)
    {
        for (int k = 0; k < getFaceSize(i); ++k)
        {
          float vtx[32];
          getInputFaceVertexData(i, k, vtx);

          //                                    pos[0]       pos[1]       pos[2]          uv[0]        uv[1]          n[0]         n[1]         n[2]
          file << "data["<<i<<", "<<k<<"] = ("<<vtx[0]<<" "<<vtx[1]<<" "<<vtx[2]<<") ("<<vtx[3]<<" "<<vtx[4]<<") ("<<vtx[5]<<" "<<vtx[6]<<" "<<vtx[7]<<")\n";
        }
    }

    file.close();
  }
}


void MeshCompiler::AdjacencyList::dbgdump(std::ofstream& file) const
{
  if (file.is_open())
  {
    for (int i = 0; i < num; ++i)
    {
      for (int k = 0; k < getCount(i); ++k)
        file << "adj["<<i<<"]["<<k<<"] = "<<getAdj(i,k)<<"\n";
    }
  }
}

void MeshCompiler::VertexElementInput::getElementData(int _idx, void* _dst, const VertexElement* _desc) const
{
  // attribute data not set by user, skip
  if (count == 0 || !data)
  {
    memset(_dst, 0, attrSize);
    return;
  }

  //   assert(_idx >= 0);
  //   assert(_idx < count);

  if (_idx < 0 || _idx >= count)
  {
    memset(_dst, 0, attrSize);
    return;
  }

  const void* dataAdr = static_cast<const char*>(data) + (size_t)(_idx * stride);

  if (fmt == 0 || elementSize == -1 || fmt == _desc->type_)
    memcpy(_dst, dataAdr, attrSize);
  else
  {
    // typecast data to format in vertex buffer
    int data_i[4] = {0,0,0,0};    // data in integer fmt
    double data_d[4] = {0.0, 0.0, 0.0, 0.0}; // data in floating point

    // read data
    for (int i = 0; i < elementSize; ++i)
    {
      switch (fmt)
      {
      case GL_FLOAT:  data_d[i] = (static_cast<const float*>(dataAdr))[i];  break;
      case GL_DOUBLE: data_d[i] = (static_cast<const double*>(dataAdr))[i];  break;
      
      case GL_UNSIGNED_INT:
      case GL_INT:    data_i[i] = (static_cast<const int*>(dataAdr))[i];  break;
      
      case GL_UNSIGNED_SHORT:
      case GL_SHORT:  data_i[i] = (static_cast<const short*>(dataAdr))[i];  break;

      case GL_UNSIGNED_BYTE:
      case GL_BYTE:  data_i[i] = (static_cast<const short*>(dataAdr))[i];  break;

      default: std::cerr << "MeshCompiler: unknown data format - " << fmt << std::endl; break;
      }
    }


    // zero mem
    memset(_dst, 0, attrSize);

    // typecast data
    if (fmt == GL_FLOAT || fmt == GL_DOUBLE)
    {
      for (int i = 0; i < (int)_desc->numElements_ && i < (int)elementSize; ++i)
      {
        switch (_desc->type_)
        {
        case GL_FLOAT: ((float*)_dst)[i] = (float)data_d[i]; break;
        case GL_DOUBLE: ((double*)_dst)[i] = (double)data_d[i]; break;
        
        case GL_UNSIGNED_INT:
        case GL_INT: ((int*)_dst)[i] = (int)data_d[i]; break;

        case GL_UNSIGNED_BYTE:
        case GL_BYTE: ((char*)_dst)[i] = (char)data_d[i]; break;

        case GL_UNSIGNED_SHORT:
        case GL_SHORT: ((short*)_dst)[i] = (short)data_d[i]; break;

        default: break;
        }
      }
    }
    else
    {
      for (int i = 0; i < (int)_desc->numElements_ && i < (int)elementSize; ++i)
      {
        switch (_desc->type_)
        {
        case GL_FLOAT: ((float*)_dst)[i] = (float)data_i[i]; break;
        case GL_DOUBLE: ((double*)_dst)[i] = (double)data_i[i]; break;

        case GL_UNSIGNED_INT:
        case GL_INT: ((int*)_dst)[i] = (int)data_i[i]; break;

        case GL_UNSIGNED_BYTE:
        case GL_BYTE: ((char*)_dst)[i] = (char)data_i[i]; break;

        case GL_UNSIGNED_SHORT:
        case GL_SHORT: ((short*)_dst)[i] = (short)data_i[i]; break;

        default: break;
        }
      }
    }


  }
}


void MeshCompiler::dbgdumpObj(const char* _filename) const
{
  // dump resulting mesh to obj file

  std::ofstream file;
  file.open(_filename);

  if (file.is_open())
  {
    for (size_t i = 0; i < numDrawVerts_; ++i)
    {
      float vertex[64];

      getVertex(i, vertex);

      if (inputIDPos_ >= 0)
      {
        float* pos = vertex + (size_t)decl_.getElement(inputIDPos_)->pointer_ / 4;
        file << "v "<<pos[0]<<" "<<pos[1]<<" "<<pos[2]<<"\n";
      }

      if (inputIDNorm_ >= 0)
      {
        float* pos = vertex + (size_t)decl_.getElement(inputIDNorm_)->pointer_ / 4;
        file << "vn "<<pos[0]<<" "<<pos[1]<<" "<<pos[2]<<"\n";
      }

      if (inputIDTexC_ >= 0)
      {
        float* pos = vertex  + (size_t)decl_.getElement(inputIDTexC_)->pointer_ / 4;
        file << "vt "<<pos[0]<<" "<<pos[1]<<"\n";
      }
    }


    for (int i = 0; i < numTris_; ++i)
    {
      if (!inputIDTexC_ && !inputIDNorm_)
        file << "f "<<indices_[i*3]+1<<" "<<indices_[i*3+1]+1<<" "<<indices_[i*3+2]+1<<"\n";

      else if (!inputIDTexC_)
        file << "f "<<indices_[i*3]+1<<"/"<<indices_[i*3]+1<<" "<<indices_[i*3+1]+1<<"/"<<indices_[i*3+1]+1<<" "<<indices_[i*3+2]+1<<"/"<<indices_[i*3+2]+1<<"\n";

      else
        file << "f "<<indices_[i*3]+1<<"/"<<indices_[i*3]+1<<"/"<<indices_[i*3]+1<<" "<<indices_[i*3+1]+1<<"/"<<indices_[i*3+1]+1<<"/"<<indices_[i*3+1]+1<<" "<<indices_[i*3+2]+1<<"/"<<indices_[i*3+2]+1<<"/"<<indices_[i*3+2]+1<<"\n";
    }


    file.close();
  }
}


void MeshCompiler::dbgdumpInputObj( const char* _filename ) const
{
  // dump input mesh to obj file

  const int nVerts = (inputIDPos_ >= 0) ? input_[inputIDPos_].count : 0;
//  const int nNormals = (inputIDNorm_ >= 0) ? input_[inputIDNorm_].count : 0;
//  const int nTexC = (inputIDTexC_ >= 0) ? input_[inputIDTexC_].count : 0;

  
  std::ofstream file;
  file.open(_filename);

  if (file.is_open())
  {
    // write vertex data

    for (int i = 0; i < nVerts; ++i)
    {
      float pos[16];
      const VertexElement* el = decl_.getElement(inputIDPos_);
      input_[inputIDPos_].getElementData(i, pos, el);

      file << "v "<<pos[0]<<" "<<pos[1]<<" "<<pos[2]<<"\n";
    }

    if (inputIDTexC_ >= 0)
    {
      float pos[16];
      const VertexElementInput* vei = &input_[inputIDTexC_];
      const VertexElement* el = decl_.getElement(inputIDTexC_);

      for (int i = 0; i < vei->count; ++i)
      {
        vei->getElementData(i, pos, el);
        file << "vt "<<pos[0]<<" "<<pos[1]<<"\n";
      }
    }


    if (inputIDNorm_ >= 0)
    {
      float pos[16];
      const VertexElementInput* vei = &input_[inputIDNorm_];
      const VertexElement* el = decl_.getElement(inputIDNorm_);

      for (int i = 0; i < vei->count; ++i)
      {
        vei->getElementData(i, pos, el);
        file << "vn "<<pos[0]<<" "<<pos[1]<<" "<<pos[2]<<"\n";
      }
    }


