src/yafraycore/ray_kdtree.cc

Go to the documentation of this file.

// search for "todo" and "IMPLEMENT" and "<<" or ">>"...

#include <yafraycore/ray_kdtree.h>
#include <core_api/material.h>
#include <core_api/scene.h>
#include <stdexcept>
//#include <math.h>
#include <limits>
#include <set>
#if ( HAVE_PTHREAD && defined (__GNUC__) )
#include <ext/mt_allocator.h>
#endif
#include <time.h>

__BEGIN_YAFRAY

#define LOWER_B 0
#define UPPER_B 2
#define BOTH_B  1

#define _TRI_CLIP 1 //tempoarily disabled
#define TRI_CLIP_THRESH 32
#define CLIP_DATA_SIZE (3*12*sizeof(double))
#define KD_BINS 1024

#define KD_MAX_STACK 64

// #define Y_MIN3(a,b,c) ( ((a)>(b)) ? ( ((b)>(c))?(c):(b)):( ((a)>(c))?(c):(a)) )
// #define Y_MAX3(a,b,c) ( ((a)<(b)) ? ( ((b)>(c))?(b):(c)):( ((a)>(c))?(a):(c)) )

#if (defined(_M_IX86) || defined(i386) || defined(_X86_))
        #define Y_FAST_INT 1
#else
        #define Y_FAST_INT 0
#endif

#define _doublemagicroundeps          (0.5 - 1.4e-11)

inline int Y_Round2Int(double val) {
#if Y_FAST_INT > 0
        union { double f; int i[2]; } u;
        u.f     = val + 6755399441055744.0; //2^52 * 1.5,  uses limited precision to floor
        return u.i[0];
#else
        return int(val);
#endif
}

inline int Y_Float2Int(double val) {
#if Y_FAST_INT > 0
        return (val<0) ?  Y_Round2Int( val+_doublemagicroundeps ) :
                   Y_Round2Int(val-_doublemagicroundeps);
#else
        return (int)val;
#endif
}

//still in old file...
//int Kd_inodes=0, Kd_leaves=0, _emptyKd_leaves=0, Kd_prims=0, _clip=0, _bad_clip=0, _null_clip=0, _early_out=0;

//bound_t getTriBound(const triangle_t tri);
//int triBoxOverlap(double boxcenter[3],double boxhalfsize[3],double triverts[3][3]);
//int triBoxClip(const double b_min[3], const double b_max[3], const double triverts[3][3], bound_t &box);

template<class T>
kdTree_t<T>::kdTree_t(const T **v, int np, int depth, int leafSize,
                        float cost_ratio, float emptyBonus)
        : costRatio(cost_ratio), eBonus(emptyBonus), maxDepth(depth)
{
        std::cout << "starting build of kd-tree ("<<np<<" prims, cr:"<<costRatio<<" eb:"<<eBonus<<")\n";
        clock_t c_start, c_end;
        c_start = clock();
        Kd_inodes=0, Kd_leaves=0, _emptyKd_leaves=0, Kd_prims=0, depthLimitReached=0, NumBadSplits=0,
                _clip=0, _bad_clip=0, _null_clip=0, _early_out=0;
        totalPrims = np;
        nextFreeNode = 0;
        allocatedNodesCount = 256;
        nodes = (rkdTreeNode<T>*)y_memalign(64, 256 * sizeof(rkdTreeNode<T>));
        if(maxDepth <= 0) maxDepth = int( 7.0f + 1.66f * log(float(totalPrims)) );
        double logLeaves = 1.442695f * log(double(totalPrims)); // = base2 log
        if(leafSize <= 0)
        {
                //maxLeafSize = int( 1.442695f * log(float(totalPrims)/62500.0) );
                int mls = int( logLeaves - 16.0 );
                if(mls <= 0) mls = 1;
                maxLeafSize = (unsigned int) mls;
        }
        else maxLeafSize = (unsigned int) leafSize;
        if(maxDepth>KD_MAX_STACK) maxDepth = KD_MAX_STACK; //to prevent our stack to overflow
        //experiment: add penalty to cost ratio to reduce memory usage on huge scenes
        if( logLeaves > 16.0 ) costRatio += 0.25*( logLeaves - 16.0 );
        allBounds = new bound_t[totalPrims + TRI_CLIP_THRESH+1];
        std::cout << "getting triangle bounds...";
        for(u_int32 i=0; i<totalPrims; i++)
        {
                allBounds[i] = v[i]->getBound();
                /* calc tree bound. Remember to upgrade bound_t class... */
                if(i) treeBound = bound_t(treeBound, allBounds[i]);
                else treeBound = allBounds[i];
        }
        //slightly(!) increase tree bound to prevent errors with prims
        //lying in a bound plane (still slight bug with trees where one dim. is 0)
        for(int i=0;i<3;i++)
        {
                double foo = (treeBound.g[i] - treeBound.a[i])*0.001;
                treeBound.a[i] -= foo, treeBound.g[i] += foo;
        }
        std::cout << "done!\n";
        // get working memory for tree construction
        boundEdge *edges[3];
        u_int32 rMemSize = 3*totalPrims; // (maxDepth+1)*totalPrims;
        u_int32 *leftPrims = new u_int32[std::max( (u_int32)2*TRI_CLIP_THRESH, totalPrims )];
        u_int32 *rightPrims = new u_int32[rMemSize]; //just a rough guess, allocating worst case is insane!
//      u_int32 *primNums = new u_int32[totalPrims]; //isn't this like...totaly unnecessary? use leftPrims?
        for (int i = 0; i < 3; ++i) edges[i] = new boundEdge[514/*2*totalPrims*/];
        clip = new int[maxDepth+2];
        cdata = (char*)y_memalign(64, (maxDepth+2)*TRI_CLIP_THRESH*CLIP_DATA_SIZE);
        
        // prepare data
        for (u_int32 i = 0; i < totalPrims; i++) leftPrims[i] = i;//primNums[i] = i;
        for (int i = 0; i < maxDepth+2; i++) clip[i] = -1;
        
