shadbuf.c

Go to the documentation of this file.

/*
 * ***** BEGIN GPL LICENSE BLOCK *****
 *
 * This program is free software; you can redistribute it and/or
 * modify it under the terms of the GNU General Public License
 * as published by the Free Software Foundation; either version 2
 * of the License, or (at your option) any later version.
 *
 * This program is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 * GNU General Public License for more details.
 *
 * You should have received a copy of the GNU General Public License
 * along with this program; if not, write to the Free Software Foundation,
 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA.
 *
 * The Original Code is Copyright (C) 2001-2002 by NaN Holding BV.
 * All rights reserved.
 *
 * Contributor(s): 2004-2006, Blender Foundation
 *
 * ***** END GPL LICENSE BLOCK *****
 */

#include <math.h>
#include <string.h>


#include "MEM_guardedalloc.h"

#include "DNA_group_types.h"
#include "DNA_lamp_types.h"
#include "DNA_material_types.h"

#include "BKE_global.h"
#include "BKE_scene.h"


#include "BLI_math.h"
#include "BLI_blenlib.h"
#include "BLI_jitter.h"
#include "BLI_memarena.h"
#include "BLI_rand.h"
#include "BLI_utildefines.h"

#include "PIL_time.h"

#include "renderpipeline.h"
#include "render_types.h"
#include "renderdatabase.h"
#include "rendercore.h"
#include "shadbuf.h"
#include "shading.h"
#include "zbuf.h"

/* XXX, could be better implemented... this is for endian issues
*/
#ifdef __BIG_ENDIAN__
#  define RCOMP 3
#  define GCOMP 2
#  define BCOMP 1
#  define ACOMP 0
#else
#  define RCOMP 0
#  define GCOMP 1
#  define BCOMP 2
#  define ACOMP 3
#endif

/* ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ */
/* defined in pipeline.c, is hardcopy of active dynamic allocated Render */
/* only to be used here in this file, it's for speed */
extern struct Render R;
/* ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ */

/* ------------------------------------------------------------------------- */

/* initshadowbuf() in convertBlenderScene.c */

/* ------------------------------------------------------------------------- */

static void copy_to_ztile(int *rectz, int size, int x1, int y1, int tile, char *r1)
{
    int len4, *rz;  
    int x2, y2;
    
    x2= x1+tile;
    y2= y1+tile;
    if(x2>=size) x2= size-1;
    if(y2>=size) y2= size-1;

    if(x1>=x2 || y1>=y2) return;

    len4= 4*(x2- x1);
    rz= rectz + size*y1 + x1;
    for(; y1<y2; y1++) {
        memcpy(r1, rz, len4);
        rz+= size;
        r1+= len4;
    }
}

#if 0
static int sizeoflampbuf(ShadBuf *shb)
{
    int num,count=0;
    char *cp;
    
    cp= shb->cbuf;
    num= (shb->size*shb->size)/256;

    while(num--) count+= *(cp++);
    
    return 256*count;
}
#endif

/* not threadsafe... */
static float *give_jitter_tab(int samp)
{
    /* these are all possible jitter tables, takes up some
     * 12k, not really bad!
     * For soft shadows, it saves memory and render time
     */
    static int tab[17]={1, 4, 9, 16, 25, 36, 49, 64, 81, 100, 121, 144, 169, 196, 225, 256};
    static float jit[1496][2];
    static char ctab[17]= {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0};
    int a, offset=0;
    
    if(samp<2) samp= 2;
    else if(samp>16) samp= 16;

    for(a=0; a<samp-1; a++) offset+= tab[a];

    if(ctab[samp]==0) {
        ctab[samp]= 1;
        BLI_initjit(jit[offset], samp*samp);
    }
        
    return jit[offset];
    
}

static void make_jitter_weight_tab(Render *re, ShadBuf *shb, short filtertype) 
{
    float *jit, totw= 0.0f;
    int samp= get_render_shadow_samples(&re->r, shb->samp);
    int a, tot=samp*samp;
    
    shb->weight= MEM_mallocN(sizeof(float)*tot, "weight tab lamp");
    
    for(jit= shb->jit, a=0; a<tot; a++, jit+=2) {
        if(filtertype==LA_SHADBUF_TENT) 
            shb->weight[a]= 0.71f - sqrt(jit[0]*jit[0] + jit[1]*jit[1]);
        else if(filtertype==LA_SHADBUF_GAUSS) 
            shb->weight[a]= RE_filter_value(R_FILTER_GAUSS, 1.8f*sqrt(jit[0]*jit[0] + jit[1]*jit[1]));
        else
            shb->weight[a]= 1.0f;
        
        totw+= shb->weight[a];
    }
    
    totw= 1.0f/totw;
    for(a=0; a<tot; a++) {
        shb->weight[a]*= totw;
    }
}

static int verg_deepsample(const void *poin1, const void *poin2)
{
    const DeepSample *ds1= (const DeepSample*)poin1;
    const DeepSample *ds2= (const DeepSample*)poin2;

    if(ds1->z < ds2->z) return -1;
    else if(ds1->z == ds2->z) return 0;
    else return 1;
}

static int compress_deepsamples(DeepSample *dsample, int tot, float epsilon)
{
    /* uses doubles to avoid overflows and other numerical issues,
       could be improved */
    DeepSample *ds, *newds;
    float v;
    double slope, slopemin, slopemax, min, max, div, newmin, newmax;
    int a, first, z, newtot= 0;

    /*if(print) {
        for(a=0, ds=dsample; a<tot; a++, ds++)
            printf("%lf,%f ", ds->z/(double)0x7FFFFFFF, ds->v);
        printf("\n");
    }*/

    /* read from and write into same array */
    ds= dsample;
    newds= dsample;
    a= 0;

    /* as long as we are not at the end of the array */
    for(a++, ds++; a<tot; a++, ds++) {
        slopemin= 0.0f;
        slopemax= 0.0f;
        first= 1;

        for(; a<tot; a++, ds++) {
            //dz= ds->z - newds->z;
            if(ds->z == newds->z) {
                /* still in same z position, simply check
                   visibility difference against epsilon */
                if(!(fabs(newds->v - ds->v) <= epsilon)) {
                    break;
                }
            }
            else {
                /* compute slopes */
                div= (double)0x7FFFFFFF/((double)ds->z - (double)newds->z);
                min= ((ds->v - epsilon) - newds->v)*div;
                max= ((ds->v + epsilon) - newds->v)*div;

                /* adapt existing slopes */
                if(first) {
                    newmin= min;
                    newmax= max;
                    first= 0;
                }
                else {
                    newmin= MAX2(slopemin, min);
                    newmax= MIN2(slopemax, max);

                    /* verify if there is still space between the slopes */
                    if(newmin > newmax) {
                        ds--;
                        a--;
                        break;
                    }
                }

                slopemin= newmin;
                slopemax= newmax;
            }
        }

        if(a == tot) {
            ds--;
            a--;
        }

        /* always previous z */
        z= ds->z;

        if(first || a==tot-1) {
            /* if slopes were not initialized, use last visibility */
            v= ds->v;
        }
        else {
            /* compute visibility at center between slopes at z */
            slope= (slopemin+slopemax)*0.5f;
            v= newds->v + slope*((z - newds->z)/(double)0x7FFFFFFF);
        }

        newds++;
        newtot++;

        newds->z= z;
        newds->v= v;
    }

    if(newtot == 0 || (newds->v != (newds-1)->v))
        newtot++;

    /*if(print) {
        for(a=0, ds=dsample; a<newtot; a++, ds++)
            printf("%lf,%f ", ds->z/(double)0x7FFFFFFF, ds->v);
        printf("\n");
    }*/

    return newtot;
}

static float deep_alpha(Render *re, int obinr, int facenr, int strand)
{
    ObjectInstanceRen *obi= &re->objectinstance[obinr];
    Material *ma;

    if(strand) {
        StrandRen *strand= RE_findOrAddStrand(obi->obr, facenr-1);
        ma= strand->buffer->ma;
    }
    else {
        VlakRen *vlr= RE_findOrAddVlak(obi->obr, (facenr-1) & RE_QUAD_MASK);
        ma= vlr->mat;
    }

    return ma->shad_alpha;
}

static void compress_deepshadowbuf(Render *re, ShadBuf *shb, APixstr *apixbuf, APixstrand *apixbufstrand)
{
    ShadSampleBuf *shsample;
    DeepSample *ds[RE_MAX_OSA], *sampleds[RE_MAX_OSA], *dsb, *newbuf;
    APixstr *ap, *apn;
    APixstrand *aps, *apns;
    float visibility;

    const int totbuf= shb->totbuf;
    const float totbuf_f= (float)shb->totbuf;
    const float totbuf_f_inv= 1.0f/totbuf_f;
    const int size= shb->size;

    int a, b, c, tot, minz, found, prevtot, newtot;
    int sampletot[RE_MAX_OSA], totsample = 0, totsamplec = 0;
    
    shsample= MEM_callocN( sizeof(ShadSampleBuf), "shad sample buf");
    BLI_addtail(&shb->buffers, shsample);

    shsample->totbuf= MEM_callocN(sizeof(int)*size*size, "deeptotbuf");
    shsample->deepbuf= MEM_callocN(sizeof(DeepSample*)*size*size, "deepbuf");

    ap= apixbuf;
    aps= apixbufstrand;
    for(a=0; a<size*size; a++, ap++, aps++) {
        /* count number of samples */
        for(c=0; c<totbuf; c++)
            sampletot[c]= 0;

        tot= 0;
        for(apn=ap; apn; apn=apn->next)
            for(b=0; b<4; b++)
                if(apn->p[b])
                    for(c=0; c<totbuf; c++)
                        if(apn->mask[b] & (1<<c))
                            sampletot[c]++;

        if(apixbufstrand) {
            for(apns=aps; apns; apns=apns->next)
                for(b=0; b<4; b++)
                    if(apns->p[b])
                        for(c=0; c<totbuf; c++)
                            if(apns->mask[b] & (1<<c))
                                sampletot[c]++;
        }

        for(c=0; c<totbuf; c++)
            tot += sampletot[c];

        if(tot == 0) {
            shsample->deepbuf[a]= NULL;
            shsample->totbuf[a]= 0;
            continue;
        }

        /* fill samples */
        ds[0]= sampleds[0]= MEM_callocN(sizeof(DeepSample)*tot*2, "deepsample");
        for(c=1; c<totbuf; c++)
            ds[c]= sampleds[c]= sampleds[c-1] + sampletot[c-1]*2;

        for(apn=ap; apn; apn=apn->next) {
            for(b=0; b<4; b++) {
                if(apn->p[b]) {
                    for(c=0; c<totbuf; c++) {
                        if(apn->mask[b] & (1<<c)) {
                            /* two entries to create step profile */
                            ds[c]->z= apn->z[b];
                            ds[c]->v= 1.0f; /* not used */
                            ds[c]++;
                            ds[c]->z= apn->z[b];
                            ds[c]->v= deep_alpha(re, apn->obi[b], apn->p[b], 0);
                            ds[c]++;
                        }
                    }
                }
            }
        }

        if(apixbufstrand) {
            for(apns=aps; apns; apns=apns->next) {
                for(b=0; b<4; b++) {
                    if(apns->p[b]) {
                        for(c=0; c<totbuf; c++) {
                            if(apns->mask[b] & (1<<c)) {
                                /* two entries to create step profile */
                                ds[c]->z= apns->z[b];
                                ds[c]->v= 1.0f; /* not used */
                                ds[c]++;
                                ds[c]->z= apns->z[b];
                                ds[c]->v= deep_alpha(re, apns->obi[b], apns->p[b], 1);
                                ds[c]++;
                            }
                        }
                    }
                }
            }
        }

        for(c=0; c<totbuf; c++) {
            /* sort by increasing z */
            qsort(sampleds[c], sampletot[c], sizeof(DeepSample)*2, verg_deepsample);

