///////////////////////////////////////////////////////////////////////////////////////////
// SplitUpStripsAndOptimize()
//
// Splits the input vector of strips (allBigStrips) into smaller, cache friendly pieces, then
// reorders these pieces to maximize cache hits
// The final strips are output through outStrips
//
void NvStripifier::SplitUpStripsAndOptimize(NvStripInfoVec &allStrips, NvStripInfoVec &outStrips,
NvEdgeInfoVec& edgeInfos, NvFaceInfoVec& outFaceList)
{
int threshold = cacheSize;
NvStripInfoVec tempStrips;
//split up strips into threshold-sized pieces
for(size_t i = 0; i < allStrips.size(); i++)
{
NvStripInfo* currentStrip;
NvStripStartInfo startInfo(NULL, NULL, false);
int actualStripSize = 0;
for(size_t j = 0; j < allStrips[i]->m_faces.size(); ++j)
{
if( !IsDegenerate(allStrips[i]->m_faces[j]) )
actualStripSize++;
}
if(actualStripSize /*allStrips[i]->m_faces.size()*/ > threshold)
{
int numTimes = actualStripSize /*allStrips[i]->m_faces.size()*/ / threshold;
int numLeftover = actualStripSize /*allStrips[i]->m_faces.size()*/ % threshold;
int degenerateCount = 0;
int j = 0;
for(; j < numTimes; j++)
{
currentStrip = new NvStripInfo(startInfo, 0, -1);
int faceCtr = j*threshold + degenerateCount;
bool bFirstTime = true;
while(faceCtr < threshold+(j*threshold)+degenerateCount)
{
if(IsDegenerate(allStrips[i]->m_faces[faceCtr]))
{
degenerateCount++;
//last time or first time through, no need for a degenerate
if( (((faceCtr + 1) != threshold+(j*threshold)+degenerateCount) ||
((j == numTimes - 1) && (numLeftover < 4) && (numLeftover > 0))) &&
!bFirstTime)
{
currentStrip->m_faces.push_back(allStrips[i]->m_faces[faceCtr++]);
}
else
{
//but, we do need to delete the degenerate, if it's marked fake, to avoid leaking
if(allStrips[i]->m_faces[faceCtr]->m_bIsFake)
{
delete allStrips[i]->m_faces[faceCtr], allStrips[i]->m_faces[faceCtr] = NULL;
}
++faceCtr;
}
}
else
{
currentStrip->m_faces.push_back(allStrips[i]->m_faces[faceCtr++]);
bFirstTime = false;
}
}
/*
for(int faceCtr = j*threshold; faceCtr < threshold+(j*threshold); faceCtr++)
{
currentStrip->m_faces.push_back(allStrips[i]->m_faces[faceCtr]);
}
*/
if(j == numTimes - 1) //last time through
{
if( (numLeftover < 4) && (numLeftover > 0) ) //way too small
{
//just add to last strip
int ctr = 0;
while(ctr < numLeftover)
{
IsDegenerate( allStrips[i]->m_faces[faceCtr] ) ? ++degenerateCount : ++ctr;
currentStrip->m_faces.push_back(allStrips[i]->m_faces[faceCtr++]);
}
numLeftover = 0;
}
}
tempStrips.push_back(currentStrip);
}
int leftOff = j * threshold + degenerateCount;
if(numLeftover != 0)
{
currentStrip = new NvStripInfo(startInfo, 0, -1);
int ctr = 0;
bool bFirstTime = true;
while(ctr < numLeftover)
{
if( !IsDegenerate(allStrips[i]->m_faces[leftOff]) )
{
ctr++;
bFirstTime = false;
currentStrip->m_faces.push_back(allStrips[i]->m_faces[leftOff++]);
}
else if(!bFirstTime)
currentStrip->m_faces.push_back(allStrips[i]->m_faces[leftOff++]);
else
{
//don't leak
if(allStrips[i]->m_faces[leftOff]->m_bIsFake)
{
delete allStrips[i]->m_faces[leftOff], allStrips[i]->m_faces[leftOff] = NULL;