        /* build tree */
        prims = v;
        std::cout << "starting recursive build...\n";
        buildTree(totalPrims, treeBound, leftPrims,
                          leftPrims, rightPrims, edges, // <= working memory
                          rMemSize, 0, 0 );
        
        // free working memory
        delete[] leftPrims;
        delete[] rightPrims;
        delete[] allBounds;
        for (int i = 0; i < 3; ++i) delete[] edges[i];
        delete[] clip;
        y_free(cdata);
        //print some stats:
        c_end = clock() - c_start;
        std::cout << "\n=== kd-tree stats ("<< float(c_end) / (float)CLOCKS_PER_SEC <<"s) ===\n";
        std::cout << "used/allocated kd-tree nodes: " << nextFreeNode << "/" << allocatedNodesCount
                << " (" << 100.f * float(nextFreeNode)/allocatedNodesCount << "%)\n";
        std::cout << "primitives in tree: " << totalPrims << std::endl;
        std::cout << "interior nodes: " << Kd_inodes << " / " << "leaf nodes: " << Kd_leaves
                << " (empty: " << _emptyKd_leaves << " = " << 100.f * float(_emptyKd_leaves)/Kd_leaves << "%)\n";
        std::cout << "leaf prims: " << Kd_prims << " (" << float(Kd_prims)/totalPrims << "x prims in tree, leaf size:"<< maxLeafSize<<")\n";
        std::cout << "   => " << float(Kd_prims)/ (Kd_leaves-_emptyKd_leaves) << " prims per non-empty leaf\n";
        std::cout << "leaves due to depth limit/bad splits: " << depthLimitReached << "/" << NumBadSplits << "\n";
        std::cout << "clipped triangles: " << _clip << " (" <<_bad_clip << " bad clips, "<<_null_clip
                <<" null clips)\n";
        //std::cout << "early outs: " << _early_out << "\n\n";
}

template<class T>
kdTree_t<T>::~kdTree_t()
{
//      std::cout << "kd-tree destructor: freeing nodes...";
        y_free(nodes);
//      std::cout << "done!\n";
        //y_free(prims); //überflüssig?
}

// ============================================================
template<class T>
void kdTree_t<T>::pigeonMinCost(u_int32 nPrims, bound_t &nodeBound, u_int32 *primIdx, splitCost_t &split)
{
        bin_t bin[ KD_BINS+1 ];
        PFLOAT d[3];
        d[0] = nodeBound.longX();
        d[1] = nodeBound.longY();
        d[2] = nodeBound.longZ();
        split.oldCost = float(nPrims);
        split.bestCost = std::numeric_limits<PFLOAT>::infinity();
        float invTotalSA = 1.0f / (d[0]*d[1] + d[0]*d[2] + d[1]*d[2]);
        PFLOAT t_low, t_up;
        int b_left, b_right;
        
        for(int axis=0;axis<3;axis++)
        {
                PFLOAT s = KD_BINS/d[axis];
                PFLOAT min = nodeBound.a[axis];
                // pigeonhole sort:
                for(unsigned int i=0; i<nPrims; ++i)
                {
                        const bound_t &bbox = allBounds[ primIdx[i] ];
                        t_low = bbox.a[axis];
                        t_up  = bbox.g[axis];
                        b_left = (int)((t_low - min)*s);
                        b_right = (int)((t_up - min)*s);
//                      b_left = Y_Round2Int( ((t_low - min)*s) );
//                      b_right = Y_Round2Int( ((t_up - min)*s) );
                        if(b_left<0) b_left=0; else if(b_left > KD_BINS) b_left = KD_BINS;
                        if(b_right<0) b_right=0; else if(b_right > KD_BINS) b_right = KD_BINS;
                        
                        if(t_low == t_up)
                        {
                                if(bin[b_left].empty() || (t_low >= bin[b_left].t && !bin[b_left].empty() ) )
                                {
                                        bin[b_left].t = t_low;
                                        bin[b_left].c_both++;
                                }
                                else
                                {
                                        bin[b_left].c_left++;
                                        bin[b_left].c_right++;
                                }
                                bin[b_left].n += 2;
                        }
                        else
                        {       
                                if(bin[b_left].empty() || (t_low > bin[b_left].t  && !bin[b_left].empty() ) )
                                {
                                        bin[b_left].t = t_low;
                                        bin[b_left].c_left += bin[b_left].c_both + bin[b_left].c_bleft;
                                        bin[b_left].c_right += bin[b_left].c_both;
                                        bin[b_left].c_both = bin[b_left].c_bleft = 0;
                                        bin[b_left].c_bleft++;
                                }
                                else if(t_low == bin[b_left].t)
                                {
                                        bin[b_left].c_bleft++;
                                }
                                else bin[b_left].c_left++;
                                bin[b_left].n++;
                                
                                bin[b_right].c_right++;
                                if(bin[b_right].empty() || t_up > bin[b_right].t)
                                {
                                        bin[b_right].t = t_up;
                                        bin[b_right].c_left += bin[b_right].c_both + bin[b_right].c_bleft;
                                        bin[b_right].c_right += bin[b_right].c_both;
                                        bin[b_right].c_both = bin[b_right].c_bleft = 0;
                                }
                                bin[b_right].n++;
                        }

                }
                
                const int axisLUT[3][3] = { {0,1,2}, {1,2,0}, {2,0,1} };
                float capArea = d[ axisLUT[1][axis] ] * d[ axisLUT[2][axis] ];
                float capPerim = d[ axisLUT[1][axis] ] + d[ axisLUT[2][axis] ];
                