    // write face data

    for (int i = 0; i < faceInput_->getNumFaces(); ++i)
    {
      std::string strLine = "f ";

      int size = faceInput_->getFaceSize(i);

      char tmp[0xff];
      for (int k = 0; k < size; ++k)
      {
        if (inputIDNorm_>=0 && inputIDTexC_>=0)
          sprintf(tmp, "%d/%d/%d ", faceInput_->getSingleFaceAttr(i, k, inputIDPos_) + 1,
          faceInput_->getSingleFaceAttr(i, k, inputIDTexC_) + 1,
          faceInput_->getSingleFaceAttr(i, k, inputIDNorm_) + 1);
        else if (inputIDNorm_ >= 0)
          sprintf(tmp, "%d//%d ", faceInput_->getSingleFaceAttr(i, k, inputIDPos_) + 1, faceInput_->getSingleFaceAttr(i, k, inputIDNorm_) + 1);
        else if (inputIDTexC_ >= 0)
          sprintf(tmp, "%d/%d ", faceInput_->getSingleFaceAttr(i, k, inputIDPos_) + 1, faceInput_->getSingleFaceAttr(i, k, inputIDTexC_) + 1);
        else
          sprintf(tmp, "%d ", faceInput_->getSingleFaceAttr(i, k, inputIDPos_) + 1);

        strLine += tmp;
      }

      file << strLine.c_str() << "\n";
    }

    file.close();
  }

}


void MeshCompiler::dbgdumpInputBin( const char* _filename, bool _seperateFiles ) const
{
  // dump input mesh to custom binary format (for debugging/profiling purposes)

  const int nVerts = (inputIDPos_ >= 0) ? input_[inputIDPos_].count : 0;
  const int nNormals = (inputIDNorm_ >= 0) ? input_[inputIDNorm_].count : 0;
  const int nTexC = (inputIDTexC_ >= 0) ? input_[inputIDTexC_].count : 0;

  if (!_seperateFiles)
  {
    /*
    simple binary format (can be used for profiling and testing with large meshes):

    int numFaces,
      numVerts,
      numNormals,
      numTexCoords,
      faceBufSize

    -------------------------

    int faceSize[numFaces];
    int faceData[attribute][faceBufSize];

    -------------------------

    float3 verts[numVerts];
    float2 texc[numTexCoords];
    float3 normals[numNormals];

    -------------------------

    adjacency chunks: vertex + face
  
    int num, bufsize;
    int* offset[num];
    uchar* count[num];
    int* buf[bufsize];

    */


    FILE* file = 0;

    file = fopen(_filename, "wb");

    if (file)
    {
      const int nFaces = faceInput_->getNumFaces();
      fwrite(&nFaces, 4, 1, file);

      fwrite(&nVerts, 4, 1, file);
      fwrite(&nNormals, 4, 1, file);
      fwrite(&nTexC, 4, 1, file);

      int faceBufSize = 0;

      for (int i = 0; i < faceInput_->getNumFaces(); ++i)
        faceBufSize += faceInput_->getFaceSize(i);

      fwrite(&faceBufSize, 4, 1, file);

      // write face data

      for (int i = 0; i < faceInput_->getNumFaces(); ++i)
      {
        int fsize = faceInput_->getFaceSize(i);
        fwrite(&fsize, 4, 1, file);
      }

      for (int i = 0; i < faceInput_->getNumFaces(); ++i)
      {
        for (int k = 0; k < faceInput_->getFaceSize(i); ++k)
        {
          int idx = faceInput_->getSingleFaceAttr(i, k, inputIDPos_);
          fwrite(&idx, 4, 1, file);
        }
      }

      if (inputIDTexC_ >= 0)
      {
        for (int i = 0; i < faceInput_->getNumFaces(); ++i)
        {
          for (int k = 0; k < faceInput_->getFaceSize(i); ++k)
          {
            int idx = faceInput_->getSingleFaceAttr(i, k, inputIDTexC_);
            fwrite(&idx, 4, 1, file);
          }
        }
      }

      if (inputIDNorm_ >= 0)
      {
        for (int i = 0; i < faceInput_->getNumFaces(); ++i)
        {
          for (int k = 0; k < faceInput_->getFaceSize(i); ++k)
          {
            int idx = faceInput_->getSingleFaceAttr(i, k, inputIDNorm_);
            fwrite(&idx, 4, 1, file);
          }
        }
      }

      // write vertex data

    
      for (int i = 0; i < nVerts; ++i)
      {
        float pos[16];
        const VertexElement* el = decl_.getElement(inputIDPos_);
        input_[inputIDPos_].getElementData(i, pos, el);

        fwrite(pos, 4, 3, file);
      }

      if (inputIDTexC_ >= 0)
      {
        float pos[16];
        const VertexElementInput* vei = &input_[inputIDTexC_];
        const VertexElement* el = decl_.getElement(inputIDTexC_);

        for (int i = 0; i < vei->count; ++i)
        {
          vei->getElementData(i, pos, el);
          fwrite(pos, 4, 2, file);
        }
      }


      if (inputIDNorm_ >= 0)
      {
        float pos[16];
        const VertexElementInput* vei = &input_[inputIDNorm_];
        const VertexElement* el = decl_.getElement(inputIDNorm_);

        for (int i = 0; i < vei->count; ++i)
        {
          vei->getElementData(i, pos, el);
          fwrite(pos, 4, 3, file);
        }
      }
    

      // write adjacency

      fwrite(&adjacencyVert_.num, 4, 1, file);
      fwrite(&adjacencyVert_.bufSize, 4, 1, file);

      fwrite(adjacencyVert_.start, 4, adjacencyVert_.num, file);
      fwrite(adjacencyVert_.count, 1, adjacencyVert_.num, file);
      fwrite(adjacencyVert_.buf, 4, adjacencyVert_.bufSize, file);

      fclose(file);
    }

  }
  else
  {
    // dump data to separate files
    FILE* file = fopen("../mesh_fsize.bin", "wb");

    for (int i = 0; i < numFaces_; ++i)
    {
      unsigned char fsize = (unsigned char)faceInput_->getFaceSize(i);
      fwrite(&fsize, 1, 1, file);
    }

    fclose(file);

    file = fopen("../mesh_fdata_pos.bin", "wb");

    for (int i = 0; i < numFaces_; ++i)
    {
      for (int k = 0; k < faceInput_->getFaceSize(i); ++k)
      {
        int idx = faceInput_->getSingleFaceAttr(i, k, inputIDPos_);
        fwrite(&idx, 4, 1, file);
      }
    }

    fclose(file);