            /* sum visibility, replacing alpha values */
            visibility= 1.0f;
            ds[c]= sampleds[c];

            for(b=0; b<sampletot[c]; b++) {
                /* two entries creating step profile */
                ds[c]->v= visibility;
                ds[c]++;

                visibility *= 1.0f-ds[c]->v;
                ds[c]->v= visibility;
                ds[c]++;
            }

            /* halfway trick, probably won't work well for volumes? */
            ds[c]= sampleds[c];
            for(b=0; b<sampletot[c]; b++) {
                if(b+1 < sampletot[c]) {
                    ds[c]->z= (ds[c]->z>>1) + ((ds[c]+2)->z>>1);
                    ds[c]++;
                    ds[c]->z= (ds[c]->z>>1) + ((ds[c]+2)->z>>1);
                    ds[c]++;
                }
                else {
                    ds[c]->z= (ds[c]->z>>1) + (0x7FFFFFFF>>1);
                    ds[c]++;
                    ds[c]->z= (ds[c]->z>>1) + (0x7FFFFFFF>>1);
                    ds[c]++;
                }
            }

            /* init for merge loop */
            ds[c]= sampleds[c];
            sampletot[c] *= 2;
        }

        shsample->deepbuf[a]= MEM_callocN(sizeof(DeepSample)*tot*2, "deepsample");
        shsample->totbuf[a]= 0;

        /* merge buffers */
        dsb= shsample->deepbuf[a];
        while(1) {
            minz= 0;
            found= 0;

            for(c=0; c<totbuf; c++) {
                if(sampletot[c] && (!found || ds[c]->z < minz)) {
                    minz= ds[c]->z;
                    found= 1;
                }
            }

            if(!found)
                break;

            dsb->z= minz;
            dsb->v= 0.0f;

            visibility= 0.0f;
            for(c=0; c<totbuf; c++) {
                if(sampletot[c] && ds[c]->z == minz) {
                    ds[c]++;
                    sampletot[c]--;
                }

                if(sampleds[c] == ds[c])
                    visibility += totbuf_f_inv;
                else
                    visibility += (ds[c]-1)->v / totbuf_f;
            }

            dsb->v= visibility;
            dsb++;
            shsample->totbuf[a]++;
        }

        prevtot= shsample->totbuf[a];
        totsample += prevtot;

        newtot= compress_deepsamples(shsample->deepbuf[a], prevtot, shb->compressthresh);
        shsample->totbuf[a]= newtot;
        totsamplec += newtot;

        if(newtot < prevtot) {
            newbuf= MEM_mallocN(sizeof(DeepSample)*newtot, "cdeepsample");
            memcpy(newbuf, shsample->deepbuf[a], sizeof(DeepSample)*newtot);
            MEM_freeN(shsample->deepbuf[a]);
            shsample->deepbuf[a]= newbuf;
        }

        MEM_freeN(sampleds[0]);
    }

    //printf("%d -> %d, ratio %f\n", totsample, totsamplec, (float)totsamplec/(float)totsample);
}

/* create Z tiles (for compression): this system is 24 bits!!! */
static void compress_shadowbuf(ShadBuf *shb, int *rectz, int square)
{
    ShadSampleBuf *shsample;
    float dist;
    uintptr_t *ztile;
    int *rz, *rz1, verg, verg1, size= shb->size;
    int a, x, y, minx, miny, byt1, byt2;
    char *rc, *rcline, *ctile, *zt;
    
    shsample= MEM_callocN( sizeof(ShadSampleBuf), "shad sample buf");
    BLI_addtail(&shb->buffers, shsample);
    
    shsample->zbuf= MEM_mallocN( sizeof(uintptr_t)*(size*size)/256, "initshadbuf2");
    shsample->cbuf= MEM_callocN( (size*size)/256, "initshadbuf3");
    
    ztile= (uintptr_t *)shsample->zbuf;
    ctile= shsample->cbuf;
    
    /* help buffer */
    rcline= MEM_mallocN(256*4+sizeof(int), "makeshadbuf2");
    
    for(y=0; y<size; y+=16) {
        if(y< size/2) miny= y+15-size/2;
        else miny= y-size/2;    
        
        for(x=0; x<size; x+=16) {
            
            /* is tile within spotbundle? */
            a= size/2;
            if(x< a) minx= x+15-a;
            else minx= x-a; 
            
            dist= sqrt( (float)(minx*minx+miny*miny) );
            
            if(square==0 && dist>(float)(a+12)) {   /* 12, tested with a onlyshadow lamp */
                a= 256; verg= 0; /* 0x80000000; */ /* 0x7FFFFFFF; */
                rz1= (&verg)+1;
            } 
            else {
                copy_to_ztile(rectz, size, x, y, 16, rcline);
                rz1= (int *)rcline;
                
                verg= (*rz1 & 0xFFFFFF00);
                
                for(a=0;a<256;a++,rz1++) {
                    if( (*rz1 & 0xFFFFFF00) !=verg) break;
                }
            }
            if(a==256) { /* complete empty tile */
                *ctile= 0;
                *ztile= *(rz1-1);
            }
            else {
                
                /* ACOMP etc. are defined to work L/B endian */
                
                rc= rcline;
                rz1= (int *)rcline;
                verg=  rc[ACOMP];
                verg1= rc[BCOMP];
                rc+= 4;
                byt1= 1; byt2= 1;
                for(a=1;a<256;a++,rc+=4) {
                    byt1 &= (verg==rc[ACOMP]);
                    byt2 &= (verg1==rc[BCOMP]);
                    
                    if(byt1==0) break;
                }
                if(byt1 && byt2) {  /* only store byte */
                    *ctile= 1;
                    *ztile= (uintptr_t)MEM_mallocN(256+4, "tile1");
                    rz= (int *)*ztile;
                    *rz= *rz1;
                    
                    zt= (char *)(rz+1);
                    rc= rcline;
                    for(a=0; a<256; a++, zt++, rc+=4) *zt= rc[GCOMP];   
                }
                else if(byt1) {     /* only store short */
                    *ctile= 2;
                    *ztile= (uintptr_t)MEM_mallocN(2*256+4,"Tile2");
                    rz= (int *)*ztile;
                    *rz= *rz1;
                    
                    zt= (char *)(rz+1);
                    rc= rcline;
                    for(a=0; a<256; a++, zt+=2, rc+=4) {
                        zt[0]= rc[BCOMP];
                        zt[1]= rc[GCOMP];
                    }
                }
                else {          /* store triple */
                    *ctile= 3;
                    *ztile= (uintptr_t)MEM_mallocN(3*256,"Tile3");

                    zt= (char *)*ztile;
                    rc= rcline;
                    for(a=0; a<256; a++, zt+=3, rc+=4) {
                        zt[0]= rc[ACOMP];
                        zt[1]= rc[BCOMP];
                        zt[2]= rc[GCOMP];
                    }
                }
            }
            ztile++;
            ctile++;
        }
    }

    MEM_freeN(rcline);
}

/* sets start/end clipping. lar->shb should be initialized */
static void shadowbuf_autoclip(Render *re, LampRen *lar)
{
    ObjectInstanceRen *obi;
    ObjectRen *obr;
    VlakRen *vlr= NULL;
    VertRen *ver= NULL;
    Material *ma= NULL;
    float minz, maxz, vec[3], viewmat[4][4], obviewmat[4][4];
    unsigned int lay = -1;
    int i, a, maxtotvert, ok= 1;
    char *clipflag;
    
    minz= 1.0e30f; maxz= -1.0e30f;
    copy_m4_m4(viewmat, lar->shb->viewmat);
    
    if(lar->mode & (LA_LAYER|LA_LAYER_SHADOW)) lay= lar->lay;

    maxtotvert= 0;
    for(obr=re->objecttable.first; obr; obr=obr->next)
        maxtotvert= MAX2(obr->totvert, maxtotvert);

    clipflag= MEM_callocN(sizeof(char)*maxtotvert, "autoclipflag");

    /* set clip in vertices when face visible */
    for(i=0, obi=re->instancetable.first; obi; i++, obi=obi->next) {
        obr= obi->obr;

        if(obi->flag & R_TRANSFORMED)
            mult_m4_m4m4(obviewmat, viewmat, obi->mat);
        else
            copy_m4_m4(obviewmat, viewmat);

        memset(clipflag, 0, sizeof(char)*obr->totvert);

        /* clear clip, is being set if face is visible (clip is calculated for real later) */
        for(a=0; a<obr->totvlak; a++) {
            if((a & 255)==0) vlr= obr->vlaknodes[a>>8].vlak;
            else vlr++;
            
            /* note; these conditions are copied from zbuffer_shadow() */
            if(vlr->mat!= ma) {
                ma= vlr->mat;
                ok= 1;
                if((ma->mode & MA_SHADBUF)==0) ok= 0;
            }
            
            if(ok && (obi->lay & lay)) {
                clipflag[vlr->v1->index]= 1;
                clipflag[vlr->v2->index]= 1;
                clipflag[vlr->v3->index]= 1;
                if(vlr->v4) clipflag[vlr->v4->index]= 1;
            }               
        }       
        
        /* calculate min and max */
        for(a=0; a< obr->totvert;a++) {
            if((a & 255)==0) ver= RE_findOrAddVert(obr, a);
            else ver++;
            
            if(clipflag[a]) {
                copy_v3_v3(vec, ver->co);
                mul_m4_v3(obviewmat, vec);
                /* Z on visible side of lamp space */
                if(vec[2] < 0.0f) {
                    float inpr, z= -vec[2];
                    
                    /* since vec is rotated in lampspace, this is how to get the cosine of angle */
                    /* precision is set 20% larger */
                    vec[2]*= 1.2f;
                    normalize_v3(vec);
                    inpr= - vec[2];

                    if(inpr>=lar->spotsi) {
                        if(z<minz) minz= z;
                        if(z>maxz) maxz= z;
                    }
                }
            }
        }
    }

    MEM_freeN(clipflag);
    
    /* set clipping min and max */
    if(minz < maxz) {
        float delta= (maxz - minz); /* threshold to prevent precision issues */
        
        //printf("minz %f maxz %f delta %f\n", minz, maxz, delta);
        if(lar->bufflag & LA_SHADBUF_AUTO_START)
            lar->shb->d= minz - delta*0.02f;    /* 0.02 is arbitrary... needs more thinking! */
        if(lar->bufflag & LA_SHADBUF_AUTO_END)
            lar->shb->clipend= maxz + delta*0.1f;
        
        /* bias was calculated as percentage, we scale it to prevent animation issues */
        delta= (lar->clipend-lar->clipsta)/(lar->shb->clipend-lar->shb->d);
        //printf("bias delta %f\n", delta);
        lar->shb->bias= (int) (delta*(float)lar->shb->bias);
    }
}

static void makeflatshadowbuf(Render *re, LampRen *lar, float *jitbuf)
{
    ShadBuf *shb= lar->shb;
    int *rectz, samples;