                unsigned int nBelow=0, nAbove=nPrims;
                // cumulate prims and evaluate cost
                for(int i=0; i<KD_BINS+1; ++i)
                {
                        if(!bin[i].empty())
                        {       
                                nBelow += bin[i].c_left;
                                nAbove -= bin[i].c_right;
                                // cost:
                                PFLOAT edget = bin[i].t;
                                if (edget > nodeBound.a[axis] &&
                                        edget < nodeBound.g[axis]) {
                                        // Compute cost for split at _i_th edge
                                        float l1 = edget - nodeBound.a[axis];
                                        float l2 = nodeBound.g[axis] - edget;
                                        float belowSA = capArea + l1*capPerim;
                                        float aboveSA = capArea + l2*capPerim;
                                        float rawCosts = (belowSA * nBelow + aboveSA * nAbove);
                                        //float eb = (nAbove == 0 || nBelow == 0) ? eBonus*rawCosts : 0.f;
                                        float eb;
                                        if(nAbove == 0) eb = (0.1f + l2/d[axis])*eBonus*rawCosts;
                                        else if(nBelow == 0) eb = (0.1f + l1/d[axis])*eBonus*rawCosts;
                                        else eb = 0.0f;
                                        float cost = costRatio + invTotalSA * (rawCosts - eb);
                                        // Update best split if this is lowest cost so far
                                        if (cost < split.bestCost)  {
                                                split.t = edget;
                                                split.bestCost = cost;
                                                split.bestAxis = axis;
                                                split.bestOffset = i; // kinda useless...
                                                split.nBelow = nBelow;
                                                split.nAbove = nAbove;
                                        }
                                }
                                nBelow += bin[i].c_both + bin[i].c_bleft;
                                nAbove -= bin[i].c_both;
                        }
                } // for all bins
                if(nBelow != nPrims || nAbove != 0)
                {
                        int c1=0, c2=0, c3=0, c4=0, c5=0;
                        std::cout << "SCREWED!!\n";
                        for(int i=0;i<KD_BINS+1;i++){ c1+= bin[i].n; std::cout << bin[i].n << " ";}
                        std::cout << "\nn total: "<< c1 << "\n";
                        for(int i=0;i<KD_BINS+1;i++){ c2+= bin[i].c_left; std::cout << bin[i].c_left << " ";}
                        std::cout << "\nc_left total: "<< c2 << "\n";
                        for(int i=0;i<KD_BINS+1;i++){ c3+= bin[i].c_bleft; std::cout << bin[i].c_bleft << " ";}
                        std::cout << "\nc_bleft total: "<< c3 << "\n";
                        for(int i=0;i<KD_BINS+1;i++){ c4+= bin[i].c_both; std::cout << bin[i].c_both << " ";}
                        std::cout << "\nc_both total: "<< c4 << "\n";
                        for(int i=0;i<KD_BINS+1;i++){ c5+= bin[i].c_right; std::cout << bin[i].c_right << " ";}
                        std::cout << "\nc_right total: "<< c5 << "\n";
                        std::cout << "\nnPrims: "<<nPrims<<" nBelow: "<<nBelow<<" nAbove: "<<nAbove<<"\n";
                        std::cout << "total left: " << c2 + c3 + c4 << "\ntotal right: " << c4 + c5 << "\n";
                        std::cout << "n/2: " << c1/2 << "\n";
                        throw std::logic_error("cost function mismatch");
                }
                for(int i=0;i<KD_BINS+1;i++) bin[i].reset();
        } // for all axis
}

// ============================================================
template<class T>
void kdTree_t<T>::minimalCost(u_int32 nPrims, bound_t &nodeBound, u_int32 *primIdx,
                const bound_t *pBounds, boundEdge *edges[3], splitCost_t &split)
{
        PFLOAT d[3];
        d[0] = nodeBound.longX();
        d[1] = nodeBound.longY();
        d[2] = nodeBound.longZ();
        split.oldCost = float(nPrims);
        split.bestCost = std::numeric_limits<PFLOAT>::infinity();
        float invTotalSA = 1.0f / (d[0]*d[1] + d[0]*d[2] + d[1]*d[2]);
        int nEdge;
        
        for(int axis=0;axis<3;axis++)
        {
                // << get edges for axis >>
                int pn;
                nEdge=0;
                //test!
                if(pBounds!=allBounds) for (unsigned int i=0; i < nPrims; i++)
                {
                        pn = primIdx[i];
                        const bound_t &bbox = pBounds[i];
                        if(bbox.a[axis] == bbox.g[axis])
                        {
                                edges[axis][nEdge] = boundEdge(bbox.a[axis], i /* pn */, BOTH_B);
                                ++nEdge;
                        }
                        else
                        {
                                edges[axis][nEdge] = boundEdge(bbox.a[axis], i /* pn */, LOWER_B);
                                edges[axis][nEdge+1] = boundEdge(bbox.g[axis], i /* pn */, UPPER_B);
                                nEdge += 2;
                        }
                }
                else for (unsigned int i=0; i < nPrims; i++)
                {
                        pn = primIdx[i];
                        const bound_t &bbox = pBounds[pn];
                        if(bbox.a[axis] == bbox.g[axis])
                        {
                                edges[axis][nEdge] = boundEdge(bbox.a[axis], pn, BOTH_B);
                                ++nEdge;
                        }
                        else
                        {
                                edges[axis][nEdge] = boundEdge(bbox.a[axis], pn, LOWER_B);
                                edges[axis][nEdge+1] = boundEdge(bbox.g[axis], pn, UPPER_B);
                                nEdge += 2;
                        }
                }
                std::sort(&edges[axis][0], &edges[axis][nEdge]);
                // Compute cost of all splits for _axis_ to find best
                const int axisLUT[3][3] = { {0,1,2}, {1,2,0}, {2,0,1} };
                float capArea = d[ axisLUT[1][axis] ] * d[ axisLUT[2][axis] ];
                float capPerim = d[ axisLUT[1][axis] ] + d[ axisLUT[2][axis] ];
                unsigned int nBelow = 0, nAbove = nPrims;
                //todo: early-out criteria: if l1 > l2*nPrims (l2 > l1*nPrims) => minimum is lowest (highest) edge!
                if(nPrims>5)
                {
                        PFLOAT edget = edges[axis][0].pos;
                        float l1 = edget - nodeBound.a[axis];
                        float l2 = nodeBound.g[axis] - edget;
                        if(l1 > l2*float(nPrims) && l2 > 0.f)
                        {
                                float rawCosts = (capArea + l2*capPerim) * nPrims;
                                float cost = costRatio + invTotalSA * (rawCosts - eBonus); //todo: use proper ebonus...
                                //optimal cost is definitely here, and nowhere else!
                                if (cost < split.bestCost)  {
                                        split.bestCost = cost;
                                        split.bestAxis = axis;
                                        split.bestOffset = 0;
                                        split.nEdge = nEdge;
                                        ++_early_out;
                                }
                                continue;
                        }
                        edget = edges[axis][nEdge-1].pos;
                        l1 = edget - nodeBound.a[axis];
                        l2 = nodeBound.g[axis] - edget;
                        if(l2 > l1*float(nPrims) && l1 > 0.f)
                        {
                                float rawCosts = (capArea + l1*capPerim) * nPrims;
                                float cost = costRatio + invTotalSA * (rawCosts - eBonus); //todo: use proper ebonus...
                                if (cost < split.bestCost)  {
                                        split.bestCost = cost;
                                        split.bestAxis = axis;
                                        split.bestOffset = nEdge-1;
                                        split.nEdge = nEdge;
                                        ++_early_out;
                                }
                                continue;
                        }
                }
                