    if (inputIDTexC_ >= 0)
    {
      file = fopen("../mesh_fdata_texc.bin", "wb");

      for (int i = 0; i < numFaces_; ++i)
      {
        for (int k = 0; k < faceInput_->getFaceSize(i); ++k)
        {
          int idx = faceInput_->getSingleFaceAttr(i, k, inputIDTexC_);
          fwrite(&idx, 4, 1, file);
        }
      }

      fclose(file);
    }

    if (inputIDNorm_ >= 0)
    {
      file = fopen("../mesh_fdata_norm.bin", "wb");

      for (int i = 0; i < numFaces_; ++i)
      {
        for (int k = 0; k < faceInput_->getFaceSize(i); ++k)
        {
          int idx = faceInput_->getSingleFaceAttr(i, k, inputIDNorm_);
          fwrite(&idx, 4, 1, file);
        }
      }

      fclose(file);
    }

    // write vertex data

    file = fopen("../mesh_vdata_pos.bin", "wb");
    for (int i = 0; i < nVerts; ++i)
    {
      float pos[16];
      const VertexElement* el = decl_.getElement(inputIDPos_);
      input_[inputIDPos_].getElementData(i, pos, el);

      fwrite(pos, 4, 3, file);
    }
    fclose(file);

    if (inputIDTexC_ >= 0)
    {
      file = fopen("../mesh_vdata_texc.bin", "wb");

      float pos[16];
      const VertexElementInput* vei = &input_[inputIDTexC_];
      const VertexElement* el = decl_.getElement(inputIDTexC_);

      for (int i = 0; i < vei->count; ++i)
      {
        vei->getElementData(i, pos, el);
        fwrite(pos, 4, 2, file);
      }

      fclose(file);
    }


    if (inputIDNorm_ >= 0)
    {
      file = fopen("../mesh_vdata_norm.bin", "wb");

      float pos[16];
      const VertexElementInput* vei = &input_[inputIDNorm_];
      const VertexElement* el = decl_.getElement(inputIDNorm_);

      for (int i = 0; i < vei->count; ++i)
      {
        vei->getElementData(i, pos, el);
        fwrite(pos, 4, 3, file);
      }

      fclose(file);
    }
  }

}


void MeshCompiler::setFaceVerts( int _i, int _numEdges, int* _v )
{
  setFaceAttrib(_i, _numEdges, _v, inputIDPos_);
}

void MeshCompiler::setFaceVerts( int _i, int _v0, int _v1, int _v2 )
{
  int tmp[] = {_v0, _v1, _v2};
  setFaceAttrib(_i, 3, tmp, inputIDPos_);
}

void MeshCompiler::setFaceNormals( int _i, int _numEdges, int* _v )
{
  setFaceAttrib(_i, _numEdges, _v, inputIDNorm_);
}

void MeshCompiler::setFaceNormals( int _i, int _v0, int _v1, int _v2 )
{
  int tmp[] = {_v0, _v1, _v2};
  setFaceAttrib(_i, 3, tmp, inputIDNorm_);
}


void MeshCompiler::setFaceTexCoords( int _i, int _numEdges, int* _v )
{
  setFaceAttrib(_i, _numEdges, _v, inputIDTexC_);
}

void MeshCompiler::setFaceTexCoords( int _i, int _v0, int _v1, int _v2 )
{
  int tmp[] = {_v0, _v1, _v2};
  setFaceAttrib(_i, 3, tmp, inputIDTexC_);
}

void MeshCompiler::getVertex( int _id, void* _out ) const
{
  int faceID, cornerID;
  int posID = mapToOriginalVertexID(_id, faceID, cornerID);

  if (faceID >= 0)
  {
    // read connnected vertex
    for (int i = 0; i < numAttributes_; ++i)
    {
      const VertexElement* el = decl_.getElement(i);

      int idx = getInputIndex(faceID, cornerID, i);

      input_[i].getElementData(idx, (char*)_out + (size_t)el->pointer_, el);
    }
  }
  else
  {
    // isolated vertex

    for (int i = 0; i < numAttributes_; ++i)
    {
      const VertexElement* el = decl_.getElement(i);
      input_[i].getElementData(posID, (char*)_out + (size_t)el->pointer_, el);
    }
  }
}

int MeshCompiler::getIndex( int _i ) const
{
  assert(_i >= 0);
  assert(_i < numTris_ * 3);
  return indices_[_i];
}

void MeshCompiler::createVertexMap(bool _keepIsolatedVerts)
{
  // vertexMap: vertex id -> (face, corner)

  // the current vertex buffer does not contain any isolated vertices
  // -> add them to end of buffer if required
  const int offsetIso = numDrawVerts_;
  if (_keepIsolatedVerts)
    numDrawVerts_ += numIsolatedVerts_; // keep
  else
    numIsolatedVerts_ = 0; // delete isolated vertices

  vertexMapFace_.resize(numDrawVerts_, -1);
  vertexMapCorner_.resize(numDrawVerts_, -1);

  for (int i = 0; i < numFaces_; ++i)
  {
    const int fsize = getFaceSize(i);
    for (int k = 0; k < fsize; ++k)
    {
      // map from (face, corner) -> vertex id is given by getInputIndexSplit(), so create the inverse
      int vertexID = getInputIndexSplit(i, k);
      vertexMapFace_[vertexID] = i;
      vertexMapCorner_[vertexID] = k;
    }
  }

  if (_keepIsolatedVerts)
  {
    // add isolated verts to end of map
    for ( size_t i = 0; i < numIsolatedVerts_; ++i)
    {
      assert(vertexMapFace_[offsetIso + i] < 0);
      vertexMapFace_[offsetIso + i] = isolatedVertices_[i];
    }
  }

  // validate
#ifdef _DEBUG
  for (int i = 0; i < numDrawVerts_; ++i)
  {
    if (i < offsetIso)
    {
      if (vertexMapFace_[i] < 0 ||
        vertexMapCorner_[i]  < 0)
        std::cerr << "mesh-assembler: vertexMap error at index: " << i << std::endl;
    }
    else if (vertexMapFace_[i] < 0)
      std::cerr << "mesh-assembler: vertexMap error at index: " << i << std::endl;
  }
#endif // _DEBUG

}


void MeshCompiler::createFaceMap()
{
  // -------------------------------
  // create tri -> face map

  triToFaceMap_.clear();
  triToFaceMap_.resize(numTris_, -1);
  for (int i = 0; i < numTris_; ++i)
  {
    int faceID = i;
    if (!triOptMap_.empty())
      faceID = triOptMap_[faceID];

    if (!triToSortFaceMap_.empty())
      faceID = triToSortFaceMap_[faceID];

    if (!faceSortMap_.empty())
      faceID = faceSortMap_[faceID];

    triToFaceMap_[i] = faceID;
  }


  // -------------------------------
  // create face -> tri map

  // offset table is necessary for variable polygon face sizes, because the map is stored in a single array

  // create offset table
  faceToTriMapOffset_.clear();
  if (!constantFaceSize_)
  {
    faceToTriMapOffset_.resize(numTris_, -1);

    int offset = 0;
    for (int i = 0; i < numFaces_; ++i)
    {
      faceToTriMapOffset_[i] = offset;

      // # tris of face
      int fsize = getFaceSize(i);
      offset += fsize - 2;
    }
  }

  // create face -> tri map
  faceToTriMap_.clear();
  faceToTriMap_.resize(numTris_, -1); 

  for (int i = 0; i < numTris_; ++i)
  {
    // face -> tri map is the inverse of tri -> face map, which is already given by mapToOriginalFaceID()
    int faceID = mapToOriginalFaceID(i);