}
leftOff++;
}
}
/*
for(size_t k = 0; k < numLeftover; k++)
{
currentStrip->m_faces.push_back(allStrips[i]->m_faces[leftOff++]);
}
*/
tempStrips.push_back(currentStrip);
}
}
else
{
//we're not just doing a tempStrips.push_back(allBigStrips[i]) because
// this way we can delete allBigStrips later to free the memory
currentStrip = new NvStripInfo(startInfo, 0, -1);
for(size_t j = 0; j < allStrips[i]->m_faces.size(); j++)
currentStrip->m_faces.push_back(allStrips[i]->m_faces[j]);
tempStrips.push_back(currentStrip);
}
}
//add small strips to face list
NvStripInfoVec tempStrips2;
RemoveSmallStrips(tempStrips, tempStrips2, outFaceList);
outStrips.clear();
//screw optimization for now
// for(size_t i = 0; i < tempStrips.size(); ++i)
// outStrips.push_back(tempStrips[i]);
if(tempStrips2.size() != 0)
{
//Optimize for the vertex cache
VertexCache* vcache = new VertexCache(cacheSize);
float bestNumHits = -1.0f;
float numHits;
int bestIndex;
bool done = false;
int firstIndex = 0;
float minCost = 10000.0f;
for(size_t i = 0; i < tempStrips2.size(); i++)
{
int numNeighbors = 0;
//find strip with least number of neighbors per face
for(size_t j = 0; j < tempStrips2[i]->m_faces.size(); j++)
{
numNeighbors += NumNeighbors(tempStrips2[i]->m_faces[j], edgeInfos);
}
float currCost = (float)numNeighbors / (float)tempStrips2[i]->m_faces.size();
if(currCost < minCost)
{
minCost = currCost;
firstIndex = i;
}
}
UpdateCacheStrip(vcache, tempStrips2[firstIndex]);
outStrips.push_back(tempStrips2[firstIndex]);
tempStrips2[firstIndex]->visited = true;
bool bWantsCW = (tempStrips2[firstIndex]->m_faces.size() % 2) == 0;
//this n^2 algo is what slows down stripification so much....
// needs to be improved
while(1)
{
bestNumHits = -1.0f;
//find best strip to add next, given the current cache
for(size_t i = 0; i < tempStrips2.size(); i++)
{
if(tempStrips2[i]->visited)
continue;
numHits = CalcNumHitsStrip(vcache, tempStrips2[i]);
if(numHits > bestNumHits)
{
bestNumHits = numHits;
bestIndex = i;
}
else if(numHits >= bestNumHits)
{
//check previous strip to see if this one requires it to switch polarity
NvStripInfo *strip = tempStrips2[i];
int nStripFaceCount = strip->m_faces.size();
NvFaceInfo tFirstFace(strip->m_faces[0]->m_v0, strip->m_faces[0]->m_v1, strip->m_faces[0]->m_v2);
// If there is a second face, reorder vertices such that the
// unique vertex is first
if (nStripFaceCount > 1)
{
int nUnique = NvStripifier::GetUniqueVertexInB(strip->m_faces[1], &tFirstFace);
if (nUnique == tFirstFace.m_v1)
{
SWAP(tFirstFace.m_v0, tFirstFace.m_v1);
}
else if (nUnique == tFirstFace.m_v2)
{
SWAP(tFirstFace.m_v0, tFirstFace.m_v2);
}
// If there is a third face, reorder vertices such that the
// shared vertex is last
if (nStripFaceCount > 2)
{
int nShared0, nShared1;
GetSharedVertices(strip->m_faces[2], &tFirstFace, &nShared0, &nShared1);
if ( (nShared0 == tFirstFace.m_v1) && (nShared1 == -1) )
{
SWAP(tFirstFace.m_v1, tFirstFace.m_v2);
}
}
}
// Check CW/CCW ordering
if (bWantsCW == IsCW(strip->m_faces[0], tFirstFace.m_v0, tFirstFace.m_v1))
{
//I like this one!