                for (int i = 0; i < nEdge; ++i) {
                        if (edges[axis][i].end == UPPER_B) --nAbove;
                        PFLOAT edget = edges[axis][i].pos;
                        if (edget > nodeBound.a[axis] &&
                                edget < nodeBound.g[axis]) {
                                // Compute cost for split at _i_th edge
                                float l1 = edget - nodeBound.a[axis];
                                float l2 = nodeBound.g[axis] - edget;
                                float belowSA = capArea + (l1)*capPerim;
                                float aboveSA = capArea + (l2)*capPerim;
                                float rawCosts = (belowSA * nBelow + aboveSA * nAbove);
                                //float eb = (nAbove == 0 || nBelow == 0) ? eBonus*rawCosts : 0.f;
                                float eb;
                                if(nAbove == 0) eb = (0.1f + l2/d[axis])*eBonus*rawCosts;
                                else if(nBelow == 0) eb = (0.1f + l1/d[axis])*eBonus*rawCosts;
                                else eb = 0.0f;
                                float cost = costRatio + invTotalSA * (rawCosts - eb);
                                // Update best split if this is lowest cost so far
                                if (cost < split.bestCost)  {
                                        split.bestCost = cost;
                                        split.bestAxis = axis;
                                        split.bestOffset = i;
                                        split.nEdge = nEdge;
                                        //delete again:
                                        split.nBelow = nBelow;
                                        split.nAbove = nAbove;
                                }
                        }
                        if (edges[axis][i].end != UPPER_B)
                        {
                                ++nBelow;
                                if (edges[axis][i].end == BOTH_B) --nAbove;
                        }
                }
                if(nBelow != nPrims || nAbove != 0) std::cout << "you screwed your new idea!\n";
//              Assert(nBelow == nPrims && nAbove == 0);
        }
}

// ============================================================
template<class T>
int kdTree_t<T>::buildTree(u_int32 nPrims, bound_t &nodeBound, u_int32 *primNums,
                u_int32 *leftPrims, u_int32 *rightPrims, boundEdge *edges[3], //working memory
                u_int32 rightMemSize, int depth, int badRefines ) // status
{
//      std::cout << "tree level: " << depth << std::endl;
        if (nextFreeNode == allocatedNodesCount) {
                int newCount = 2*allocatedNodesCount;
                newCount = (newCount > 0x100000) ? allocatedNodesCount+0x80000 : newCount;
                rkdTreeNode<T>  *n = (rkdTreeNode<T> *) y_memalign(64, newCount * sizeof(rkdTreeNode<T>));
                memcpy(n, nodes, allocatedNodesCount * sizeof(rkdTreeNode<T>));
                y_free(nodes);
                nodes = n;
                allocatedNodesCount = newCount;
        }       

#if _TRI_CLIP > 0
        if(nPrims <= TRI_CLIP_THRESH)
        {
//              exBound_t ebound(nodeBound);
                int oPrims[TRI_CLIP_THRESH], nOverl=0;
                double bHalfSize[3];
                double b_ext[2][3];
                for(int i=0; i<3; ++i)
                {
//                      bCenter[i]   = ((double)nodeBound.a[i] + (double)nodeBound.g[i])*0.5;
                        bHalfSize[i] = ((double)nodeBound.g[i] - (double)nodeBound.a[i]);
                        double temp  = ((double)treeBound.g[i] - (double)treeBound.a[i]);
                        b_ext[0][i] = nodeBound.a[i] - 0.021*bHalfSize[i] - 0.00001*temp;
                        b_ext[1][i] = nodeBound.g[i] + 0.021*bHalfSize[i] + 0.00001*temp;
//                      ebound.halfSize[i] *= 1.01;
                }
                char *c_old = cdata + (TRI_CLIP_THRESH * CLIP_DATA_SIZE * depth);
                char *c_new = cdata + (TRI_CLIP_THRESH * CLIP_DATA_SIZE * (depth+1));
                for(unsigned int i=0; i<nPrims; ++i)
                {
                        const T *ct = prims[ primNums[i] ];
                        u_int32 old_idx=0;
                        if(clip[depth] >= 0) old_idx = primNums[i+nPrims];
//                      if(old_idx > TRI_CLIP_THRESH){ std::cout << "ouch!\n"; }
//                      std::cout << "parent idx: " << old_idx << std::endl;
                        if(ct->clippingSupport())
                        {
                                if( ct->clipToBound(b_ext, clip[depth], allBounds[totalPrims+nOverl],
                                        c_old + old_idx*CLIP_DATA_SIZE, c_new + nOverl*CLIP_DATA_SIZE) )
                                {
                                        ++_clip;
                                        oPrims[nOverl] = primNums[i]; nOverl++;
                                }
                                else ++_null_clip;
                        }
                        else
                        {
                                // no clipping supported by prim, copy old bound:
                                allBounds[totalPrims+nOverl] = allBounds[ primNums[i] ]; //really??
                                oPrims[nOverl] = primNums[i]; nOverl++;
                        }
                }
                //copy back
                memcpy(primNums, oPrims, nOverl*sizeof(u_int32));
                nPrims = nOverl;
        }
#endif
        //      << check if leaf criteria met >>
        if(nPrims <= maxLeafSize || depth >= maxDepth)
        {
//              std::cout << "leaf\n";
                nodes[nextFreeNode].createLeaf(primNums, nPrims, prims, primsArena);
                nextFreeNode++;
                if( depth >= maxDepth ) depthLimitReached++; //stat
                return 0;
        }
        