    // offset into lookup table
    int offset = constantFaceSize_ ? (faceID * (maxFaceSize_ - 2)) : faceToTriMapOffset_[faceID];

    int triNo = 0;

    // search for free entry
    while (faceToTriMap_[offset + triNo] != -1 && triNo + offset < numTris_)
      ++triNo;

    assert(triNo < getFaceSize(faceID) - 2);

    faceToTriMap_[offset + triNo] = i;
  }
}



void MeshCompiler::dbgdumpAdjList( const char* _filename ) const
{
  // dump adjacency list to text file: vertex -> adjacent faces

  FILE* file = 0;
  
  file = fopen(_filename, "wt");

  if (file)
  {
    fprintf(file, "vertex-adjacency: \n");
    for (int i = 0; i < input_[inputIDPos_].count; ++i)
    {
      // sorting the adjacency list for easy comparison of adjacency input
      int count = getAdjVertexFaceCount(i);

      std::vector<int> sortedList(count);
      for (int k = 0; k < count; ++k)
        sortedList[k] = getAdjVertexFace(i, k);

      std::sort(sortedList.begin(), sortedList.end());

      for (int k = 0; k < count; ++k)
        fprintf(file, "adj[%d][%d] = %d\n", i, k, sortedList[k]);
    }


    fclose(file);
  }
}

void MeshCompiler::setFaceGroup( int _i, short _groupID )
{
  if ((int)faceGroupIDs_.size() <= std::max(numFaces_,_i))
    faceGroupIDs_.resize(std::max(numFaces_,_i+1), -1);

  faceGroupIDs_[_i] = _groupID;
}

void MeshCompiler::getVertexBuffer( void* _dst, const int _offset /*= 0*/, const int _range /*= -1*/ )
{
  // the final vertex buffer is not explicitly stored (save memory)
  // so it is created by reading the input data and placing it at the right position in the vbo

  int batchSize = _range;

  // clamp batch size
  if ((_range < 0) || ( static_cast<size_t>(_offset + batchSize) > numDrawVerts_) )
    batchSize = numDrawVerts_ - _offset;

  char* bdst = (char*)_dst;

  for (int i = 0; i < batchSize; ++i)
  {
    getVertex(i + _offset, bdst + decl_.getVertexStride() * i);
  }
}

struct MeshCompiler_EdgeTriMapKey
{
  // ordered vertex ids of the edge:  e0 < e1
  int e0, e1;

  MeshCompiler_EdgeTriMapKey(int a, int b)
    : e0(std::min(a,b)), e1(std::max(a,b)) {}

  bool operator ==(const MeshCompiler_EdgeTriMapKey& k) const
  {
    return e0 == k.e0 && e1 == k.e1;
  }

  bool operator <(const MeshCompiler_EdgeTriMapKey& k) const
  {
    if (e0 < k.e0)
      return true;

    if (e0 > k.e0)
      return false;

    return e1 < k.e1;
  }
};


#ifdef ACG_MC_USE_STL_HASH
// requires c++0x
struct MeshCompiler_EdgeTriMapKey_Hash
{
  std::size_t operator()(const MeshCompiler_EdgeTriMapKey& k) const
  {
    return ((std::hash<int>()(k.e0) << 1) ^ std::hash<int>()(k.e1)) >> 1;
  }
};
#else
uint qHash(const MeshCompiler_EdgeTriMapKey& k)
{
  return ((::qHash(k.e0) << 1) ^ ::qHash(k.e1)) >> 1;
}
#endif // ACG_MC_USE_STL_HASH


void MeshCompiler::getIndexAdjBuffer(void* _dst, const int _borderIndex /* = -1 */)
{
  int* idst = (int*)_dst;

  for (int i = 0; i < numTris_; ++i)
  {
    // initialize all triangles
    idst[i*6] = getIndex(i*3);
    idst[i*6+2] = getIndex(i*3+1);
    idst[i*6+4] = getIndex(i*3+2);
    
    idst[i*6+1] = _borderIndex;
    idst[i*6+3] = _borderIndex;
    idst[i*6+5] = _borderIndex;
  }

  // map from edge -> adjacent tris
  QHash<MeshCompiler_EdgeTriMapKey, std::pair<int, int> > EdgeAdjMap;

  for (int i = 0; i < numTris_; ++i)
  {
    // get input positions of triangle
    int PosIdx[3];

    for (int k = 0; k < 3; ++k)
    {
      int faceId, cornerId;
      PosIdx[k] = mapToOriginalVertexID(getIndex(i*3+k), faceId, cornerId);
    }

    // build edge->triangle adjacency map
    for (int e = 0; e < 3; ++e)
    {
      MeshCompiler_EdgeTriMapKey edge( PosIdx[e], PosIdx[(e+1)%3] );

      QHash< MeshCompiler_EdgeTriMapKey, std::pair<int, int> >::iterator itKey;

      itKey = EdgeAdjMap.find(edge);

      if (itKey == EdgeAdjMap.end())
        EdgeAdjMap[edge] = std::pair<int, int>(i, -1);
      else
        itKey.value().second = i;
    }

  }

  // use edge->triangle adjacency map to build index buffer with adjacency
  for (QHash< MeshCompiler_EdgeTriMapKey, std::pair<int, int> >::iterator it = EdgeAdjMap.begin(); it != EdgeAdjMap.end(); ++it)
  {
    // two adjacent tris of current edge
    int Tris[2] = {it.value().first, it.value().second};

    // get input positions of triangles
    int PosIdx[6];

    for (int k = 0; k < 3; ++k)
    {
      int faceId, cornerId;

      PosIdx[k] = mapToOriginalVertexID(getIndex(Tris[0]*3+k), faceId, cornerId);

      if (Tris[1] >= 0)
        PosIdx[3+k] = mapToOriginalVertexID(getIndex(Tris[1]*3+k), faceId, cornerId);
    }


    // find edge and opposite corner wrt. to adjacent tris

    int TriEdges[2] = {-1,-1};
    int TriOppositeVerts[2] = {-1,-1};

    for (int e = 0; e < 3; ++e)
    {
      MeshCompiler_EdgeTriMapKey edge0( PosIdx[e], PosIdx[(e+1)%3] );

      if (edge0 == it.key())
      {
        TriEdges[0] = e;
        TriOppositeVerts[1] = (e+2)%3;
      }
    }

    if (Tris[1] >= 0)
    {
      for (int e = 0; e < 3; ++e)
      {
        MeshCompiler_EdgeTriMapKey edge1( PosIdx[3 + e], PosIdx[3 + ((e+1)%3)] );

        if (edge1 == it.key())
        {
          TriEdges[1] = e;
          TriOppositeVerts[0] = (e+2)%3;
        }
      }
    }

    // store adjacency in buffer

    for (int i = 0; i < 2; ++i)
    {
      if (Tris[i] >= 0)
      {
        int e = TriEdges[i];

        // opposite vertex
        int ov = _borderIndex;

        if (TriOppositeVerts[i] >= 0 && Tris[(i+1)%2] >= 0)
          ov = getIndex(Tris[(i+1)%2]*3 + TriOppositeVerts[i]);

        idst[Tris[i] * 6 + e*2 + 1] = ov;
      }
    }

  }

  return;

  /*

  // alternative algorithm without hashing with multi threading
  // in case hashing method is too slow, the following unfinished implementation can be completed