    /* zbuffering */
    rectz= MEM_mapallocN(sizeof(int)*shb->size*shb->size, "makeshadbuf");
    
    for(samples=0; samples<shb->totbuf; samples++) {
        zbuffer_shadow(re, shb->persmat, lar, rectz, shb->size, jitbuf[2*samples], jitbuf[2*samples+1]);
        /* create Z tiles (for compression): this system is 24 bits!!! */
        compress_shadowbuf(shb, rectz, lar->mode & LA_SQUARE);

        if(re->test_break(re->tbh))
            break;
    }
    
    MEM_freeN(rectz);
}

static void makedeepshadowbuf(Render *re, LampRen *lar, float *jitbuf)
{
    ShadBuf *shb= lar->shb;
    APixstr *apixbuf;
    APixstrand *apixbufstrand= NULL;
    ListBase apsmbase= {NULL, NULL};

    /* zbuffering */
    apixbuf= MEM_callocN(sizeof(APixstr)*shb->size*shb->size, "APixbuf");
    if(re->totstrand)
        apixbufstrand= MEM_callocN(sizeof(APixstrand)*shb->size*shb->size, "APixbufstrand");

    zbuffer_abuf_shadow(re, lar, shb->persmat, apixbuf, apixbufstrand, &apsmbase, shb->size,
        shb->totbuf, (float(*)[2])jitbuf);

    /* create Z tiles (for compression): this system is 24 bits!!! */
    compress_deepshadowbuf(re, shb, apixbuf, apixbufstrand);
    
    MEM_freeN(apixbuf);
    if(apixbufstrand)
        MEM_freeN(apixbufstrand);
    freepsA(&apsmbase);
}

void makeshadowbuf(Render *re, LampRen *lar)
{
    ShadBuf *shb= lar->shb;
    float wsize, *jitbuf, twozero[2]= {0.0f, 0.0f}, angle, temp;
    
    if(lar->bufflag & (LA_SHADBUF_AUTO_START|LA_SHADBUF_AUTO_END))
        shadowbuf_autoclip(re, lar);
    
    /* just to enforce identical behaviour of all irregular buffers */
    if(lar->buftype==LA_SHADBUF_IRREGULAR)
        shb->size= 1024;
    
    /* matrices and window: in winmat the transformation is being put,
        transforming from observer view to lamp view, including lamp window matrix */
    
    angle= saacos(lar->spotsi);
    temp= 0.5f*shb->size*cos(angle)/sin(angle);
    shb->pixsize= (shb->d)/temp;
    wsize= shb->pixsize*(shb->size/2.0f);
    
    perspective_m4( shb->winmat,-wsize, wsize, -wsize, wsize, shb->d, shb->clipend);
    mult_m4_m4m4(shb->persmat, shb->winmat, shb->viewmat);

    if(ELEM3(lar->buftype, LA_SHADBUF_REGULAR, LA_SHADBUF_HALFWAY, LA_SHADBUF_DEEP)) {
        shb->totbuf= lar->buffers;

        /* jitter, weights - not threadsafe! */
        BLI_lock_thread(LOCK_CUSTOM1);
        shb->jit= give_jitter_tab(get_render_shadow_samples(&re->r, shb->samp));
        make_jitter_weight_tab(re, shb, lar->filtertype);
        BLI_unlock_thread(LOCK_CUSTOM1);
        
        if(shb->totbuf==4) jitbuf= give_jitter_tab(2);
        else if(shb->totbuf==9) jitbuf= give_jitter_tab(3);
        else jitbuf= twozero;
        
        /* zbuffering */
        if(lar->buftype == LA_SHADBUF_DEEP) {
            makedeepshadowbuf(re, lar, jitbuf);
            shb->totbuf= 1;
        }
        else
            makeflatshadowbuf(re, lar, jitbuf);

        /* printf("lampbuf %d\n", sizeoflampbuf(shb)); */
    }
}

static void *do_shadow_thread(void *re_v)
{
    Render *re= (Render*)re_v;
    LampRen *lar;

    do {
        BLI_lock_thread(LOCK_CUSTOM1);
        for(lar=re->lampren.first; lar; lar=lar->next) {
            if(lar->shb && !lar->thread_assigned) {
                lar->thread_assigned= 1;
                break;
            }
        }
        BLI_unlock_thread(LOCK_CUSTOM1);

        /* if type is irregular, this only sets the perspective matrix and autoclips */
        if(lar) {
            makeshadowbuf(re, lar);
            BLI_lock_thread(LOCK_CUSTOM1);
            lar->thread_ready= 1;
            BLI_unlock_thread(LOCK_CUSTOM1);
        }
    } while(lar && !re->test_break(re->tbh));

    return NULL;
}

static volatile int g_break= 0;
static int thread_break(void *UNUSED(arg))
{
    return g_break;
}

void threaded_makeshadowbufs(Render *re)
{
    ListBase threads;
    LampRen *lar;
    int a, totthread= 0;
    int (*test_break)(void *);

    /* count number of threads to use */
    if(G.rendering) {
        for(lar=re->lampren.first; lar; lar= lar->next)
            if(lar->shb)
                totthread++;
        
        totthread= MIN2(totthread, re->r.threads);
    }
    else
        totthread= 1; /* preview render */

    if(totthread <= 1) {
        for(lar=re->lampren.first; lar; lar= lar->next) {
            if(re->test_break(re->tbh)) break;
            if(lar->shb) {
                /* if type is irregular, this only sets the perspective matrix and autoclips */
                makeshadowbuf(re, lar);
            }
        }
    }
    else {
        /* swap test break function */
        test_break= re->test_break;
        re->test_break= thread_break;

        for(lar=re->lampren.first; lar; lar= lar->next) {
            lar->thread_assigned= 0;
            lar->thread_ready= 0;
        }

        BLI_init_threads(&threads, do_shadow_thread, totthread);
        
        for(a=0; a<totthread; a++)
            BLI_insert_thread(&threads, re);

        /* keep rendering as long as there are shadow buffers not ready */
        do {
            if((g_break=test_break(re->tbh)))
                break;

            PIL_sleep_ms(50);

            BLI_lock_thread(LOCK_CUSTOM1);
            for(lar=re->lampren.first; lar; lar= lar->next)
                if(lar->shb && !lar->thread_ready)
                    break;
            BLI_unlock_thread(LOCK_CUSTOM1);
        } while(lar);
    
        BLI_end_threads(&threads);

        /* unset threadsafety */
        re->test_break= test_break;
        g_break= 0;
    }
}

void freeshadowbuf(LampRen *lar)
{
    if(lar->shb) {
        ShadBuf *shb= lar->shb;
        ShadSampleBuf *shsample;
        int b, v;
        
        for(shsample= shb->buffers.first; shsample; shsample= shsample->next) {
            if(shsample->deepbuf) {
                v= shb->size*shb->size;
                for(b=0; b<v; b++)
                    if(shsample->deepbuf[b])
                        MEM_freeN(shsample->deepbuf[b]);
                    
                MEM_freeN(shsample->deepbuf);
                MEM_freeN(shsample->totbuf);
            }
            else {
                intptr_t *ztile= shsample->zbuf;
                char *ctile= shsample->cbuf;
                
                v= (shb->size*shb->size)/256;
                for(b=0; b<v; b++, ztile++, ctile++)
                    if(*ctile) MEM_freeN((void *) *ztile);
                
                MEM_freeN(shsample->zbuf);
                MEM_freeN(shsample->cbuf);
            }
        }
        BLI_freelistN(&shb->buffers);
        
        if(shb->weight) MEM_freeN(shb->weight);
        MEM_freeN(lar->shb);
        
        lar->shb= NULL;
    }
}


static int firstreadshadbuf(ShadBuf *shb, ShadSampleBuf *shsample, int **rz, int xs, int ys, int nr)
{
    /* return a 1 if fully compressed shadbuf-tile && z==const */
    int ofs;
    char *ct;

    if(shsample->deepbuf)
        return 0;

    /* always test borders of shadowbuffer */
    if(xs<0) xs= 0; else if(xs>=shb->size) xs= shb->size-1;
    if(ys<0) ys= 0; else if(ys>=shb->size) ys= shb->size-1;

    /* calc z */
    ofs= (ys>>4)*(shb->size>>4) + (xs>>4);
    ct= shsample->cbuf+ofs;
    if(*ct==0) {
        if(nr==0) {
            *rz= *( (int **)(shsample->zbuf+ofs) );
            return 1;
        }
        else if(*rz!= *( (int **)(shsample->zbuf+ofs) )) return 0;
        
        return 1;
    }
    
    return 0;
}

static float readdeepvisibility(DeepSample *dsample, int tot, int z, int bias, float *biast)
{
    DeepSample *ds, *prevds;
    float t;
    int a;

    /* tricky stuff here; we use ints which can overflow easily with bias values */

    ds= dsample;
    for(a=0; a<tot && (z-bias > ds->z); a++, ds++)
        ;

    if(a == tot) {
        if(biast)
            *biast= 0.0f;
        return (ds-1)->v; /* completely behind all samples */
    }
    
    /* check if this read needs bias blending */
    if(biast) {
        if(z > ds->z)
            *biast= (float)(z - ds->z)/(float)bias;
        else
            *biast= 0.0f;
    }

    if(a == 0)
        return 1.0f; /* completely in front of all samples */

    /* converting to float early here because ds->z - prevds->z can overflow */
    prevds= ds-1;
    t= ((float)(z-bias) - (float)prevds->z)/((float)ds->z - (float)prevds->z);
    return t*ds->v + (1.0f-t)*prevds->v;
}

static float readdeepshadowbuf(ShadBuf *shb, ShadSampleBuf *shsample, int bias, int xs, int ys, int zs)
{
    float v, biasv, biast;
    int ofs, tot;

    if(zs < - 0x7FFFFE00 + bias)
        return 1.0; /* extreme close to clipstart */

    /* calc z */
    ofs= ys*shb->size + xs;
    tot= shsample->totbuf[ofs];
    if(tot == 0)
        return 1.0f;

    v= readdeepvisibility(shsample->deepbuf[ofs], tot, zs, bias, &biast);

    if(biast != 0.0f) {
        /* in soft bias area */
        biasv= readdeepvisibility(shsample->deepbuf[ofs], tot, zs, 0, 0);

        biast= biast*biast;
        return (1.0f-biast)*v + biast*biasv;
    }

    return v;
}

/* return 1.0 : fully in light */
static float readshadowbuf(ShadBuf *shb, ShadSampleBuf *shsample, int bias, int xs, int ys, int zs) 
{
    float temp;
    int *rz, ofs;
    int zsamp=0;
    char *ct, *cz;

    /* simpleclip */
    /* if(xs<0 || ys<0) return 1.0; */
    /* if(xs>=shb->size || ys>=shb->size) return 1.0; */
    
    /* always test borders of shadowbuffer */
    if(xs<0) xs= 0; else if(xs>=shb->size) xs= shb->size-1;
    if(ys<0) ys= 0; else if(ys>=shb->size) ys= shb->size-1;

    if(shsample->deepbuf)
        return readdeepshadowbuf(shb, shsample, bias, xs, ys, zs);