        //<< calculate cost for all axes and chose minimum >>
        splitCost_t split;
        float baseBonus=eBonus;
        eBonus *= 1.1 - (float)depth/(float)maxDepth;
        if(nPrims > 128) pigeonMinCost(nPrims, nodeBound, primNums, split);
#if _TRI_CLIP > 0
        else if (nPrims > TRI_CLIP_THRESH) minimalCost(nPrims, nodeBound, primNums, allBounds, edges, split);
        else minimalCost(nPrims, nodeBound, primNums, allBounds+totalPrims, edges, split);
#else
        else minimalCost(nPrims, nodeBound, primNums, allBounds, edges, split);
#endif
        eBonus=baseBonus; //restore eBonus
        //<< if (minimum > leafcost) increase bad refines >>
        if (split.bestCost > split.oldCost) ++badRefines;
        if ((split.bestCost > 1.6f * split.oldCost && nPrims < 16) ||
                split.bestAxis == -1 || badRefines == 2) {
                nodes[nextFreeNode].createLeaf(primNums, nPrims, prims, primsArena);
                nextFreeNode++;
                if( badRefines == 2) ++NumBadSplits; //stat
                return 0;
        }
        
        //todo: check working memory for child recursive calls
        u_int32 remainingMem, *morePrims = 0, *nRightPrims;
        u_int32 *oldRightPrims = rightPrims;
        if(nPrims > rightMemSize || 2*TRI_CLIP_THRESH > rightMemSize ) // *possibly* not enough, get some more
        {
//              std::cout << "buildTree: more memory allocated!\n";
                remainingMem = nPrims * 3;
                morePrims = new u_int32[remainingMem];
                nRightPrims = morePrims;
        }
        else
        {
                nRightPrims = oldRightPrims;
                remainingMem = rightMemSize;
        }
        
        // Classify primitives with respect to split
        PFLOAT splitPos;
        int n0 = 0, n1 = 0;
        if(nPrims > 128) // we did pigeonhole
        {
                int pn;
                for (unsigned int i=0; i<nPrims; i++)
                {
                        pn = primNums[i];
                        if(allBounds[ pn ].a[split.bestAxis] >= split.t) nRightPrims[n1++] = pn;
                        else
                        {
                                leftPrims[n0++] = pn;
                                if(allBounds[ pn ].g[split.bestAxis] > split.t) nRightPrims[n1++] = pn;
                        }
                }
                splitPos = split.t;
                if (n0!= split.nBelow || n1 != split.nAbove) std::cout << "oops!\n";
/*              static int foo=0;
                if(foo<10)
                {
                        std::cout << "best axis:"<<split.bestAxis<<", rel. split:" <<(split.t - nodeBound.a[split.bestAxis])/(nodeBound.g[split.bestAxis]-nodeBound.a[split.bestAxis]);
                        std::cout << "\nleft Prims:"<<n0<<", right Prims:"<<n1<<", total Prims:"<<nPrims<<" (level:"<<depth<<")\n";
                        foo++;
                }*/
        }
        else if(nPrims <= TRI_CLIP_THRESH)
        {
                int cindizes[TRI_CLIP_THRESH];
                u_int32 oldPrims[TRI_CLIP_THRESH];
//              std::cout << "memcpy\n";
                memcpy(oldPrims, primNums, nPrims*sizeof(int));
                
                for (int i=0; i<split.bestOffset; ++i)
                        if (edges[split.bestAxis][i].end != UPPER_B)
                        {
                                cindizes[n0] = edges[split.bestAxis][i].primNum;
//                              if(cindizes[n0] >nPrims) std::cout << "error!!\n";
                                leftPrims[n0] = oldPrims[cindizes[n0]];
                                ++n0;
                        }
//              std::cout << "append n0\n";
                for(int i=0; i<n0; ++i){ leftPrims[n0+i] = cindizes[i]; /* std::cout << cindizes[i] << " "; */ }
                if (edges[split.bestAxis][split.bestOffset].end == BOTH_B)
                {
                        cindizes[n1] = edges[split.bestAxis][split.bestOffset].primNum;
//                      if(cindizes[n1] >nPrims) std::cout << "error!!\n";
                        nRightPrims[n1] = oldPrims[cindizes[n1]];
                        ++n1;
                }
                for (int i=split.bestOffset+1; i<split.nEdge; ++i)
                        if (edges[split.bestAxis][i].end != LOWER_B)
                        {
                                cindizes[n1] = edges[split.bestAxis][i].primNum;
//                              if(cindizes[n1] >nPrims) std::cout << "error!!\n";
                                nRightPrims[n1] = oldPrims[cindizes[n1]];
                                ++n1;
                        }
//              std::cout << "\nappend n1\n";
                for(int i=0; i<n1; ++i){ nRightPrims[n1+i] = cindizes[i]; /* std::cout << cindizes[i] << " "; */ }
                splitPos = edges[split.bestAxis][split.bestOffset].pos;
//              std::cout << "\ndone\n";
        }
        else //we did "normal" cost function
        {       
                for (int i=0; i<split.bestOffset; ++i)
                        if (edges[split.bestAxis][i].end != UPPER_B)
                                leftPrims[n0++] = edges[split.bestAxis][i].primNum;
                if (edges[split.bestAxis][split.bestOffset].end == BOTH_B)
                        nRightPrims[n1++] = edges[split.bestAxis][split.bestOffset].primNum;
                for (int i=split.bestOffset+1; i<split.nEdge; ++i)
                        if (edges[split.bestAxis][i].end != LOWER_B)
                                nRightPrims[n1++] = edges[split.bestAxis][i].primNum;
                splitPos = edges[split.bestAxis][split.bestOffset].pos;
        }
//      std::cout << "split axis: " << split.bestAxis << ", pos: " << splitPos << "\n";
        //advance right prims pointer
        remainingMem -= n1;
        
        
        u_int32 curNode = nextFreeNode;
        nodes[curNode].createInterior(split.bestAxis, splitPos);
        ++nextFreeNode;
        bound_t boundL = nodeBound, boundR = nodeBound;
        switch(split.bestAxis){
                case 0: boundL.setMaxX(splitPos); boundR.setMinX(splitPos); break;
                case 1: boundL.setMaxY(splitPos); boundR.setMinY(splitPos); break;
                case 2: boundL.setMaxZ(splitPos); boundR.setMinZ(splitPos); break;
        }