  // requires vertex-face adjacency

  computeAdjacency();

#ifdef  USE_OPENMP
#pragma omp parallel for
#endif
  for (int i = 0; i < numFaces_; ++i)
  {
    const int fsize = getFaceSize(i);

    // for each edge of current face
    for (int e = 0; e < fsize; ++e)
    {
      int vertexId = getInputIndex(i, e, inputIDPos_);

      MeshCompiler_EdgeTriMapKey edge( getInputIndex(i, e, inputIDPos_), 
        getInputIndex(i, (e+1)%fsize, inputIDPos_) );


      int numAdjFaces = adjacencyVert_.getCount(vertexId);

      // search for adjacent tri with shared edge
      for (int a = 0; a < numAdjFaces; ++a)
      {
        int adjFaceID = adjacencyVert_.getAdj(vertexId, a);

        const int adjFaceSize = getFaceSize(adjFaceID);

        // for each edge of adjacent face
        for (int ae = 0; ae < adjFaceSize; ++ae)
        {
          // get start/end indices of adjacent half-edge
          MeshCompiler_EdgeTriMapKey adjEdge( getInputIndex(adjFaceID, ae, inputIDPos_), 
            getInputIndex(adjFaceID, (ae+1)%adjFaceSize, inputIDPos_) );

          // check whether this is the same edge
          if (adjEdge == edge)
          {
            // find these half-edges in output mesh




          }
        }
      }

    }


    // add adjacency for newly inserted edges (triangulation of n-polys)

  }

  // delete adjacency list since its no longer needed
  adjacencyVert_.clear();

*/
}

void MeshCompiler::getIndexAdjBuffer_MT(void* _dst, const int _borderIndex /* = -1 */)
{
  // alternative algorithm without hashing, but with multi-threading support

  // compute vertex-position -> triangle adjacency
  //  vertex-position: unsplit position IDs of input mesh
  //  triangles:  triangles in output mesh after build()

  const int numVerts = input_[inputIDPos_].count;
  AdjacencyList outputVertexTriAdj;

  outputVertexTriAdj.init(numVerts);

  // count # adjacent tris per vertex
  for (int i = 0; i < numTris_; ++i)
  {
    int faceID, cornerID;

    for (int k = 0; k < 3; ++k)
    {
      int vertex = getIndex(i * 3 + k);
      int posID = mapToOriginalVertexID(vertex, faceID, cornerID);

      outputVertexTriAdj.count[posID]++;
    }
  }

  // count num of needed entries
  int nCounter = 0;

  for (int i = 0; i < numVerts; ++i)
  {
    outputVertexTriAdj.start[i] = nCounter; // save start indices

    nCounter += outputVertexTriAdj.count[i];

    outputVertexTriAdj.count[i] = 0; // count gets recomputed in next step
  }

  // alloc memory
  outputVertexTriAdj.buf = new int[nCounter];
  outputVertexTriAdj.bufSize = nCounter;

  // build adjacency list
  for (int i = 0; i < numTris_; ++i)
  {
    int faceID, cornerID;

    for (int k = 0; k < 3; ++k)
    {
      int vertex = getIndex(i * 3 + k);
      int posID = mapToOriginalVertexID(vertex, faceID, cornerID);

      int adjIdx = outputVertexTriAdj.start[posID] + outputVertexTriAdj.count[posID]++;

      outputVertexTriAdj.buf[ adjIdx ] = i;
    }
  }


  // assemble index buffer

  int* idst = (int*)_dst;

  for (int i = 0; i < numTris_; ++i)
  {
    // initialize all triangles
    idst[i*6] = getIndex(i*3);
    idst[i*6+2] = getIndex(i*3+1);
    idst[i*6+4] = getIndex(i*3+2);

    idst[i*6+1] = _borderIndex;
    idst[i*6+3] = _borderIndex;
    idst[i*6+5] = _borderIndex;
  }

#ifdef USE_OPENMP
#pragma omp parallel for
#endif
  for (int vertexID = 0; vertexID < numVerts; ++vertexID)
  {
    const int numAdjTris = outputVertexTriAdj.getCount(vertexID);


    for (int adjID0 = 0; adjID0 < numAdjTris; ++adjID0)
    {
      const int triID0 = outputVertexTriAdj.getAdj(vertexID, adjID0);

      int TriPos0[3];
      for (int k = 0; k < 3; ++k)
      {
        int faceid, cornerid;
        TriPos0[k] = mapToOriginalVertexID( getIndex(triID0*3+k), faceid, cornerid );
      }

      for (int adjID1 = adjID0+1; adjID1 < numAdjTris; ++adjID1) //for (int triID1 = 0; triID1 < numTris_; ++triID1)
      {
        const int triID1 = outputVertexTriAdj.getAdj(vertexID, adjID1);

        int TriPos1[3];
        for (int k = 0; k < 3; ++k)
        {
          int faceid, cornerid;
          TriPos1[k] = mapToOriginalVertexID( getIndex(triID1*3+k), faceid, cornerid );
        }

        // find shared edge
        for (int e0 = 0; e0 < 3; ++e0)
        {
          MeshCompiler_EdgeTriMapKey edge0(TriPos0[e0], TriPos0[(e0+1)%3]);

          for (int e1 = 0; e1 < 3; ++e1)
          {
            MeshCompiler_EdgeTriMapKey edge1(TriPos1[e1], TriPos1[(e1+1)%3]);

            if (edge0 == edge1)
            {
              // found shared edge

              int oppVertex0 = getIndex( triID1*3 + ((e1+2)%3) );
              int oppVertex1 = getIndex( triID0*3 + ((e0+2)%3) );

              idst[triID0*6 + e0*2 +1] = oppVertex0;
              idst[triID1*6 + e1*2 + 1] = oppVertex1;
            }
          }

        }

      }

    }

  }
}

void MeshCompiler::getIndexAdjBuffer_BruteForce( void* _dst, const int _borderIndex /*= -1*/ )
{
  // simple, but slow algorithm for creating a reference and testing the improved implementations (getIndexAdjBuffer and getIndexAdjBuffer_MT)

  int* idst = (int*)_dst;

  for (int i = 0; i < numTris_; ++i)
  {
    // initialize all triangles
    idst[i*6] = getIndex(i*3);
    idst[i*6+2] = getIndex(i*3+1);
    idst[i*6+4] = getIndex(i*3+2);

    idst[i*6+1] = _borderIndex;
    idst[i*6+3] = _borderIndex;
    idst[i*6+5] = _borderIndex;
  }

  // build brute-force adjacency list and compare -- O(n^2)
#ifdef USE_OPENMP
#pragma omp parallel for
#endif
  for (int triID0 = 0; triID0 < numTris_; ++triID0)
  {
    int TriPos0[3];
    for (int k = 0; k < 3; ++k)
    {
      int faceid, cornerid;
      TriPos0[k] = mapToOriginalVertexID( getIndex(triID0*3+k), faceid, cornerid );
    }

    for (int triID1 = triID0 + 1; triID1 < numTris_; ++triID1)
    {
      int TriPos1[3];
      for (int k = 0; k < 3; ++k)
      {
        int faceid, cornerid;
        TriPos1[k] = mapToOriginalVertexID( getIndex(triID1*3+k), faceid, cornerid );
      }