    /* calc z */
    ofs= (ys>>4)*(shb->size>>4) + (xs>>4);
    ct= shsample->cbuf+ofs;
    rz= *( (int **)(shsample->zbuf+ofs) );

    if(*ct==3) {
        ct= ((char *)rz)+3*16*(ys & 15)+3*(xs & 15);
        cz= (char *)&zsamp;
        cz[ACOMP]= ct[0];
        cz[BCOMP]= ct[1];
        cz[GCOMP]= ct[2];
    }
    else if(*ct==2) {
        ct= ((char *)rz);
        ct+= 4+2*16*(ys & 15)+2*(xs & 15);
        zsamp= *rz;
    
        cz= (char *)&zsamp;
        cz[BCOMP]= ct[0];
        cz[GCOMP]= ct[1];
    }
    else if(*ct==1) {
        ct= ((char *)rz);
        ct+= 4+16*(ys & 15)+(xs & 15);
        zsamp= *rz;

        cz= (char *)&zsamp;
        cz[GCOMP]= ct[0];

    }
    else {
        /* got warning on this for 64 bits.... */
        /* but it's working code! in this case rz is not a pointer but zvalue (ton) */
        zsamp= GET_INT_FROM_POINTER(rz);
    }

    /* tricky stuff here; we use ints which can overflow easily with bias values */
    
    if(zsamp > zs) return 1.0;      /* absolute no shadow */
    else if(zs < - 0x7FFFFE00 + bias) return 1.0;   /* extreme close to clipstart */
    else if(zsamp < zs-bias) return 0.0 ;   /* absolute in shadow */
    else {                  /* soft area */
        
        temp=  ( (float)(zs- zsamp) )/(float)bias;
        return 1.0f - temp*temp;
            
    }
}

static void shadowbuf_project_co(float *x, float *y, float *z, ShadBuf *shb, const float co[3])
{
    float hco[4], size= 0.5f*(float)shb->size;

    copy_v3_v3(hco, co);
    hco[3]= 1.0f;

    mul_m4_v4(shb->persmat, hco);

    *x= size*(1.0f+hco[0]/hco[3]);
    *y= size*(1.0f+hco[1]/hco[3]);
    if(z) *z= (hco[2]/hco[3]);
}

/* the externally called shadow testing (reading) function */
/* return 1.0: no shadow at all */
float testshadowbuf(Render *re, ShadBuf *shb, const float co[3], const float dxco[3], const float dyco[3], float inp, float mat_bias)
{
    ShadSampleBuf *shsample;
    float fac, dco[3], dx[3], dy[3], shadfac=0.0f;
    float xs1, ys1, zs1, *jit, *weight, xres, yres, biasf;
    int xs, ys, zs, bias, *rz;
    short a, num;
    
    /* crash preventer */
    if(shb->buffers.first==NULL)
        return 1.0f;
    
    /* when facing away, assume fully in shadow */
    if(inp <= 0.0f)
        return 0.0f;

    /* project coordinate to pixel space */
    shadowbuf_project_co(&xs1, &ys1, &zs1, shb, co);

    /* clip z coordinate, z is projected so that (-1.0, 1.0) matches
       (clipstart, clipend), so we can do this simple test */
    if(zs1>=1.0f)
        return 0.0f;
    else if(zs1<= -1.0f)
        return 1.0f;

    zs= ((float)0x7FFFFFFF)*zs1;

    /* take num*num samples, increase area with fac */
    num= get_render_shadow_samples(&re->r, shb->samp);
    num= num*num;
    fac= shb->soft;
    
    /* compute z bias */
    if(mat_bias!=0.0f) biasf= shb->bias*mat_bias;
    else biasf= shb->bias;
    /* with inp==1.0, bias is half the size. correction value was 1.1, giving errors 
       on cube edges, with one side being almost frontal lighted (ton)  */
    bias= (1.5f-inp*inp)*biasf;
    
    /* in case of no filtering we can do things simpler */
    if(num==1) {
        for(shsample= shb->buffers.first; shsample; shsample= shsample->next)
            shadfac += readshadowbuf(shb, shsample, bias, (int)xs1, (int)ys1, zs);
        
        return shadfac/(float)shb->totbuf;
    }

    /* calculate filter size */
    add_v3_v3v3(dco, co, dxco);
    shadowbuf_project_co(&dx[0], &dx[1], NULL, shb, dco);
    dx[0]= xs1 - dx[0];
    dx[1]= ys1 - dx[1];

    add_v3_v3v3(dco, co, dyco);
    shadowbuf_project_co(&dy[0], &dy[1], NULL, shb, dco);
    dy[0]= xs1 - dy[0];
    dy[1]= ys1 - dy[1];
    
    xres= fac*(fabs(dx[0]) + fabs(dy[0]));
    yres= fac*(fabs(dx[1]) + fabs(dy[1]));
    if(xres<1.0f) xres= 1.0f;
    if(yres<1.0f) yres= 1.0f;
    
    /* make xs1/xs1 corner of sample area */
    xs1 -= xres*0.5f;
    ys1 -= yres*0.5f;

    /* in case we have a constant value in a tile, we can do quicker lookup */
    if(xres<16.0f && yres<16.0f) {
        shsample= shb->buffers.first;
        if(firstreadshadbuf(shb, shsample, &rz, (int)xs1, (int)ys1, 0)) {
            if(firstreadshadbuf(shb, shsample, &rz, (int)(xs1+xres), (int)ys1, 1)) {
                if(firstreadshadbuf(shb, shsample, &rz, (int)xs1, (int)(ys1+yres), 1)) {
                    if(firstreadshadbuf(shb, shsample, &rz, (int)(xs1+xres), (int)(ys1+yres), 1)) {
                        return readshadowbuf(shb, shsample, bias,(int)xs1, (int)ys1, zs);
                    }
                }
            }
        }
    }
    
    /* full jittered shadow buffer lookup */
    for(shsample= shb->buffers.first; shsample; shsample= shsample->next) {
        jit= shb->jit;
        weight= shb->weight;
        
        for(a=num; a>0; a--, jit+=2, weight++) {
            /* instead of jit i tried random: ugly! */
            /* note: the plus 0.5 gives best sampling results, jit goes from -0.5 to 0.5 */
            /* xs1 and ys1 are already corrected to be corner of sample area */
            xs= xs1 + xres*(jit[0] + 0.5f);
            ys= ys1 + yres*(jit[1] + 0.5f);
            
            shadfac+= *weight * readshadowbuf(shb, shsample, bias, xs, ys, zs);
        }
    }

    /* Renormalizes for the sample number: */
    return shadfac/(float)shb->totbuf;
}

/* different function... sampling behind clipend can be LIGHT, bias is negative! */
/* return: light */
static float readshadowbuf_halo(ShadBuf *shb, ShadSampleBuf *shsample, int xs, int ys, int zs)
{
    float temp;
    int *rz, ofs;
    int bias, zbias, zsamp;
    char *ct, *cz;

    /* negative! The other side is more important */
    bias= -shb->bias;
    
    /* simpleclip */
    if(xs<0 || ys<0) return 0.0;
    if(xs>=shb->size || ys>=shb->size) return 0.0;

    /* calc z */
    ofs= (ys>>4)*(shb->size>>4) + (xs>>4);
    ct= shsample->cbuf+ofs;
    rz= *( (int **)(shsample->zbuf+ofs) );

    if(*ct==3) {
        ct= ((char *)rz)+3*16*(ys & 15)+3*(xs & 15);
        cz= (char *)&zsamp;
        zsamp= 0;
        cz[ACOMP]= ct[0];
        cz[BCOMP]= ct[1];
        cz[GCOMP]= ct[2];
    }
    else if(*ct==2) {
        ct= ((char *)rz);
        ct+= 4+2*16*(ys & 15)+2*(xs & 15);
        zsamp= *rz;
    
        cz= (char *)&zsamp;
        cz[BCOMP]= ct[0];
        cz[GCOMP]= ct[1];
    }
    else if(*ct==1) {
        ct= ((char *)rz);
        ct+= 4+16*(ys & 15)+(xs & 15);
        zsamp= *rz;

        cz= (char *)&zsamp;
        cz[GCOMP]= ct[0];

    }
    else {
        /* same as before */
        /* still working code! (ton) */
        zsamp= GET_INT_FROM_POINTER(rz);
    }

    /* NO schadow when sampled at 'eternal' distance */

    if(zsamp >= 0x7FFFFE00) return 1.0; 

    if(zsamp > zs) return 1.0;      /* absolute no shadww */
    else {
        /* bias is negative, so the (zs-bias) can be beyond 0x7fffffff */
        zbias= 0x7fffffff - zs;
        if(zbias > -bias) {
            if( zsamp < zs-bias) return 0.0 ;   /* absolute in shadow */
        }
        else return 0.0 ;   /* absolute shadow */
    }

    /* soft area */
    
    temp=  ( (float)(zs- zsamp) )/(float)bias;
    return 1.0f - temp*temp;
}


float shadow_halo(LampRen *lar, const float p1[3], const float p2[3])
{
    /* p1 p2 already are rotated in spot-space */
    ShadBuf *shb= lar->shb;
    ShadSampleBuf *shsample;
    float co[4], siz;
    float labda, labdao, labdax, labday, ldx, ldy;
    float zf, xf1, yf1, zf1, xf2, yf2, zf2;
    float count, lightcount;
    int x, y, z, xs1, ys1;
    int dx = 0, dy = 0;
    
    siz= 0.5f*(float)shb->size;
    
    co[0]= p1[0];
    co[1]= p1[1];
    co[2]= p1[2]/lar->sh_zfac;
    co[3]= 1.0;
    mul_m4_v4(shb->winmat, co); /* rational hom co */
    xf1= siz*(1.0f+co[0]/co[3]);
    yf1= siz*(1.0f+co[1]/co[3]);
    zf1= (co[2]/co[3]);


    co[0]= p2[0];
    co[1]= p2[1];
    co[2]= p2[2]/lar->sh_zfac;
    co[3]= 1.0;
    mul_m4_v4(shb->winmat, co); /* rational hom co */
    xf2= siz*(1.0f+co[0]/co[3]);
    yf2= siz*(1.0f+co[1]/co[3]);
    zf2= (co[2]/co[3]);

    /* the 2dda (a pixel line formula) */

    xs1= (int)xf1;
    ys1= (int)yf1;

    if(xf1 != xf2) {
        if(xf2-xf1 > 0.0f) {
            labdax= (xf1-xs1-1.0f)/(xf1-xf2);
            ldx= -shb->shadhalostep/(xf1-xf2);
            dx= shb->shadhalostep;
        }
        else {
            labdax= (xf1-xs1)/(xf1-xf2);
            ldx= shb->shadhalostep/(xf1-xf2);
            dx= -shb->shadhalostep;
        }
    }
    else {
        labdax= 1.0;
        ldx= 0.0;
    }

    if(yf1 != yf2) {
        if(yf2-yf1 > 0.0f) {
            labday= (yf1-ys1-1.0f)/(yf1-yf2);
            ldy= -shb->shadhalostep/(yf1-yf2);
            dy= shb->shadhalostep;
        }
        else {
            labday= (yf1-ys1)/(yf1-yf2);
            ldy= shb->shadhalostep/(yf1-yf2);
            dy= -shb->shadhalostep;
        }
    }
    else {
        labday= 1.0;
        ldy= 0.0;
    }
    
    x= xs1;
    y= ys1;
    labda= count= lightcount= 0.0;