#if _TRI_CLIP > 0
        if(nPrims <= TRI_CLIP_THRESH)
        {
                remainingMem -= n1;
                //<< recurse below child >>
                clip[depth+1] = split.bestAxis;
                buildTree(n0, boundL, leftPrims, leftPrims, nRightPrims+2*n1, edges, remainingMem, depth+1, badRefines);
                clip[depth+1] |= 1<<2;
                //<< recurse above child >>
                nodes[curNode].setRightChild (nextFreeNode);
                buildTree(n1, boundR, nRightPrims, leftPrims, nRightPrims+2*n1, edges, remainingMem, depth+1, badRefines);
                clip[depth+1] = -1;
        }
        else
        {
#endif
                //<< recurse below child >>
                buildTree(n0, boundL, leftPrims, leftPrims, nRightPrims+n1, edges, remainingMem, depth+1, badRefines);
                //<< recurse above child >>
                nodes[curNode].setRightChild (nextFreeNode);
                buildTree(n1, boundR, nRightPrims, leftPrims, nRightPrims+n1, edges, remainingMem, depth+1, badRefines);
#if _TRI_CLIP > 0
        }
#endif
        // free additional working memory, if present
        if(morePrims) delete[] morePrims;
        return 1;
}
        


//============================
template<class T>
bool kdTree_t<T>::Intersect(const ray_t &ray, PFLOAT dist, T **tr, PFLOAT &Z, void *udat) const
{
        Z=dist;
//      std::cout << "kdTree_t::Intersect: Z="<<Z<<"\n";
        PFLOAT a, b, t; // entry/exit/splitting plane signed distance
        PFLOAT t_hit;
        
        if (!treeBound.cross(ray, a, b, dist))
        { return false; }
        
        unsigned char udat1[PRIM_DAT_SIZE], udat2[PRIM_DAT_SIZE];
        void *c_udat=(void*)&udat1[0], *t_udat=(void*)&udat2[0];
        vector3d_t invDir(1.0/ray.dir.x, 1.0/ray.dir.y, 1.0/ray.dir.z); //was 1.f!
//      int rayId = curMailboxId++;
        bool hit = false;
        
        rKdStack<T> stack[KD_MAX_STACK];
        const rkdTreeNode<T> *farChild, *currNode;
        currNode = nodes;
        
        int enPt = 0;
        stack[enPt].t = a;
        
        //distinguish between internal and external origin
        if(a >= 0.0) // ray with external origin
                stack[enPt].pb = ray.from + ray.dir * a;
        else // ray with internal origin
                stack[enPt].pb = ray.from;
        
        // setup initial entry and exit poimt in stack
        int exPt = 1; // pointer to stack
        stack[exPt].t = b;
        stack[exPt].pb = ray.from + ray.dir * b;
        stack[exPt].node = 0; // "nowhere", termination flag
        
        //loop, traverse kd-Tree until object intersection or ray leaves tree bound
        while (currNode != NULL)
        {
                if (dist < stack[enPt].t) break;
                // loop until leaf is found
                while( !currNode->IsLeaf() )
                {
                        int axis = currNode->SplitAxis();
                        PFLOAT splitVal = currNode->SplitPos();
                        
                        if(stack[enPt].pb[axis] <= splitVal){
                                if(stack[exPt].pb[axis] <= splitVal)
                                {
                                        currNode++;
                                        continue;
                                }
                                if(stack[exPt].pb[axis] == splitVal)
                                {
                                        currNode = &nodes[currNode->getRightChild()];
                                        continue;
                                }
                                // case N4
                                farChild = &nodes[currNode->getRightChild()];
                                currNode ++;
                        }
                        else
                        {
                                if(splitVal < stack[exPt].pb[axis])
                                {
                                        currNode = &nodes[currNode->getRightChild()];
                                        continue;
                                }
                                farChild = currNode+1;
                                currNode = &nodes[currNode->getRightChild()];
                        }
                        // traverse both children
                        
                        t = (splitVal - ray.from[axis]) * invDir[axis];
                        
                        // setup the new exit point
                        int tmp = exPt;
                        exPt++;
                        
                        // possibly skip current entry point so not to overwrite the data
                        if (exPt == enPt) exPt++;
                        // push values onto the stack //todo: lookup table
                        static const int npAxis[2][3] = { {1, 2, 0}, {2, 0, 1} };
                        int nextAxis = npAxis[0][axis];//(axis+1)%3;
                        int prevAxis = npAxis[1][axis];//(axis+2)%3;
                        stack[exPt].prev = tmp;
                        stack[exPt].t = t;
                        stack[exPt].node = farChild;
                        stack[exPt].pb[axis] = splitVal;
                        stack[exPt].pb[nextAxis] = ray.from[nextAxis] + t * ray.dir[nextAxis];
                        stack[exPt].pb[prevAxis] = ray.from[prevAxis] + t * ray.dir[prevAxis];
                }
                                 