      // find shared edge
      for (int e0 = 0; e0 < 3; ++e0)
      {
        MeshCompiler_EdgeTriMapKey edge0(TriPos0[e0], TriPos0[(e0+1)%3]);

        for (int e1 = 0; e1 < 3; ++e1)
        {
          MeshCompiler_EdgeTriMapKey edge1(TriPos1[e1], TriPos1[(e1+1)%3]);

          if (edge0 == edge1)
          {
            // found shared edge

            int oppVertex0 = getIndex( triID1*3 + ((e1+2)%3) );
            int oppVertex1 = getIndex( triID0*3 + ((e0+2)%3) );

            idst[triID0*6 + e0*2 +1] = oppVertex0;
            idst[triID1*6 + e1*2 + 1] = oppVertex1;
          }
        }

      }

    }

  }

}


int MeshCompiler::mapToOriginalFaceID( const int _i ) const
{
//   int faceID = _i;
//   if (!triOptMap_.empty())
//     faceID = triOptMap_[faceID];
// 
//   if (!triToSortFaceMap_.empty())
//     faceID = triToSortFaceMap_[faceID];
// 
//   if (!faceSortMap_.empty())
//     faceID = faceSortMap_[faceID];
// 
//   return faceID;

  return triToFaceMap_[_i];
}

const int* MeshCompiler::mapToOriginalFaceIDPtr() const
{
  if (triToFaceMap_.empty())
    return 0;
  else 
    return &triToFaceMap_[0];
}



int MeshCompiler::mapToOriginalVertexID( const size_t _i, int& _faceID, int& _cornerID ) const
{
  int positionID = -1;

  if (_i < numDrawVerts_ - numIsolatedVerts_)
  {
    // connected vertex
    _faceID = vertexMapFace_[_i];
    _cornerID = vertexMapCorner_[_i];

    positionID = getInputIndex(_faceID, _cornerID, inputIDPos_);
  }
  else
  {
    // isolated vertex: vertexMap stores input position id instead of face
    positionID = vertexMapFace_[_i];

    _faceID = -1;
    _cornerID = -1;
  }

  return positionID;
}

int MeshCompiler::mapToDrawVertexID( const int _faceID, const int _cornerID ) const
{
  return getInputIndexSplit(_faceID, _cornerID);
}

int MeshCompiler::mapToDrawTriID( const int _faceID, const int _k /*= 0*/, int* _numTrisOut /*= 0*/ ) const
{
  const int fsize = getFaceSize(_faceID);
  assert(_k < fsize - 2);

  if (_numTrisOut)
    *_numTrisOut = fsize - 2;

  int offset = constantFaceSize_ ? (_faceID * (maxFaceSize_ - 2)) : faceToTriMapOffset_[_faceID];

  return faceToTriMap_[offset + _k];
}

size_t MeshCompiler::getMemoryUsage(bool _printConsole) const
{
  /*
  Short evaluation of memory footprint (version at 27.12.13)

  Memory consumption was measured with xyzrgb_dragon.ply from the Stanford repository (7.2 mil tris, 3.6 mil verts).
   Options used for build(): welding - true, optimize - true, require per face attributes - true,
   Adjacency information was generated by MeshCompiler.
   Input data by user: 275 mb (indices, float3 positions, float3 normals)

  Breakdown of memory footprint in build():
  1. 467 mb - adjacency lists ( vertex-face list 98 mb, face-face list 369 mb)
  2. 103 mb - vertex weld map
  3.  82 mb - each of: faceBufSplit_, indices_
  4.  51 mb - vertexMap
  5.  41 mb - splitter_
  6.  27 mb - each of: faceToTriMap_, triToFaceMap_, faceSortMap_, ... ie. any per face map
  7.  13 mb - faceGroupIds
  8.   6 mb - faceRotCount

  Peak memory load: 563 mb after computing face-face adjacency for forceUnsharedVertex() (adjacency lists and faceBufSplit_ were allocated at that time)
  Final memory load: 292 mb measured after build(), mem allocations: [faceGroupIds, faceBufSplit, vertexMap, faceToTriMap, triToFaceMap, indices]

  Update: (12.01.14)
  forceUnsharedVertex() has been modified and no longer requires a face-face adjacency list, which was the biggest memory buffer allocated by build()
  */


  size_t usage = faceStart_.size() * 4;
  usage += faceSize_.size() * 1;
  usage += faceData_.size() * 4;
  usage += faceGroupIDs_.size() * 2;
  usage += faceBufSplit_.size() * 4;
  usage += faceSortMap_.size() * 4;

  usage += triIndexBuffer_.size() * 4;
  
  usage += subsetIDMap_.size() * sizeof(Subset);

  usage += vertexWeldMapFace_.size() * 5;

  usage += adjacencyVert_.bufSize * 4; // buf
  usage += adjacencyVert_.num * 4; // start
  usage += adjacencyVert_.num * 1; // count


  if (splitter_)
    usage += splitter_->splits.size() * 4;

  usage += triToSortFaceMap_.size() * 4;
  usage += triOptMap_.size() * 4;
  usage += vertexMapFace_.size() * 5;
  usage += faceToTriMap_.size() * 4;
  usage += faceToTriMapOffset_.size() * 4;
  usage += triToFaceMap_.size() * 4;

  usage += indices_.size() * 4; // indices_



  if (_printConsole)
  {
    const int byteToMB = 1024 * 1024;

    std::cout << "faceStart_: " << sizeof(std::pair<int, unsigned char>) << std::endl;

    std::cout << "faceStart_: " << faceStart_.size() * 4 / byteToMB << std::endl;
    std::cout << "faceSize_: " << faceSize_.size() * 1 / byteToMB << std::endl;
    std::cout << "faceData_: " << faceData_.size() * 4 / byteToMB << std::endl;
    std::cout << "faceGroupIDs_: " << faceGroupIDs_.size() * 2 / byteToMB << std::endl;
    std::cout << "faceBufSplit_: " << faceBufSplit_.size() * 4 / byteToMB << std::endl;

    std::cout << "faceSortMap_: " << faceSortMap_.size() * 4 / byteToMB << std::endl;

    std::cout << "triIndexBuffer_: " << triIndexBuffer_.size() * 4 / byteToMB << std::endl;

    std::cout << "vertexWeldMap_: " << vertexWeldMapFace_.size() * 5 / byteToMB << std::endl;

    std::cout << "adjacencyVert_: buf = " << adjacencyVert_.bufSize * 4 / byteToMB <<
      ", offset = " << adjacencyVert_.num * 4 / byteToMB <<
      ", count = "  << adjacencyVert_.num * 1 / byteToMB << std::endl;

    if (splitter_)
      std::cout << "splitter_: " << splitter_->splits.size() * 4 / byteToMB << std::endl;

    std::cout << "triToSortFaceMap_: " << triToSortFaceMap_.size() *  4 / byteToMB << std::endl;
    std::cout << "triOptMap_: " << triOptMap_.size() *  4 / byteToMB << std::endl;
    std::cout << "vertexMap_: " << vertexMapFace_.size() * 5 / byteToMB << std::endl;
    std::cout << "faceToTriMapOffset_: " << faceToTriMapOffset_.size() * 4 / byteToMB << std::endl;
    std::cout << "faceToTriMap_: " << faceToTriMap_.size() * 4 / byteToMB << std::endl;
    std::cout << "triToFaceMap_: " << triToFaceMap_.size() * 4 / byteToMB << std::endl;


    std::cout << "indices_: " << 3 * numTris_ * 4 / byteToMB << std::endl;