/* printf("start %x %x  \n", (int)(0x7FFFFFFF*zf1), (int)(0x7FFFFFFF*zf2)); */

    while(1) {
        labdao= labda;
        
        if(labdax==labday) {
            labdax+= ldx;
            x+= dx;
            labday+= ldy;
            y+= dy;
        }
        else {
            if(labdax<labday) {
                labdax+= ldx;
                x+= dx;
            } else {
                labday+= ldy;
                y+= dy;
            }
        }
        
        labda= MIN2(labdax, labday);
        if(labda==labdao || labda>=1.0f) break;
        
        zf= zf1 + labda*(zf2-zf1);
        count+= (float)shb->totbuf;

        if(zf<= -1.0f) lightcount += 1.0f;  /* close to the spot */
        else {
        
            /* make sure, behind the clipend we extend halolines. */
            if(zf>=1.0f) z= 0x7FFFF000;
            else z= (int)(0x7FFFF000*zf);
            
            for(shsample= shb->buffers.first; shsample; shsample= shsample->next)
                lightcount+= readshadowbuf_halo(shb, shsample, x, y, z);
            
        }
    }
    
    if(count!=0.0f) return (lightcount/count);
    return 0.0f;
    
}


/* ********************* Irregular Shadow Buffer (ISB) ************* */
/* ********** storage of all view samples in a raster of lists ***** */

/* based on several articles describing this method, like:
The Irregular Z-Buffer and its Application to Shadow Mapping
Gregory S. Johnson - William R. Mark - Christopher A. Burns 
and
Alias-Free Shadow Maps
Timo Aila and Samuli Laine
*/

/* bsp structure (actually kd tree) */

#define BSPMAX_SAMPLE   128
#define BSPMAX_DEPTH    32

/* aligned with struct rctf */
typedef struct Boxf {
    float xmin, xmax;
    float ymin, ymax;
    float zmin, zmax;
} Boxf;

typedef struct ISBBranch {
    struct ISBBranch *left, *right;
    float divider[2];
    Boxf box;
    short totsamp, index, full, unused;
    ISBSample **samples;
} ISBBranch;

typedef struct BSPFace {
    Boxf box;
    float *v1, *v2, *v3, *v4;
    int obi;        /* object for face lookup */
    int facenr;     /* index to retrieve VlakRen */
    int type;       /* only for strand now */
    short shad_alpha, is_full;
    
    /* strand caching data, optimize for point_behind_strand() */
    float radline, radline_end, len;
    float vec1[3], vec2[3], rc[3];
} BSPFace;

/* boxes are in lamp projection */
static void init_box(Boxf *box)
{
    box->xmin= 1000000.0f;
    box->xmax= 0;
    box->ymin= 1000000.0f;
    box->ymax= 0;
    box->zmin= 0x7FFFFFFF;
    box->zmax= - 0x7FFFFFFF;
}

/* use v1 to calculate boundbox */
static void bound_boxf(Boxf *box, const float v1[3])
{
    if(v1[0] < box->xmin) box->xmin= v1[0];
    if(v1[0] > box->xmax) box->xmax= v1[0];
    if(v1[1] < box->ymin) box->ymin= v1[1];
    if(v1[1] > box->ymax) box->ymax= v1[1];
    if(v1[2] < box->zmin) box->zmin= v1[2];
    if(v1[2] > box->zmax) box->zmax= v1[2];
}

/* use v1 to calculate boundbox */
static void bound_rectf(rctf *box, const float v1[2])
{
    if(v1[0] < box->xmin) box->xmin= v1[0];
    if(v1[0] > box->xmax) box->xmax= v1[0];
    if(v1[1] < box->ymin) box->ymin= v1[1];
    if(v1[1] > box->ymax) box->ymax= v1[1];
}


/* halfway splitting, for initializing a more regular tree */
static void isb_bsp_split_init(ISBBranch *root, MemArena *mem, int level)
{
    
    /* if level > 0 we create new branches and go deeper*/
    if(level > 0) {
        ISBBranch *left, *right;
        int i;
        
        /* splitpoint */
        root->divider[0]= 0.5f*(root->box.xmin+root->box.xmax);
        root->divider[1]= 0.5f*(root->box.ymin+root->box.ymax);
        
        /* find best splitpoint */
        if(root->box.xmax-root->box.xmin > root->box.ymax-root->box.ymin)
            i= root->index= 0;
        else
            i= root->index= 1;
        
        left= root->left= BLI_memarena_alloc(mem, sizeof(ISBBranch));
        right= root->right= BLI_memarena_alloc(mem, sizeof(ISBBranch));
        
        /* box info */
        left->box= root->box;
        right->box= root->box;
        if(i==0) {
            left->box.xmax= root->divider[0];
            right->box.xmin= root->divider[0];
        }
        else {
            left->box.ymax= root->divider[1];
            right->box.ymin= root->divider[1];
        }
        isb_bsp_split_init(left, mem, level-1);
        isb_bsp_split_init(right, mem, level-1);
    }
    else {
        /* we add sample array */
        root->samples= BLI_memarena_alloc(mem, BSPMAX_SAMPLE*sizeof(void *));
    }
}

/* note; if all samples on same location we just spread them over 2 new branches */
static void isb_bsp_split(ISBBranch *root, MemArena *mem)
{
    ISBBranch *left, *right;
    ISBSample *samples[BSPMAX_SAMPLE];
    int a, i;

    /* splitpoint */
    root->divider[0]= root->divider[1]= 0.0f;
    for(a=BSPMAX_SAMPLE-1; a>=0; a--) {
        root->divider[0]+= root->samples[a]->zco[0];
        root->divider[1]+= root->samples[a]->zco[1];
    }
    root->divider[0]/= BSPMAX_SAMPLE;
    root->divider[1]/= BSPMAX_SAMPLE;
    
    /* find best splitpoint */
    if(root->box.xmax-root->box.xmin > root->box.ymax-root->box.ymin)
        i= root->index= 0;
    else
        i= root->index= 1;
    
    /* new branches */
    left= root->left= BLI_memarena_alloc(mem, sizeof(ISBBranch));
    right= root->right= BLI_memarena_alloc(mem, sizeof(ISBBranch));

    /* new sample array */
    left->samples= BLI_memarena_alloc(mem, BSPMAX_SAMPLE*sizeof(void *));
    right->samples= samples; // tmp
    
    /* split samples */
    for(a=BSPMAX_SAMPLE-1; a>=0; a--) {
        int comp= 0;
        /* this prevents adding samples all to 1 branch when divider is equal to samples */
        if(root->samples[a]->zco[i] == root->divider[i])
            comp= a & 1;
        else if(root->samples[a]->zco[i] < root->divider[i])
            comp= 1;
        
        if(comp==1) {
            left->samples[left->totsamp]= root->samples[a];
            left->totsamp++;
        }
        else {
            right->samples[right->totsamp]= root->samples[a];
            right->totsamp++;
        }
    }
    
    /* copy samples from tmp */
    memcpy(root->samples, samples, right->totsamp*(sizeof(void *)));
    right->samples= root->samples;
    root->samples= NULL;
    
    /* box info */
    left->box= root->box;
    right->box= root->box;
    if(i==0) {
        left->box.xmax= root->divider[0];
        right->box.xmin= root->divider[0];
    }
    else {
        left->box.ymax= root->divider[1];
        right->box.ymin= root->divider[1];
    }
}

/* inserts sample in main tree, also splits on threshold */
/* returns 1 if error */
static int isb_bsp_insert(ISBBranch *root, MemArena *memarena, ISBSample *sample)
{
    ISBBranch *bspn= root;
    float *zco= sample->zco;
    int i= 0;
    
    /* debug counter, also used to check if something was filled in ever */
    root->totsamp++;
    
    /* going over branches until last one found */
    while(bspn->left) {
        if(zco[bspn->index] <= bspn->divider[bspn->index])
            bspn= bspn->left;
        else
            bspn= bspn->right;
        i++;
    }
    /* bspn now is the last branch */
    
    if(bspn->totsamp==BSPMAX_SAMPLE) {
        printf("error in bsp branch\n");    /* only for debug, cannot happen */
        return 1;
    }
    
    /* insert */
    bspn->samples[bspn->totsamp]= sample;
    bspn->totsamp++;

    /* split if allowed and needed */
    if(bspn->totsamp==BSPMAX_SAMPLE) {
        if(i==BSPMAX_DEPTH) {
            bspn->totsamp--;    /* stop filling in... will give errors */
            return 1;
        }
        isb_bsp_split(bspn, memarena);
    }
    return 0;
}

/* initialize vars in face, for optimal point-in-face test */
static void bspface_init_strand(BSPFace *face) 
{
    
    face->radline= 0.5f* len_v2v2(face->v1, face->v2);
    
    mid_v3_v3v3(face->vec1, face->v1, face->v2);
    if(face->v4)
        mid_v3_v3v3(face->vec2, face->v3, face->v4);
    else
        copy_v3_v3(face->vec2, face->v3);
    
    face->rc[0]= face->vec2[0]-face->vec1[0];
    face->rc[1]= face->vec2[1]-face->vec1[1];
    face->rc[2]= face->vec2[2]-face->vec1[2];
    
    face->len= face->rc[0]*face->rc[0]+ face->rc[1]*face->rc[1];
    
    if(face->len!=0.0f) {
        face->radline_end= face->radline/sqrt(face->len);
        face->len= 1.0f/face->len;
    }
}

/* brought back to a simple 2d case */
static int point_behind_strand(const float p[3], BSPFace *face)
{
    /* v1 - v2 is radius, v1 - v3 length */
    float dist, rc[2], pt[2];
    
    /* using code from dist_to_line_segment_v2(), distance vec to line-piece */

    if(face->len==0.0f) {
        rc[0]= p[0]-face->vec1[0];
        rc[1]= p[1]-face->vec1[1];
        dist= (float)(sqrt(rc[0]*rc[0]+ rc[1]*rc[1]));
        
        if(dist < face->radline)
            return 1;
    }
    else {
        float labda= ( face->rc[0]*(p[0]-face->vec1[0]) + face->rc[1]*(p[1]-face->vec1[1]) )*face->len;
        
        if(labda > -face->radline_end && labda < 1.0f+face->radline_end) { 
            /* hesse for dist: */
            //dist= (float)(fabs( (p[0]-vec2[0])*rc[1] + (p[1]-vec2[1])*rc[0])/len);
            
            pt[0]= labda*face->rc[0]+face->vec1[0];
            pt[1]= labda*face->rc[1]+face->vec1[1];
            
            rc[0]= pt[0]-p[0];
            rc[1]= pt[1]-p[1];
            dist= (float)sqrt(rc[0]*rc[0]+ rc[1]*rc[1]);
            
            if(dist < face->radline) {
                float zval= face->vec1[2] + labda*face->rc[2];
                if(p[2] > zval)
                    return 1;
            }
        }
    }
    return 0;
}


/* return 1 if inside. code derived from src/parametrizer.c */
static int point_behind_tria2d(const float p[3], const float v1[3], const float v2[3], const float v3[3])
{
    float a[2], c[2], h[2], div;
    float u, v;
    
    a[0] = v2[0] - v1[0];
    a[1] = v2[1] - v1[1];
    c[0] = v3[0] - v1[0];
    c[1] = v3[1] - v1[1];
    
    div = a[0]*c[1] - a[1]*c[0];
    if(div==0.0f)
        return 0;
    
    h[0] = p[0] - v1[0];
    h[1] = p[1] - v1[1];
    
    div = 1.0f/div;
    
    u = (h[0]*c[1] - h[1]*c[0])*div;
    if(u >= 0.0f) {
        v = (a[0]*h[1] - a[1]*h[0])*div;
        if(v >= 0.0f) {
            if( u + v <= 1.0f) {
                /* inside, now check if point p is behind */
                float z=  (1.0f-u-v)*v1[2] + u*v2[2] + v*v3[2];
                if(z <= p[2])
                    return 1;
            }
        }
    }
    
    return 0;
}

#if 0
/* tested these calls, but it gives inaccuracy, 'side' cannot be found reliably using v3 */