                // Check for intersections inside leaf node
                u_int32 nPrimitives = currNode->nPrimitives();
                if (nPrimitives == 1) {
                        T *mp = currNode->onePrimitive;
//                      if (mp->lastMailboxId != rayId) {
//                              mp->lastMailboxId = rayId;
                                if (mp->intersect(ray, &t_hit, t_udat))
                                {
//                                      std::cout<<"kdTree_t: hit! t="<<t_hit<<" Z="<<Z<<"\n";
                                        if(t_hit < Z && t_hit >= ray.tmin /*0.f*/ /*stack[enPt].t*/)
                                        {
                                                Z = t_hit;
                                                *tr = mp;
                                                std::swap(t_udat, c_udat);
                                                hit = true;
                                        }
                                }
//                      }
                }
                else {
                        T **prims = currNode->primitives;
                        for (u_int32 i = 0; i < nPrimitives; ++i) {
                                T *mp = prims[i];
//                              if (mp->lastMailboxId != rayId) {
//                                      mp->lastMailboxId = rayId;
                                        if (mp->intersect(ray, &t_hit, t_udat))
                                        {
//                                              std::cout<<"kdTree_t: hit! t="<<t_hit<<" Z="<<Z<<"\n";
                                                if(t_hit < Z && t_hit >= ray.tmin /*0.f*/ /*stack[enPt].t*/)
                                                {
                                                        Z = t_hit;
                                                        *tr = mp;
                                                        std::swap(t_udat, c_udat);
                                                        hit = true;
                                                }
                                        }
//                              }
                        }
                }
                
                if(hit && Z <= stack[exPt].t){ memcpy(udat, c_udat, PRIM_DAT_SIZE); return true;}
                
                enPt = exPt;
                currNode = stack[exPt].node;
                exPt = stack[enPt].prev;
                                
        } // while
//      if(hit) return true;
        memcpy(udat, c_udat, PRIM_DAT_SIZE);
        return hit; //false;
}

template<class T>
bool kdTree_t<T>::IntersectS(const ray_t &ray, PFLOAT dist, T **tr) const
{
        PFLOAT a, b, t; // entry/exit/splitting plane signed distance
        PFLOAT t_hit;
        
        if (!treeBound.cross(ray, a, b, dist))
                return false;
        
        unsigned char udat[PRIM_DAT_SIZE];
        vector3d_t invDir(1.f/ray.dir.x, 1.f/ray.dir.y, 1.f/ray.dir.z);
//      int rayId = curMailboxId++;
//      bool hit = false;
        
        rKdStack<T> stack[KD_MAX_STACK];
        const rkdTreeNode<T> *farChild, *currNode;
        currNode = nodes;
        
        int enPt = 0;
        stack[enPt].t = a;
        
        //distinguish between internal and external origin
        if(a >= 0.0) // ray with external origin
                stack[enPt].pb = ray.from + ray.dir * a;
        else // ray with internal origin
                stack[enPt].pb = ray.from;
        
        // setup initial entry and exit poimt in stack
        int exPt = 1; // pointer to stack
        stack[exPt].t = b;
        stack[exPt].pb = ray.from + ray.dir * b;
        stack[exPt].node = 0; // "nowhere", termination flag
        
        //loop, traverse kd-Tree until object intersection or ray leaves tree bound
        while (currNode != NULL)
        {
                if (dist < stack[enPt].t /*a*/) break;
                // loop until leaf is found
                while( !currNode->IsLeaf() )
                {
                        int axis = currNode->SplitAxis();
                        PFLOAT splitVal = currNode->SplitPos();
                        
                        if(stack[enPt].pb[axis] <= splitVal){
                                if(stack[exPt].pb[axis] <= splitVal)
                                {
                                        currNode++;
                                        continue;
                                }
                                if(stack[exPt].pb[axis] == splitVal)
                                {
                                        currNode = &nodes[currNode->getRightChild()];
                                        continue;
                                }
                                // case N4
                                farChild = &nodes[currNode->getRightChild()];
                                currNode ++;
                        }
                        else
                        {
                                if(splitVal < stack[exPt].pb[axis])
                                {
                                        currNode = &nodes[currNode->getRightChild()];
                                        continue;
                                }
                                farChild = currNode+1;
                                currNode = &nodes[currNode->getRightChild()];
                        }
                        // traverse both children
                        
                        t = (splitVal - ray.from[axis]) * invDir[axis];
                        
                        // setup the new exit point
                        int tmp = exPt;
                        exPt++;
                        
                        // possibly skip current entry point so not to overwrite the data
                        if (exPt == enPt) exPt++;
                        // push values onto the stack //todo: lookup table
                        static const int npAxis[2][3] = { {1, 2, 0}, {2, 0, 1} };
                        int nextAxis = npAxis[0][axis];//(axis+1)%3;
                        int prevAxis = npAxis[1][axis];//(axis+2)%3;
                        stack[exPt].prev = tmp;
                        stack[exPt].t = t;
                        stack[exPt].node = farChild;
                        stack[exPt].pb[axis] = splitVal;
                        stack[exPt].pb[nextAxis] = ray.from[nextAxis] + t * ray.dir[nextAxis];
                        stack[exPt].pb[prevAxis] = ray.from[prevAxis] + t * ray.dir[prevAxis];
                }
                                 
                // Check for intersections inside leaf node
                u_int32 nPrimitives = currNode->nPrimitives();
                if (nPrimitives == 1) {
                        T *mp = currNode->onePrimitive;
//                      if (mp->lastMailboxId != rayId) {
//                              mp->lastMailboxId = rayId;
                                if (mp->intersect(ray, &t_hit, (void*)&udat[0]))
                                {
//                                      hit = true;
                                        if(t_hit < dist && t_hit > 0.f ) // '>=' ?
                                        {
                                                *tr = mp;
                                                return true;
                                        }
                                }
//                      }
                }
                else {
                        T **prims = currNode->primitives;
                        for (u_int32 i = 0; i < nPrimitives; ++i) {
                                T *mp = prims[i];
//                              if (mp->lastMailboxId != rayId) {
//                                      mp->lastMailboxId = rayId;
                                        if (mp->intersect(ray, &t_hit, (void*)&udat[0]))
                                        {
                                                if(t_hit < dist && t_hit > 0.f )
                                                {
//                                                      hit = true;
                                                        *tr = mp;
                                                        return true;
                                                }
                                        }
//                              }
                        }
                }
                
//              if(hit && dist <= stack[exPt].t) return true;
                
                enPt = exPt;
                currNode = stack[exPt].node;
                exPt = stack[enPt].prev;
                                