  }


  return usage;
}

std::string MeshCompiler::checkInputData() const
{
  // find common errors and give a short description on what needs to be fixed for the given input

  if (!faceInput_)
    return "Error: no face input data present\n";

  std::stringstream strm;

  int faceV[16];

  for (int a = 0; a < numAttributes_; ++a)
  {
    if (input_[a].count <= 0)
      strm << "Warning: input buffer is not initialized: buffer id " << a << "\n";
  }

  for (int i = 0; i < faceInput_->getNumFaces(); ++i)
  {
    if (faceInput_->getFaceSize(i) < 3)
      strm << "Error: face size too small: face " << i << ", size " << faceInput_->getFaceSize(i) << " must be at least 3\n";

    std::map<int, int> facePositions;

    for (int k = 0; k < faceInput_->getFaceSize(i); ++k)
    {
      getInputFaceVertex(i, k, faceV);

      for (int a = 0; a < numAttributes_; ++a)
      {
        if (input_[a].count > 0)
        {
          // index boundary check
          if (faceV[a] >= input_[a].count)
            strm << "Error: input index (face/corner/attribute: " << i << "/" << k << "/" << a << ") invalid: " << faceV[a] << " >= buffer size (" << input_[a].count << ")\n";
          if (faceV[a] < 0)
            strm << "Error: input index (face/corner/attribute: " << i << "/" << k << "/" << a << ") invalid: " << faceV[a] << "\n";
        }
      }

      if (numAttributes_)
        facePositions[faceV[0]] = k;
    }

    // degenerate check
    if ((int)facePositions.size() != faceInput_->getFaceSize(i))
      strm << "Warning: degenerate face " << i << "\n";

    // empty face
    if (!faceInput_->getFaceSize(i))
      strm << "Warning: empty face " << i << "\n";
  }

  return strm.str();
}

void MeshCompiler::prepareData()
{
  // update face count, in case not provided by user
  numFaces_ = faceInput_->getNumFaces();
  numIndices_ = faceInput_->getNumIndices();

  // clip empty faces from the end (user estimated too many faces)
  for (int i = numFaces_-1; i >= 0 && !faceInput_->getFaceSize(i); --i)
    --numFaces_;

  // update size of each attribute (necessary for correct handling of missing input buffers)
  for (unsigned int i = 0; i < decl_.getNumElements(); ++i)
  {
    int attrSize = (int)VertexDeclaration::getElementSize(decl_.getElement(i));
    assert(input_[i].attrSize <= 0 || input_[i].attrSize == attrSize);
    input_[i].attrSize = attrSize;
  }

  // build internal face-offset table for storing interleaved index buffer

  int minFaceSize = 99999999;
  for (int i = 0; i < numFaces_; ++i)
  {
    const int fsize = faceInput_->getFaceSize(i);

    maxFaceSize_ = std::max((int)maxFaceSize_, fsize);
    minFaceSize  = std::min(minFaceSize, fsize);
  }
  
  // faces with less than 3 vertices (lines or points) shouldn't be included in the first place
  // these parts should be handled separately, as they don't need all the costly preprocessing of MeshCompiler
  if (minFaceSize < 3)
    std::cout << "error: input contains faces with less than 3 vertices! MeshCompiler only works for pure surface meshes!" << std::endl;

  assert(minFaceSize >= 3);

  if (minFaceSize < (int)maxFaceSize_)
  {
    int curOffset = 0;

    // internal face-offset buffer required for mesh processing
    faceStart_.resize(numFaces_, -1);
    faceSize_.resize(numFaces_, 0);

    for (int i = 0; i < numFaces_; ++i)
    {
      faceSize_[i] = faceInput_->getFaceSize(i);
      faceStart_[i] = curOffset;
      curOffset += faceSize_[i];
    }

    if (numIndices_ <= 0)
      numIndices_ = curOffset;
  }
  else
  {
    constantFaceSize_ = true;
    numIndices_ = maxFaceSize_ * numFaces_;
  }


  // reset mappings from previous builds()
  triOptMap_.clear();
  triToSortFaceMap_.clear();
  faceSortMap_.clear();

  faceToTriMap_.clear();
  faceToTriMapOffset_.clear();

  triToFaceMap_.clear();
  vertexMapFace_.clear();
  vertexMapCorner_.clear();

  faceBufSplit_.clear();

  isolatedVertices_.clear();
  numIsolatedVerts_ = 0;

  if (!provokingVertexSetByUser_)
    provokingVertex_ = -1;


#ifdef _DEBUG
  for (int i = 0; i < numFaces_; ++i)
  {
    for (int k = 0; k < getFaceSize(i); ++k)
    {
      int v0 = faceInput_->getSingleFaceAttr(i, k, inputIDPos_);
      int v1 = getInputIndex(i, k, inputIDPos_);

      assert(v0 == v1);
    }
  }
#endif // _DEBUG
}

void MeshCompiler::setFaceInput( MeshCompilerFaceInput* _faceInput )
{
  faceInput_ = _faceInput;
}

void MeshCompiler::setIndexBufferInterleaved( int _numTris, int _indexSize, const void* _indices )
{
  assert(_indices);
  assert(_indexSize);

  setNumFaces(_numTris, _numTris * 3);

  for (int i = 0; i < _numTris; ++i)
  {
    int tri[3] = {-1, -1, -1};

    for (int k = 0; k < 3; ++k)
    {
      int offs = i * 3 + k;

      switch(_indexSize)
      {
      case 1: tri[k] = ((char*)_indices)[offs]; break;
      case 2: tri[k] = ((short*)_indices)[offs]; break;
      case 4: tri[k] = ((int*)_indices)[offs]; break;
      case 8: tri[k] = (int)((long long*)_indices)[offs]; break;
      default: break;
      }
    }

    setFaceVerts(i, 3, tri);
  }
}

int MeshCompiler::getNumFaces() const
{
  return numFaces_;
}

void MeshCompiler::setProvokingVertex( int _v )
{
  provokingVertex_ = _v;
  provokingVertexSetByUser_ = true;
}

int MeshCompiler::getAdjVertexFaceCount( int _vertexID ) const
{
  return adjacencyVert_.num ? adjacencyVert_.getCount(_vertexID) : faceInput_->getVertexAdjCount(_vertexID);
}

int MeshCompiler::getAdjVertexFace( int _vertexID, int _k ) const
{
  return adjacencyVert_.num ? adjacencyVert_.getAdj(_vertexID, _k) : faceInput_->getVertexAdjFace(_vertexID, _k);
}

bool MeshCompiler::isTriangleMesh() const
{
  return maxFaceSize_ == 3;
}

bool MeshCompiler::isFaceEdge( const int _triID, const int _edge ) const
{
  assert(_edge >= 0);
  assert(_edge < 3);

  if (maxFaceSize_ <= 3) return true;

  // brute force: search for triangle edge in input face

  const int faceID = mapToOriginalFaceID(_triID);
  const int fsize = getFaceSize(faceID);

  // get all draw vertices of face
  std::vector<int> FaceVerts(fsize);

  for (int i = 0; i < fsize; ++i)
    FaceVerts[i] = mapToDrawVertexID(faceID, i);

  int edgeStart = -1;
  int edgeEnd = -1;

  switch (_edge)
  {
  case 0: edgeStart = 0; edgeEnd = 1; break;
  case 1: edgeStart = 1; edgeEnd = 2; break;
  case 2: edgeStart = 2; edgeEnd = 0; break;
  default: break;
  }

  // access index buffer of triangle
  edgeStart = getIndex(_triID * 3 + edgeStart);
  edgeEnd = getIndex(_triID * 3 + edgeEnd);