/* check if line v1-v2 has all rect points on other side of point v3 */
static int rect_outside_line(rctf *rect, const float v1[3], const float v2[3], const float v3[3])
{
    float a, b, c;
    int side;
    
    /* line formula for v1-v2 */
    a= v2[1]-v1[1];
    b= v1[0]-v2[0];
    c= -a*v1[0] - b*v1[1];
    side= a*v3[0] + b*v3[1] + c < 0.0f;
    
    /* the four quad points */
    if( side==(rect->xmin*a + rect->ymin*b + c >= 0.0f) )
        if( side==(rect->xmax*a + rect->ymin*b + c >= 0.0f) )
            if( side==(rect->xmax*a + rect->ymax*b + c >= 0.0f) )
                if( side==(rect->xmin*a + rect->ymax*b + c >= 0.0f) )
                    return 1;
    return 0;
}

/* check if one of the triangle edges separates all rect points on 1 side */
static int rect_isect_tria(rctf *rect, const float v1[3], const float v2[3], const float v3[3])
{
    if(rect_outside_line(rect, v1, v2, v3))
        return 0;
    if(rect_outside_line(rect, v2, v3, v1))
        return 0;
    if(rect_outside_line(rect, v3, v1, v2))
        return 0;
    return 1;
}
#endif

/* if face overlaps a branch, it executes func. recursive */
static void isb_bsp_face_inside(ISBBranch *bspn, BSPFace *face)
{
    
    /* are we descending? */
    if(bspn->left) {
        /* hrmf, the box struct cannot be addressed with index */
        if(bspn->index==0) {
            if(face->box.xmin <= bspn->divider[0])
                isb_bsp_face_inside(bspn->left, face);
            if(face->box.xmax > bspn->divider[0])
                isb_bsp_face_inside(bspn->right, face);
        }
        else {
            if(face->box.ymin <= bspn->divider[1])
                isb_bsp_face_inside(bspn->left, face);
            if(face->box.ymax > bspn->divider[1])
                isb_bsp_face_inside(bspn->right, face);
        }
    }
    else {
        /* else: end branch reached */
        int a;
        
        if(bspn->totsamp==0) return;
        
        /* check for nodes entirely in shadow, can be skipped */
        if(bspn->totsamp==bspn->full)
            return;
        
        /* if bsp node is entirely in front of face, give up */
        if(bspn->box.zmax < face->box.zmin)
            return;
        
        /* if face boundbox is outside of branch rect, give up */
        if(0==BLI_isect_rctf((rctf *)&face->box, (rctf *)&bspn->box, NULL))
            return;
        
        /* test all points inside branch */
        for(a=bspn->totsamp-1; a>=0; a--) {
            ISBSample *samp= bspn->samples[a];
            
            if((samp->facenr!=face->facenr || samp->obi!=face->obi) && samp->shadfac) {
                if(face->box.zmin < samp->zco[2]) {
                    if(BLI_in_rctf((rctf *)&face->box, samp->zco[0], samp->zco[1])) {
                        int inshadow= 0;
                        
                        if(face->type) {
                            if(point_behind_strand(samp->zco, face)) 
                                inshadow= 1;
                        }
                        else if( point_behind_tria2d(samp->zco, face->v1, face->v2, face->v3))
                            inshadow= 1;
                        else if(face->v4 && point_behind_tria2d(samp->zco, face->v1, face->v3, face->v4))
                            inshadow= 1;

                        if(inshadow) {
                            *(samp->shadfac) += face->shad_alpha;
                            /* optimize; is_full means shad_alpha==4096 */
                            if(*(samp->shadfac) >= 4096 || face->is_full) {
                                bspn->full++;
                                samp->shadfac= NULL;
                            }
                        }
                    }
                }
            }
        }
    }
}

/* based on available samples, recalculate the bounding box for bsp nodes, recursive */
static void isb_bsp_recalc_box(ISBBranch *root)
{
    if(root->left) {
        isb_bsp_recalc_box(root->left);
        isb_bsp_recalc_box(root->right);
    }
    else if(root->totsamp) {
        int a;
        
        init_box(&root->box);
        for(a=root->totsamp-1; a>=0; a--)
            bound_boxf(&root->box, root->samples[a]->zco);
    }   
}

/* callback function for zbuf clip */
static void isb_bsp_test_strand(ZSpan *zspan, int obi, int zvlnr, float *v1, float *v2, float *v3, float *v4)
{
    BSPFace face;
    
    face.v1= v1;
    face.v2= v2;
    face.v3= v3;
    face.v4= v4;
    face.obi= obi;
    face.facenr= zvlnr & ~RE_QUAD_OFFS;
    face.type= R_STRAND;
    if(R.osa)
        face.shad_alpha= (short)ceil(4096.0f*zspan->shad_alpha/(float)R.osa);
    else
        face.shad_alpha= (short)ceil(4096.0f*zspan->shad_alpha);
    
    face.is_full= (zspan->shad_alpha==1.0f);
    
    /* setup boundbox */
    init_box(&face.box);
    bound_boxf(&face.box, v1);
    bound_boxf(&face.box, v2);
    bound_boxf(&face.box, v3);
    if(v4)
        bound_boxf(&face.box, v4);
    
    /* optimize values */
    bspface_init_strand(&face);
    
    isb_bsp_face_inside((ISBBranch *)zspan->rectz, &face);
    
}

/* callback function for zbuf clip */
static void isb_bsp_test_face(ZSpan *zspan, int obi, int zvlnr, float *v1, float *v2, float *v3, float *v4) 
{
    BSPFace face;
    
    face.v1= v1;
    face.v2= v2;
    face.v3= v3;
    face.v4= v4;
    face.obi= obi;
    face.facenr= zvlnr & ~RE_QUAD_OFFS;
    face.type= 0;
    if(R.osa)
        face.shad_alpha= (short)ceil(4096.0f*zspan->shad_alpha/(float)R.osa);
    else
        face.shad_alpha= (short)ceil(4096.0f*zspan->shad_alpha);
    
    face.is_full= (zspan->shad_alpha==1.0f);
    
    /* setup boundbox */
    init_box(&face.box);
    bound_boxf(&face.box, v1);
    bound_boxf(&face.box, v2);
    bound_boxf(&face.box, v3);
    if(v4)
        bound_boxf(&face.box, v4);

    isb_bsp_face_inside((ISBBranch *)zspan->rectz, &face);
}

static int testclip_minmax(const float ho[4], const float minmax[4])
{
    float wco= ho[3];
    int flag= 0;
    
    if( ho[0] > minmax[1]*wco) flag = 1;
    else if( ho[0]< minmax[0]*wco) flag = 2;
    
    if( ho[1] > minmax[3]*wco) flag |= 4;
    else if( ho[1]< minmax[2]*wco) flag |= 8;
    
    return flag;
}

/* main loop going over all faces and check in bsp overlaps, fill in shadfac values */
static void isb_bsp_fillfaces(Render *re, LampRen *lar, ISBBranch *root)
{
    ObjectInstanceRen *obi;
    ObjectRen *obr;
    ShadBuf *shb= lar->shb;
    ZSpan zspan, zspanstrand;
    VlakRen *vlr= NULL;
    Material *ma= NULL;
    float minmaxf[4], winmat[4][4];
    int size= shb->size;
    int i, a, ok=1, lay= -1;
    
    /* further optimize, also sets minz maxz */
    isb_bsp_recalc_box(root);
    
    /* extra clipping for minmax */
    minmaxf[0]= (2.0f*root->box.xmin - size-2.0f)/size;
    minmaxf[1]= (2.0f*root->box.xmax - size+2.0f)/size;
    minmaxf[2]= (2.0f*root->box.ymin - size-2.0f)/size;
    minmaxf[3]= (2.0f*root->box.ymax - size+2.0f)/size;
    
    if(lar->mode & (LA_LAYER|LA_LAYER_SHADOW)) lay= lar->lay;
    
    /* (ab)use zspan, since we use zbuffer clipping code */
    zbuf_alloc_span(&zspan, size, size, re->clipcrop);
    
    zspan.zmulx=  ((float)size)/2.0f;
    zspan.zmuly=  ((float)size)/2.0f;
    zspan.zofsx= -0.5f;
    zspan.zofsy= -0.5f;
    
    /* pass on bsp root to zspan */
    zspan.rectz= (int *)root;
    
    /* filling methods */
    zspanstrand= zspan;
    //  zspan.zbuflinefunc= zbufline_onlyZ;
    zspan.zbuffunc= isb_bsp_test_face;
    zspanstrand.zbuffunc= isb_bsp_test_strand;
    
    for(i=0, obi=re->instancetable.first; obi; i++, obi=obi->next) {
        obr= obi->obr;

        if(obi->flag & R_TRANSFORMED)
            mult_m4_m4m4(winmat, shb->persmat, obi->mat);
        else
            copy_m4_m4(winmat, shb->persmat);

        for(a=0; a<obr->totvlak; a++) {
            
            if((a & 255)==0) vlr= obr->vlaknodes[a>>8].vlak;
            else vlr++;
            
            /* note, these conditions are copied in shadowbuf_autoclip() */
            if(vlr->mat!= ma) {
                ma= vlr->mat;
                ok= 1;
                if((ma->mode & MA_SHADBUF)==0) ok= 0;
                if(ma->material_type == MA_TYPE_WIRE) ok= 0;
                zspanstrand.shad_alpha= zspan.shad_alpha= ma->shad_alpha;
            }
            
            if(ok && (obi->lay & lay)) {
                float hoco[4][4];
                int c1, c2, c3, c4=0;
                int d1, d2, d3, d4=0;
                int partclip;
                
                /* create hocos per face, it is while render */
                projectvert(vlr->v1->co, winmat, hoco[0]); d1= testclip_minmax(hoco[0], minmaxf);
                projectvert(vlr->v2->co, winmat, hoco[1]); d2= testclip_minmax(hoco[1], minmaxf);
                projectvert(vlr->v3->co, winmat, hoco[2]); d3= testclip_minmax(hoco[2], minmaxf);
                if(vlr->v4) {
                    projectvert(vlr->v4->co, winmat, hoco[3]); d4= testclip_minmax(hoco[3], minmaxf);
                }

                /* minmax clipping */
                if(vlr->v4) partclip= d1 & d2 & d3 & d4;
                else partclip= d1 & d2 & d3;
                
                if(partclip==0) {
                    
                    /* window clipping */
                    c1= testclip(hoco[0]); 
                    c2= testclip(hoco[1]); 
                    c3= testclip(hoco[2]); 
                    if(vlr->v4)
                        c4= testclip(hoco[3]); 
                    
                    /* ***** NO WIRE YET */         
                    if(ma->material_type == MA_TYPE_WIRE) {
                        if(vlr->v4)
                            zbufclipwire(&zspan, i, a+1, vlr->ec, hoco[0], hoco[1], hoco[2], hoco[3], c1, c2, c3, c4);
                        else
                            zbufclipwire(&zspan, i, a+1, vlr->ec, hoco[0], hoco[1], hoco[2], 0, c1, c2, c3, 0);
                    }
                    else if(vlr->v4) {
                        if(vlr->flag & R_STRAND)
                            zbufclip4(&zspanstrand, i, a+1, hoco[0], hoco[1], hoco[2], hoco[3], c1, c2, c3, c4);
                        else
                            zbufclip4(&zspan, i, a+1, hoco[0], hoco[1], hoco[2], hoco[3], c1, c2, c3, c4);
                    }
                    else
                        zbufclip(&zspan, i, a+1, hoco[0], hoco[1], hoco[2], c1, c2, c3);
                    