bestIndex = i;
}
}
}
if(bestNumHits == -1.0f)
break;
tempStrips2[bestIndex]->visited = true;
UpdateCacheStrip(vcache, tempStrips2[bestIndex]);
outStrips.push_back(tempStrips2[bestIndex]);
bWantsCW = (tempStrips2[bestIndex]->m_faces.size() % 2 == 0) ? bWantsCW : !bWantsCW;
}
delete vcache;
}
}
///////////////////////////////////////////////////////////////////////////////////////////
// SplitUpStripsAndOptimize()
//
// Splits the input _vector_ of strips (allBigStrips) into smaller, cache friendly pieces, then
// reorders these pieces to maximize cache hits
// The final strips are output through outStrips
//
void NvStripifier::SplitUpStripsAndOptimize(NvStripInfoVec &allStrips, NvStripInfoVec &outStrips,
NvEdgeInfoVec& edgeInfos, NvFaceInfoVec& outFaceList)
{
int threshold = cacheSize;
NvStripInfoVec tempStrips;
//split up strips into threshold-sized pieces
int i;
for(i = 0; i < allStrips.size(); i++)
{
NvStripInfo* currentStrip;
NvStripStartInfo startInfo(NULL, NULL, false);
if(allStrips[i]->m_faces.size() > threshold)
{
int numTimes = allStrips[i]->m_faces.size() / threshold;
int numLeftover = allStrips[i]->m_faces.size() % threshold;
int j;
for(j = 0; j < numTimes; j++)
{
currentStrip = xr_new<NvStripInfo> (startInfo, 0, -1);
for(int faceCtr = j*threshold; faceCtr < threshold+(j*threshold); faceCtr++) {
currentStrip->m_faces.push_back(allStrips[i]->m_faces[faceCtr]);
}
tempStrips.push_back(currentStrip);
}
int leftOff = j * threshold;
if(numLeftover != 0)
{
currentStrip = xr_new<NvStripInfo> (startInfo, 0, -1);
for(int k = 0; k < numLeftover; k++)
{
currentStrip->m_faces.push_back(allStrips[i]->m_faces[leftOff++]);
}
tempStrips.push_back(currentStrip);
}
}
else
{
//we're not just doing a tempStrips.push_back(allBigStrips[i]) because
// this way we can _delete allBigStrips later to xr_free the memory
currentStrip = xr_new<NvStripInfo> (startInfo, 0, -1);
for(int j = 0; j < allStrips[i]->m_faces.size(); j++)
currentStrip->m_faces.push_back(allStrips[i]->m_faces[j]);
tempStrips.push_back(currentStrip);
}
}
//add small strips to face list
NvStripInfoVec tempStrips2;
RemoveSmallStrips(tempStrips, tempStrips2, outFaceList);
outStrips.clear();
if(tempStrips2.size() != 0)
{
//Optimize for the vertex cache
VertexCache* vcache = xr_new<VertexCache> (cacheSize);
float bestNumHits = -1.0f;
float numHits = 0;
int bestIndex = 0;
int firstIndex = 0;
float minCost = 10000.0f;
for(i = 0; i < tempStrips2.size(); i++)
{
int numNeighbors = 0;
//find strip with least number of neighbors per face
for(int j = 0; j < tempStrips2[i]->m_faces.size(); j++)
{
numNeighbors += NumNeighbors(tempStrips2[i]->m_faces[j], edgeInfos);
}
float currCost = (float)numNeighbors / (float)tempStrips2[i]->m_faces.size();
if(currCost < minCost)
{
minCost = currCost;
firstIndex = i;
}
}
UpdateCacheStrip(vcache, tempStrips2[firstIndex]);
outStrips.push_back(tempStrips2[firstIndex]);
tempStrips2[firstIndex]->visited = true;
//this n^2 algo is what slows down stripification so much....
// needs to be improved
while(1)
{
bestNumHits = -1.0f;
//find best strip to add next, given the current cache
for(int i = 0; i < tempStrips2.size(); i++)
{
if(tempStrips2[i]->visited)
continue;
numHits = CalcNumHitsStrip(vcache, tempStrips2[i]);
if(numHits > bestNumHits)
{
bestNumHits = numHits;
bestIndex = i;
}
}
if(bestNumHits == -1.0f)
break;
tempStrips2[bestIndex]->visited = true;
UpdateCacheStrip(vcache, tempStrips2[bestIndex]);
outStrips.push_back(tempStrips2[bestIndex]);
}
xr_delete(vcache);
}
}