  // search for edge in face vertices
  for (int e = 0; e < fsize; ++e)
  {
    if (FaceVerts[e] == edgeStart && FaceVerts[(e+1)%fsize] == edgeEnd)
      return true;
    if (FaceVerts[e] == edgeEnd && FaceVerts[(e+1)%fsize] == edgeStart)
      return true;
  }

  return false;
}

void MeshCompilerDefaultFaceInput::dbgWriteToObjFile(FILE* _file, int _posAttrID, int _normalAttrID, int _texcAttrID)
{
  for (int i = 0; i < numFaces_; ++i)
  {
    std::string strLine = "f ";

    int offset = faceOffset_[i];
    int size = faceSize_[i];

    char tmp[0xff];
    for (int k = 0; k < size; ++k)
    {
      if (_normalAttrID>=0 && _texcAttrID>=0)
        sprintf(tmp, "%d/%d/%d ", faceData_[_posAttrID][offset+k]+1, 
                                    faceData_[_texcAttrID][offset+k]+1,
                                    faceData_[_normalAttrID][offset+k]+1);
      else if (_normalAttrID >= 0)
        sprintf(tmp, "%d//%d ", faceData_[_posAttrID][offset + k]+1, faceData_[_normalAttrID][offset+k]+1);
      else if (_texcAttrID >= 0)
        sprintf(tmp, "%d/%d ", faceData_[_posAttrID][offset + k]+1, faceData_[_texcAttrID][offset+k]+1);
      else
        sprintf(tmp, "%d ", faceData_[_posAttrID][offset + k]+1);

      strLine += tmp;
    }

    fprintf(_file, "%s\n", strLine.c_str());
  }
}

bool MeshCompilerDefaultFaceInput::getFaceAttr( int _faceID, int _attrID, int* _out ) const
{
  int offset = faceOffset_[_faceID];
  int fsize = faceSize_[_faceID];

  for (int i = 0; i < fsize; ++i)
    _out[i] = faceData_[_attrID][offset+i];

  return true;
}

int MeshCompilerDefaultFaceInput::getSingleFaceAttr( const int _faceID, const int _faceCorner, const int _attrID ) const
{
  // Offset + _faceCorner
  const int pos = faceOffset_[_faceID] + _faceCorner;
  assert(_faceCorner < getFaceSize(_faceID));


  if (faceData_[_attrID].empty() || pos >= int(faceData_[_attrID].size()))
    return -1;

  return faceData_[_attrID][pos];
}

void MeshCompilerDefaultFaceInput::setFaceData( int _faceID, int _size, int* _data, int _attrID /*= 0*/ )
{
  // update face count
  if (numFaces_ <= _faceID)
    numFaces_ = _faceID + 1;

  // reserve mem
  if (faceData_[_attrID].capacity() == 0)
    faceData_[_attrID].reserve(numIndices_);

  if ((int)faceOffset_.size() <= _faceID)
    faceOffset_.resize(_faceID+1, -1);

  if ((int)faceSize_.size() <= _faceID)
    faceSize_.resize(_faceID+1, -1);

  faceSize_[_faceID] = _size;

  if (faceOffset_[_faceID] < 0)
  {
    // append new face data to stream
    faceOffset_[_faceID] = faceData_[_attrID].size();
    for (int i = 0; i < _size; ++i)
      faceData_[_attrID].push_back(_data[i]);
  }
  else
  {
    int offs = faceOffset_[_faceID];

    if ((int)faceData_[_attrID].size() == offs)
    {
      for (int i = 0; i < _size; ++i)
        faceData_[_attrID].push_back(_data[i]);
    }
    else
    {
      // resize array
      if ((int)faceData_[_attrID].size() < offs + _size + 1)
        faceData_[_attrID].resize(offs + _size + 1, -1);

      // update existing face data
      for (int i = 0; i < _size; ++i)
        faceData_[_attrID][offs + i] = _data[i];
    }
  }
}

int MeshCompilerFaceInput::getSingleFaceAttr( const int _faceID, const int _faceCorner, const int _attrID )  const
{
  const int fsize = getFaceSize(_faceID);

  assert(_faceID < getNumFaces());
  assert(_faceID >= 0);
  assert(_faceCorner >= 0);
  assert(_faceCorner < fsize);

  int retVal = -1;

  int* tmpPtr = getFaceAttr(_faceID, _attrID);

  if (tmpPtr)
    retVal = tmpPtr[_faceCorner];
  else
  {
    tmpPtr = new int[fsize];
    if (getFaceAttr(_faceID, _attrID, tmpPtr))
      retVal = tmpPtr[_faceCorner];
    delete [] tmpPtr;
  }

  return retVal;
}


bool MeshCompilerVertexCompare::equalVertex( const void* v0, const void* v1, const VertexDeclaration* _decl )
{
  // compare vertex data of two vertices

  assert(_decl);
  assert(v0);
  assert(v1);

  const int nElements = (int)_decl->getNumElements();
  for (int i = 0; i < nElements; ++i)
  {
    const VertexElement* el = _decl->getElement(i);

    // pointer to element data
    const void* el_0 = static_cast<const char*>(v0) + (size_t)el->pointer_;
    const void* el_1 = static_cast<const char*>(v1) + (size_t)el->pointer_;

    // interpret element based on declaration
    switch ( el->type_ )
    {
    case GL_DOUBLE:
      {
        const double* d0 = static_cast<const double*>(el_0);
        const double* d1 = static_cast<const double*>(el_1);

        double diff = 0.0;

        for (int k = 0; k < (int)el->numElements_; ++k)
          diff += fabs(d0[k] - d1[k]);

        if (diff > d_eps_)
          return false;

      } break;


    case GL_FLOAT:
      {
        const float* f0 = static_cast<const float*>(el_0);
        const float* f1 = static_cast<const float*>(el_1);

        float diff = 0.0;

        for (int k = 0; k < (int)el->numElements_; ++k)
        {
          // sign-bit manipulation vs. fabsf  - no significant performance difference
//           float e = f0[k] - f1[k];
//           int e_abs = *((int*)(&e)) & 0x7FFFFFFF;
//           diff += *((float*)&e_abs);

          diff += fabsf(f0[k] - f1[k]);
        }

        if (diff > f_eps_)
          return false;
      } break;

    case GL_INT:
    case GL_UNSIGNED_INT:
      {
        const int* i0 = static_cast<const int*>(el_0);
        const int* i1 = static_cast<const int*>(el_1);

        for (int k = 0; k < (int)el->numElements_; ++k)
        {
          if (i0[k] != i1[k])
            return false;
        }
      } break;

    case GL_SHORT:
    case GL_UNSIGNED_SHORT:
      {
        const short* i0 = static_cast<const short*>(el_0);
        const short* i1 = static_cast<const short*>(el_1);

        for (int k = 0; k < (int)el->numElements_; ++k)
        {
          if (i0[k] != i1[k])
            return false;
        }
      } break;

    case GL_BYTE:
    case GL_UNSIGNED_BYTE:
      {
        const char* i0 = static_cast<const char*>(el_0);
        const char* i1 = static_cast<const char*>(el_1);

        for (int k = 0; k < (int)el->numElements_; ++k)
        {
          if (i0[k] != i1[k])
            return false;
        }
      } break;

    default: std::cerr << "MeshCompiler: equalVertex() comparision not implemented for type: " << el->type_ << std::endl;
    }
  }

  return true;
}




}