        } // while
//      if(hit) return true;
        return false;
}

/*=============================================================
        allow for transparent shadows.
=============================================================*/

template<class T>
bool kdTree_t<T>::IntersectTS(renderState_t &state, const ray_t &ray, int maxDepth, PFLOAT dist, T **tr, color_t &filt) const
{
        PFLOAT a, b, t; // entry/exit/splitting plane signed distance
        PFLOAT t_hit;
        
        if (!treeBound.cross(ray, a, b, dist))
                return false;
        
        /* unsigned char */double udat[PRIM_DAT_SIZE];
        vector3d_t invDir(1.f/ray.dir.x, 1.f/ray.dir.y, 1.f/ray.dir.z);
//      int rayId = curMailboxId++;
//      bool hit = false;
        int depth=0;
//      filt = color_t(1.0);
#if ( HAVE_PTHREAD && defined (__GNUC__) )
        std::set<const T *, std::less<const T *>, __gnu_cxx::__mt_alloc<const T *> > filtered;
#else
        std::set<const T *> filtered;
#endif
        rKdStack<T> stack[KD_MAX_STACK];
        const rkdTreeNode<T> *farChild, *currNode;
        currNode = nodes;
        
        int enPt = 0;
        stack[enPt].t = a;
        
        //distinguish between internal and external origin
        if(a >= 0.0) // ray with external origin
                stack[enPt].pb = ray.from + ray.dir * a;
        else // ray with internal origin
                stack[enPt].pb = ray.from;
        
        // setup initial entry and exit poimt in stack
        int exPt = 1; // pointer to stack
        stack[exPt].t = b;
        stack[exPt].pb = ray.from + ray.dir * b;
        stack[exPt].node = 0; // "nowhere", termination flag
        
        //loop, traverse kd-Tree until object intersection or ray leaves tree bound
        while (currNode != NULL)
        {
                if (dist < stack[enPt].t /*a*/) break;
                // loop until leaf is found
                while( !currNode->IsLeaf() )
                {
                        int axis = currNode->SplitAxis();
                        PFLOAT splitVal = currNode->SplitPos();
                        
                        if(stack[enPt].pb[axis] <= splitVal){
                                if(stack[exPt].pb[axis] <= splitVal)
                                {
                                        currNode++;
                                        continue;
                                }
                                if(stack[exPt].pb[axis] == splitVal)
                                {
                                        currNode = &nodes[currNode->getRightChild()];
                                        continue;
                                }
                                // case N4
                                farChild = &nodes[currNode->getRightChild()];
                                currNode ++;
                        }
                        else
                        {
                                if(splitVal < stack[exPt].pb[axis])
                                {
                                        currNode = &nodes[currNode->getRightChild()];
                                        continue;
                                }
                                farChild = currNode+1;
                                currNode = &nodes[currNode->getRightChild()];
                        }
                        // traverse both children
                        
                        t = (splitVal - ray.from[axis]) * invDir[axis];
                        
                        // setup the new exit point
                        int tmp = exPt;
                        exPt++;
                        
                        // possibly skip current entry point so not to overwrite the data
                        if (exPt == enPt) exPt++;
                        // push values onto the stack //todo: lookup table
                        static const int npAxis[2][3] = { {1, 2, 0}, {2, 0, 1} };
                        int nextAxis = npAxis[0][axis];//(axis+1)%3;
                        int prevAxis = npAxis[1][axis];//(axis+2)%3;
                        stack[exPt].prev = tmp;
                        stack[exPt].t = t;
                        stack[exPt].node = farChild;
                        stack[exPt].pb[axis] = splitVal;
                        stack[exPt].pb[nextAxis] = ray.from[nextAxis] + t * ray.dir[nextAxis];
                        stack[exPt].pb[prevAxis] = ray.from[prevAxis] + t * ray.dir[prevAxis];
                }
                                 
                // Check for intersections inside leaf node
                u_int32 nPrimitives = currNode->nPrimitives();
//              PFLOAT tmax = (stack[exPt].t < dist) ? stack[exPt].t : dist;
//              PFLOAT tmin = (stack[enPt].t > ray.tmin) ? stack[enPt].t : ray.tmin;
                if (nPrimitives == 1) {
                        T *mp = currNode->onePrimitive;
                        if (mp->intersect(ray, &t_hit, (void*)&udat[0]))
                        {
                                if(t_hit < /* tmax */ dist && t_hit >= /* tmin */ ray.tmin ) // '>=' ?
                                {
//                                      *tr = mp;
                                        const material_t *mat = mp->getMaterial();
                                        if(!mat->isTransparent() ) return true;
                                        if(filtered.insert(mp).second)
                                        {
                                                if(depth>=maxDepth) return true;
                                                point3d_t h=ray.from + t_hit*ray.dir;
                                                surfacePoint_t sp;
                                                mp->getSurface(sp, h, (void*)&udat[0]);
                                                filt *= mat->getTransparency(state, sp, ray.dir);
                                                ++depth;
                                        }
                                }
                        }
                }
                else {
                        T **prims = currNode->primitives;
                        for (u_int32 i = 0; i < nPrimitives; ++i) {
                                T *mp = prims[i];
                                if (mp->intersect(ray, &t_hit, (void*)&udat[0]))
                                {
                                        if(t_hit < /* tmax */ dist && t_hit >= /* tmin */ ray.tmin )
                                        {
//                                              *tr = mp;
                                                const material_t *mat = mp->getMaterial();
                                                if(!mat->isTransparent() ) return true;
                                                if(filtered.insert(mp).second)
                                                {
                                                        if(depth>=maxDepth) return true;
                                                        point3d_t h=ray.from + t_hit*ray.dir;
                                                        surfacePoint_t sp;
                                                        mp->getSurface(sp, h, (void*)&udat[0]);
                                                        filt *= mat->getTransparency(state, sp, ray.dir);
                                                        ++depth;
                                                }
                                        }
                                }
                        }
                }
                
//              if(hit && dist <= stack[exPt].t) return true;
                
                enPt = exPt;
                currNode = stack[exPt].node;
                exPt = stack[enPt].prev;
                                
        } // while
//      if(hit) return true;
        return false;
}

// explicit instantiation of template:
template class kdTree_t<triangle_t>;
template class kdTree_t<primitive_t>;

__END_YAFRAY

---