                }
            }
        }
    }
    
    zbuf_free_span(&zspan);
}

/* returns 1 when the viewpixel is visible in lampbuffer */
static int viewpixel_to_lampbuf(ShadBuf *shb, ObjectInstanceRen *obi, VlakRen *vlr, float x, float y, float co_r[3])
{
    float hoco[4], v1[3], nor[3];
    float dface, fac, siz;
    
    RE_vlakren_get_normal(&R, obi, vlr, nor);
    copy_v3_v3(v1, vlr->v1->co);
    if(obi->flag & R_TRANSFORMED)
        mul_m4_v3(obi->mat, v1);

    /* from shadepixel() */
    dface= v1[0]*nor[0] + v1[1]*nor[1] + v1[2]*nor[2];
    hoco[3]= 1.0f;
    
    /* ortho viewplane cannot intersect using view vector originating in (0,0,0) */
    if(R.r.mode & R_ORTHO) {
        /* x and y 3d coordinate can be derived from pixel coord and winmat */
        float fx= 2.0f/(R.winx*R.winmat[0][0]);
        float fy= 2.0f/(R.winy*R.winmat[1][1]);
        
        hoco[0]= (x - 0.5f*R.winx)*fx - R.winmat[3][0]/R.winmat[0][0];
        hoco[1]= (y - 0.5f*R.winy)*fy - R.winmat[3][1]/R.winmat[1][1];
        
        /* using a*x + b*y + c*z = d equation, (a b c) is normal */
        if(nor[2]!=0.0f)
            hoco[2]= (dface - nor[0]*hoco[0] - nor[1]*hoco[1])/nor[2];
        else
            hoco[2]= 0.0f;
    }
    else {
        float div, view[3];
        
        calc_view_vector(view, x, y);
        
        div= nor[0]*view[0] + nor[1]*view[1] + nor[2]*view[2];
        if (div==0.0f) 
            return 0;
        
        fac= dface/div;
        
        hoco[0]= fac*view[0];
        hoco[1]= fac*view[1];
        hoco[2]= fac*view[2];
    }
    
    /* move 3d vector to lampbuf */
    mul_m4_v4(shb->persmat, hoco);  /* rational hom co */
    
    /* clip We can test for -1.0/1.0 because of the properties of the
     * coordinate transformations. */
    fac= fabs(hoco[3]);
    if(hoco[0]<-fac || hoco[0]>fac)
        return 0;
    if(hoco[1]<-fac || hoco[1]>fac)
        return 0;
    if(hoco[2]<-fac || hoco[2]>fac)
        return 0;
    
    siz= 0.5f*(float)shb->size;
    co_r[0]= siz*(1.0f+hoco[0]/hoco[3]) -0.5f;
    co_r[1]= siz*(1.0f+hoco[1]/hoco[3]) -0.5f;
    co_r[2]= ((float)0x7FFFFFFF)*(hoco[2]/hoco[3]);
    
    /* XXXX bias, much less than normal shadbuf, or do we need a constant? */
    co_r[2] -= 0.05f*shb->bias;
    
    return 1;
}

/* storage of shadow results, solid osa and transp case */
static void isb_add_shadfac(ISBShadfacA **isbsapp, MemArena *mem, int obi, int facenr, short shadfac, short samples)
{
    ISBShadfacA *new;
    float shadfacf;
    
    /* in osa case, the samples were filled in with factor 1.0/R.osa. if fewer samples we have to correct */
    if(R.osa)
        shadfacf= ((float)shadfac*R.osa)/(4096.0f*samples);
    else
        shadfacf= ((float)shadfac)/(4096.0f);
    
    new= BLI_memarena_alloc(mem, sizeof(ISBShadfacA));
    new->obi= obi;
    new->facenr= facenr & ~RE_QUAD_OFFS;
    new->shadfac= shadfacf;
    if(*isbsapp)
        new->next= (*isbsapp);
    else
        new->next= NULL;
    
    *isbsapp= new;
}

/* adding samples, solid case */
static int isb_add_samples(RenderPart *pa, ISBBranch *root, MemArena *memarena, ISBSample **samplebuf)
{
    int xi, yi, *xcos, *ycos;
    int sample, bsp_err= 0;
    
    /* bsp split doesn't like to handle regular sequences */
    xcos= MEM_mallocN( pa->rectx*sizeof(int), "xcos");
    ycos= MEM_mallocN( pa->recty*sizeof(int), "ycos");
    for(xi=0; xi<pa->rectx; xi++)
        xcos[xi]= xi;
    for(yi=0; yi<pa->recty; yi++)
        ycos[yi]= yi;
    BLI_array_randomize(xcos, sizeof(int), pa->rectx, 12345);
    BLI_array_randomize(ycos, sizeof(int), pa->recty, 54321);
    
    for(sample=0; sample<(R.osa?R.osa:1); sample++) {
        ISBSample *samp= samplebuf[sample], *samp1;
        
        for(yi=0; yi<pa->recty; yi++) {
            int y= ycos[yi];
            for(xi=0; xi<pa->rectx; xi++) {
                int x= xcos[xi];
                samp1= samp + y*pa->rectx + x;
                if(samp1->facenr)
                    bsp_err |= isb_bsp_insert(root, memarena, samp1);
            }
            if(bsp_err) break;
        }
    }   
    
    MEM_freeN(xcos);
    MEM_freeN(ycos);

    return bsp_err;
}

/* solid version */
/* lar->shb, pa->rectz and pa->rectp should exist */
static void isb_make_buffer(RenderPart *pa, LampRen *lar)
{
    ShadBuf *shb= lar->shb;
    ISBData *isbdata;
    ISBSample *samp, *samplebuf[16];    /* should be RE_MAX_OSA */
    ISBBranch root;
    MemArena *memarena;
    intptr_t *rd;
    int *recto, *rectp, x, y, sindex, sample, bsp_err=0;
    
    /* storage for shadow, per thread */
    isbdata= shb->isb_result[pa->thread];
    
    /* to map the shi->xs and ys coordinate */
    isbdata->minx= pa->disprect.xmin;
    isbdata->miny= pa->disprect.ymin;
    isbdata->rectx= pa->rectx;
    isbdata->recty= pa->recty;
    
    /* branches are added using memarena (32k branches) */
    memarena = BLI_memarena_new(0x8000 * sizeof(ISBBranch), "isb arena");
    BLI_memarena_use_calloc(memarena);
    
    /* samplebuf is in camera view space (pixels) */
    for(sample=0; sample<(R.osa?R.osa:1); sample++)
        samplebuf[sample]= MEM_callocN(sizeof(ISBSample)*pa->rectx*pa->recty, "isb samplebuf");
    
    /* for end result, ISBSamples point to this in non OSA case, otherwise to pixstruct->shadfac */
    if(R.osa==0)
        isbdata->shadfacs= MEM_callocN(pa->rectx*pa->recty*sizeof(short), "isb shadfacs");
    
    /* setup bsp root */
    memset(&root, 0, sizeof(ISBBranch));
    root.box.xmin= (float)shb->size;
    root.box.ymin= (float)shb->size;
    
    /* create the sample buffers */
    for(sindex=0, y=0; y<pa->recty; y++) {
        for(x=0; x<pa->rectx; x++, sindex++) {
            
            /* this makes it a long function, but splitting it out would mean 10+ arguments */
            /* first check OSA case */
            if(R.osa) {
                rd= pa->rectdaps + sindex;
                if(*rd) {
                    float xs= (float)(x + pa->disprect.xmin);
                    float ys= (float)(y + pa->disprect.ymin);
                    
                    for(sample=0; sample<R.osa; sample++) {
                        PixStr *ps= (PixStr *)(*rd);
                        int mask= (1<<sample);
                        
                        while(ps) {
                            if(ps->mask & mask)
                                break;
                            ps= ps->next;
                        }
                        if(ps && ps->facenr>0) {
                            ObjectInstanceRen *obi= &R.objectinstance[ps->obi];
                            ObjectRen *obr= obi->obr;
                            VlakRen *vlr= RE_findOrAddVlak(obr, (ps->facenr-1) & RE_QUAD_MASK);
                            
                            samp= samplebuf[sample] + sindex;
                            /* convert image plane pixel location to lamp buffer space */
                            if(viewpixel_to_lampbuf(shb, obi, vlr, xs + R.jit[sample][0], ys + R.jit[sample][1], samp->zco)) {
                                samp->obi= ps->obi;
                                samp->facenr= ps->facenr & ~RE_QUAD_OFFS;
                                ps->shadfac= 0;
                                samp->shadfac= &ps->shadfac;
                                bound_rectf((rctf *)&root.box, samp->zco);
                            }
                        }
                    }
                }
            }
            else {
                rectp= pa->rectp + sindex;
                recto= pa->recto + sindex;
                if(*rectp>0) {
                    ObjectInstanceRen *obi= &R.objectinstance[*recto];
                    ObjectRen *obr= obi->obr;
                    VlakRen *vlr= RE_findOrAddVlak(obr, (*rectp-1) & RE_QUAD_MASK);
                    float xs= (float)(x + pa->disprect.xmin);
                    float ys= (float)(y + pa->disprect.ymin);
                    
                    samp= samplebuf[0] + sindex;
                    /* convert image plane pixel location to lamp buffer space */
                    if(viewpixel_to_lampbuf(shb, obi, vlr, xs, ys, samp->zco)) {
                        samp->obi= *recto;
                        samp->facenr= *rectp & ~RE_QUAD_OFFS;
                        samp->shadfac= isbdata->shadfacs + sindex;
                        bound_rectf((rctf *)&root.box, samp->zco);
                    }
                }
            }
        }
    }
    
    /* simple method to see if we have samples */
    if(root.box.xmin != (float)shb->size) {
        /* now create a regular split, root.box has the initial bounding box of all pixels */
        /* split bsp 8 levels deep, in regular grid (16 x 16) */
        isb_bsp_split_init(&root, memarena, 8);
        
        /* insert all samples in BSP now */
        bsp_err= isb_add_samples(pa, &root, memarena, samplebuf);
            
        if(bsp_err==0) {
            /* go over all faces and fill in shadow values */
            
            isb_bsp_fillfaces(&R, lar, &root);  /* shb->persmat should have been calculated */
            
            /* copy shadow samples to persistent buffer, reduce memory overhead */
            if(R.osa) {
                ISBShadfacA **isbsa= isbdata->shadfaca= MEM_callocN(pa->rectx*pa->recty*sizeof(void *), "isb shadfacs");
                
                isbdata->memarena = BLI_memarena_new(0x8000 * sizeof(ISBSampleA), "isb arena");
                BLI_memarena_use_calloc(isbdata->memarena);

                for(rd= pa->rectdaps, x=pa->rectx*pa->recty; x>0; x--, rd++, isbsa++) {
                    
                    if(*rd) {
                        PixStr *ps= (PixStr *)(*rd);
                        while(ps) {
                            if(ps->shadfac)
                                isb_add_shadfac(isbsa, isbdata->memarena, ps->obi, ps->facenr, ps->shadfac, count_mask(ps->mask));
                            ps= ps->next;
                        }
                    }
                }
            }
        }
    }
    else {
        if(isbdata->shadfacs) {
            MEM_freeN(isbdata->shadfacs);
            isbdata->shadfacs= NULL;
        }
    }

    /* free BSP */
    BLI_memarena_free(memarena);
    
    /* free samples */
    for(x=0; x<(R.osa?R.osa:1); x++)
        MEM_freeN(samplebuf[x]);
    
    if(bsp_err) printf("error in filling bsp\n");
}

/* add sample to buffer, isbsa is the root sample in a buffer */
static ISBSampleA *isb_alloc_sample_transp(ISBSampleA **isbsa, MemArena *mem)
{
    ISBSampleA *new;
    
    new= BLI_memarena_alloc(mem, sizeof(ISBSampleA));
    if(*isbsa)
        new->next= (*isbsa);
    else
        new->next= NULL;
    
    *isbsa= new;
    return new;
}

/* adding samples in BSP, transparent case */
static int isb_add_samples_transp(RenderPart *pa, ISBBranch *root, MemArena *memarena, ISBSampleA ***samplebuf)
{
    int xi, yi, *xcos, *ycos;
    int sample, bsp_err= 0;
    
    /* bsp split doesn't like to handle regular sequences */
    xcos= MEM_mallocN( pa->rectx*sizeof(int), "xcos");
    ycos= MEM_mallocN( pa->recty*sizeof(int), "ycos");
    for(xi=0; xi<pa->rectx; xi++)
        xcos[xi]= xi;
    for(yi=0; yi<pa->recty; yi++)
        ycos[yi]= yi;
    BLI_array_randomize(xcos, sizeof(int), pa->rectx, 12345);
    BLI_array_randomize(ycos, sizeof(int), pa->recty, 54321);
    
    for(sample=0; sample<(R.osa?R.osa:1); sample++) {
        ISBSampleA **samp= samplebuf[sample], *samp1;
        
        for(yi=0; yi<pa->recty; yi++) {
            int y= ycos[yi];
            for(xi=0; xi<pa->rectx; xi++) {
                int x= xcos[xi];
                
                samp1= *(samp + y*pa->rectx + x);
                while(samp1) {
                    bsp_err |= isb_bsp_insert(root, memarena, (ISBSample *)samp1);
                    samp1= samp1->next;
                }
            }
            if(bsp_err) break;
        }
    }   
    
    MEM_freeN(xcos);
    MEM_freeN(ycos);
    
    return bsp_err;
}


/* Ztransp version */
/* lar->shb, pa->rectz and pa->rectp should exist */
static void isb_make_buffer_transp(RenderPart *pa, APixstr *apixbuf, LampRen *lar)
{
    ShadBuf *shb= lar->shb;
    ISBData *isbdata;
    ISBSampleA *samp, **samplebuf[16];  /* MAX_OSA */
    ISBBranch root;
    MemArena *memarena;
    APixstr *ap;
    int x, y, sindex, sample, bsp_err=0;
    
    /* storage for shadow, per thread */
    isbdata= shb->isb_result[pa->thread];
    
    /* to map the shi->xs and ys coordinate */
    isbdata->minx= pa->disprect.xmin;
    isbdata->miny= pa->disprect.ymin;
    isbdata->rectx= pa->rectx;
    isbdata->recty= pa->recty;
    
    /* branches are added using memarena (32k branches) */
    memarena = BLI_memarena_new(0x8000 * sizeof(ISBBranch), "isb arena");
    BLI_memarena_use_calloc(memarena);
    
    /* samplebuf is in camera view space (pixels) */
    for(sample=0; sample<(R.osa?R.osa:1); sample++)
        samplebuf[sample]= MEM_callocN(sizeof(void *)*pa->rectx*pa->recty, "isb alpha samplebuf");
    
    /* setup bsp root */
    memset(&root, 0, sizeof(ISBBranch));
    root.box.xmin= (float)shb->size;
    root.box.ymin= (float)shb->size;

    /* create the sample buffers */
    for(ap= apixbuf, sindex=0, y=0; y<pa->recty; y++) {
        for(x=0; x<pa->rectx; x++, sindex++, ap++) {
            
            if(ap->p[0]) {
                APixstr *apn;
                float xs= (float)(x + pa->disprect.xmin);
                float ys= (float)(y + pa->disprect.ymin);
                
                for(apn=ap; apn; apn= apn->next) {
                    int a;
                    for(a=0; a<4; a++) {
                        if(apn->p[a]) {
                            ObjectInstanceRen *obi= &R.objectinstance[apn->obi[a]];
                            ObjectRen *obr= obi->obr;
                            VlakRen *vlr= RE_findOrAddVlak(obr, (apn->p[a]-1) & RE_QUAD_MASK);
                            float zco[3];
                            
                            /* here we store shadfac, easier to create the end storage buffer. needs zero'ed, multiple shadowbufs use it */
                            apn->shadfac[a]= 0;
                            
                            if(R.osa) {
                                for(sample=0; sample<R.osa; sample++) {
                                    int mask= (1<<sample);
                                    
                                    if(apn->mask[a] & mask) {
                                        
                                        /* convert image plane pixel location to lamp buffer space */
                                        if(viewpixel_to_lampbuf(shb, obi, vlr, xs + R.jit[sample][0], ys + R.jit[sample][1], zco)) {
                                            samp= isb_alloc_sample_transp(samplebuf[sample] + sindex, memarena);
                                            samp->obi= apn->obi[a];
                                            samp->facenr= apn->p[a] & ~RE_QUAD_OFFS;
                                            samp->shadfac= &apn->shadfac[a];
                                            
                                            copy_v3_v3(samp->zco, zco);
                                            bound_rectf((rctf *)&root.box, samp->zco);
                                        }
                                    }
                                }
                            }
                            else {
                                
                                /* convert image plane pixel location to lamp buffer space */
                                if(viewpixel_to_lampbuf(shb, obi, vlr, xs, ys, zco)) {
                                    
                                    samp= isb_alloc_sample_transp(samplebuf[0] + sindex, memarena);
                                    samp->obi= apn->obi[a];
                                    samp->facenr= apn->p[a] & ~RE_QUAD_OFFS;
                                    samp->shadfac= &apn->shadfac[a];
                                    
                                    copy_v3_v3(samp->zco, zco);
                                    bound_rectf((rctf *)&root.box, samp->zco);
                                }
                            }
                        }
                    }
                }
            }
        }
    }
    
    /* simple method to see if we have samples */
    if(root.box.xmin != (float)shb->size) {
        /* now create a regular split, root.box has the initial bounding box of all pixels */
        /* split bsp 8 levels deep, in regular grid (16 x 16) */
        isb_bsp_split_init(&root, memarena, 8);
        
        /* insert all samples in BSP now */
        bsp_err= isb_add_samples_transp(pa, &root, memarena, samplebuf);
        
        if(bsp_err==0) {
            ISBShadfacA **isbsa;
            
            /* go over all faces and fill in shadow values */
            isb_bsp_fillfaces(&R, lar, &root);  /* shb->persmat should have been calculated */
            
            /* copy shadow samples to persistent buffer, reduce memory overhead */
            isbsa= isbdata->shadfaca= MEM_callocN(pa->rectx*pa->recty*sizeof(void *), "isb shadfacs");
            
            isbdata->memarena = BLI_memarena_new(0x8000 * sizeof(ISBSampleA), "isb arena");
            
            for(ap= apixbuf, x=pa->rectx*pa->recty; x>0; x--, ap++, isbsa++) {
                    
                if(ap->p[0]) {
                    APixstr *apn;
                    for(apn=ap; apn; apn= apn->next) {
                        int a;
                        for(a=0; a<4; a++) {
                            if(apn->p[a] && apn->shadfac[a]) {
                                if(R.osa)
                                    isb_add_shadfac(isbsa, isbdata->memarena, apn->obi[a], apn->p[a], apn->shadfac[a], count_mask(apn->mask[a]));
                                else
                                    isb_add_shadfac(isbsa, isbdata->memarena, apn->obi[a], apn->p[a], apn->shadfac[a], 0);
                            }
                        }
                    }
                }
            }
        }
    }

    /* free BSP */
    BLI_memarena_free(memarena);

    /* free samples */
    for(x=0; x<(R.osa?R.osa:1); x++)
        MEM_freeN(samplebuf[x]);

    if(bsp_err) printf("error in filling bsp\n");
}



/* exported */

/* returns amount of light (1.0 = no shadow) */
/* note, shadepixel() rounds the coordinate, not the real sample info */
float ISB_getshadow(ShadeInput *shi, ShadBuf *shb)
{
    /* if raytracing, we can't accept irregular shadow */
    if(shi->depth==0) {
        ISBData *isbdata= shb->isb_result[shi->thread];
        
        if(isbdata) {
            if(isbdata->shadfacs || isbdata->shadfaca) {
                int x= shi->xs - isbdata->minx;
                
                if(x >= 0 && x < isbdata->rectx) {
                    int y= shi->ys - isbdata->miny;
            
                    if(y >= 0 && y < isbdata->recty) {
                        if(isbdata->shadfacs) {
                            short *sp= isbdata->shadfacs + y*isbdata->rectx + x;
                            return *sp>=4096?0.0f:1.0f - ((float)*sp)/4096.0f;
                        }
                        else {
                            int sindex= y*isbdata->rectx + x;
                            int obi= shi->obi - R.objectinstance;
                            ISBShadfacA *isbsa= *(isbdata->shadfaca + sindex);
                            
                            while(isbsa) {
                                if(isbsa->facenr==shi->facenr+1 && isbsa->obi==obi)
                                    return isbsa->shadfac>=1.0f?0.0f:1.0f - isbsa->shadfac;
                                isbsa= isbsa->next;
                            }
                        }
                    }
                }
            }
        }
    }
    return 1.0f;
}

/* part is supposed to be solid zbuffered (apixbuf==NULL) or transparent zbuffered */
void ISB_create(RenderPart *pa, APixstr *apixbuf)
{
    GroupObject *go;
    
    /* go over all lamps, and make the irregular buffers */
    for(go=R.lights.first; go; go= go->next) {
        LampRen *lar= go->lampren;
        
        if(lar->type==LA_SPOT && lar->shb && lar->buftype==LA_SHADBUF_IRREGULAR) {
            
            /* create storage for shadow, per thread */
            lar->shb->isb_result[pa->thread]= MEM_callocN(sizeof(ISBData), "isb data");
            
            if(apixbuf)
                isb_make_buffer_transp(pa, apixbuf, lar);
            else
                isb_make_buffer(pa, lar);
        }
    }
}


/* end of part rendering, free stored shadow data for this thread from all lamps */
void ISB_free(RenderPart *pa)
{
    GroupObject *go;
    
    /* go over all lamps, and free the irregular buffers */
    for(go=R.lights.first; go; go= go->next) {
        LampRen *lar= go->lampren;
        
        if(lar->type==LA_SPOT && lar->shb && lar->buftype==LA_SHADBUF_IRREGULAR) {
            ISBData *isbdata= lar->shb->isb_result[pa->thread];

            if(isbdata) {
                if(isbdata->shadfacs)
                    MEM_freeN(isbdata->shadfacs);
                if(isbdata->shadfaca)
                    MEM_freeN(isbdata->shadfaca);
                
                if(isbdata->memarena)
                    BLI_memarena_free(isbdata->memarena);
                
                MEM_freeN(isbdata);
                lar->shb->isb_result[pa->thread]= NULL;
            }
        }
    }
}