source: tags/arb-6.0/GDE/RAxML/fastSearch.c

/*  RAxML-VI-HPC (version 2.2) a program for sequential and parallel estimation of phylogenetic trees 
 *  Copyright August 2006 by Alexandros Stamatakis
 *
 *  Partially derived from
 *  fastDNAml, a program for estimation of phylogenetic trees from sequences by Gary J. Olsen
 *  
 *  and 
 *
 *  Programs of the PHYLIP package by Joe Felsenstein.
 *
 *  This program is free software; you may redistribute it and/or modify its
 *  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.
 * 
 *
 *  For any other enquiries send an Email to Alexandros Stamatakis
 *  Alexandros.Stamatakis@epfl.ch
 *
 *  When publishing work that is based on the results from RAxML-VI-HPC please cite:
 *
 *  Alexandros Stamatakis:"RAxML-VI-HPC: maximum likelihood-based phylogenetic analyses with thousands of taxa and mixed models". 
 *  Bioinformatics 2006; doi: 10.1093/bioinformatics/btl446
 */

#ifndef WIN32
#include <sys/times.h>
#include <sys/types.h>
#include <sys/time.h>
#include <unistd.h> 
#endif

#include <math.h>
#include <time.h> 
#include <stdlib.h>
#include <stdio.h>
#include <ctype.h>
#include <string.h>

#include "axml.h"

#define POSTERIOR_THRESHOLD 0.90

extern int Thorough;
extern double masterTime;


extern char permFileName[1024], resultFileName[1024], 
  logFileName[1024], checkpointFileName[1024], infoFileName[1024], run_id[128], workdir[1024], bootStrapFile[1024], bootstrapFileName[1024], 
  bipartitionsFileName[1024],bipartitionsFileNameBranchLabels[1024]; 



typedef struct 
{
  nodeptr p;
  double lh;
} scores;


typedef struct 
{
  scores *s;
  int count;
  int maxCount; 
} insertions;

static void addInsertion(nodeptr p, double lh, insertions *ins)
{ 
  if(ins->count < ins->maxCount)
    {
      ins->s[ins->count].lh = lh;
      ins->s[ins->count].p = p;
      ins->count = ins->count + 1;               
    }
  else
    {        
      ins->s = rax_realloc(ins->s, sizeof(scores) * ins->maxCount * 2, FALSE);     

      ins->maxCount *= 2;

      ins->s[ins->count].lh = lh;
      ins->s[ins->count].p = p;
      ins->count = ins->count + 1; 
    }
}


/* the two functions below just compute the subtree insertion likelihood score into 
   a branch using the thorough method.
*/

static void insertFast (tree *tr, nodeptr p, nodeptr q, int numBranches)
{
  nodeptr  r, s;
  int i;
  
  r = q->back;
  s = p->back;
      
  for(i = 0; i < numBranches; i++)
    tr->lzi[i] = q->z[i];
  
  if(Thorough)
    {
      double  zqr[NUM_BRANCHES], zqs[NUM_BRANCHES], zrs[NUM_BRANCHES], lzqr, lzqs, lzrs, lzsum, lzq, lzr, lzs, lzmax;      
      double defaultArray[NUM_BRANCHES];        
      double e1[NUM_BRANCHES], e2[NUM_BRANCHES], e3[NUM_BRANCHES];
      double *qz;
      
      qz = q->z;
      
      for(i = 0; i < numBranches; i++)
        defaultArray[i] = defaultz;
      
      makenewzGeneric(tr, q, r, qz, iterations, zqr, FALSE);           
      makenewzGeneric(tr, q, s, defaultArray, iterations, zqs, FALSE);                  
      makenewzGeneric(tr, r, s, defaultArray, iterations, zrs, FALSE);
      
      
      for(i = 0; i < numBranches; i++)
        {
          lzqr = (zqr[i] > zmin) ? log(zqr[i]) : log(zmin); 
          lzqs = (zqs[i] > zmin) ? log(zqs[i]) : log(zmin);
          lzrs = (zrs[i] > zmin) ? log(zrs[i]) : log(zmin);
          lzsum = 0.5 * (lzqr + lzqs + lzrs);
          
          lzq = lzsum - lzrs;
          lzr = lzsum - lzqs;
          lzs = lzsum - lzqr;
          lzmax = log(zmax);
          
          if      (lzq > lzmax) {lzq = lzmax; lzr = lzqr; lzs = lzqs;} 
          else if (lzr > lzmax) {lzr = lzmax; lzq = lzqr; lzs = lzrs;}
          else if (lzs > lzmax) {lzs = lzmax; lzq = lzqs; lzr = lzrs;}          
          
          e1[i] = exp(lzq);
          e2[i] = exp(lzr);
          e3[i] = exp(lzs);
        }
      hookup(p->next,       q, e1, numBranches);
      hookup(p->next->next, r, e2, numBranches);
      hookup(p,             s, e3, numBranches);                  
    }
  else
    {       
      double  z[NUM_BRANCHES]; 
      
      for(i = 0; i < numBranches; i++)
        {
          z[i] = sqrt(q->z[i]);      
          
          if(z[i] < zmin) 
            z[i] = zmin;
          if(z[i] > zmax)
            z[i] = zmax;
        }
      
      hookup(p->next,       q, z, tr->numBranches);
      hookup(p->next->next, r, z, tr->numBranches);                              
    }
  
  newviewGeneric(tr, p);
  
  if(Thorough)
    {          
      localSmooth(tr, p, smoothings);   
      
      for(i = 0; i < numBranches; i++)
        {
          tr->lzq[i] = p->next->z[i];
          tr->lzr[i] = p->next->next->z[i];
          tr->lzs[i] = p->z[i];            
        }           
    }
}




static double testInsertFast (tree *tr, nodeptr p, nodeptr q, insertions *ins, boolean veryFast)
{
  double  qz[NUM_BRANCHES], pz[NUM_BRANCHES];
  nodeptr  r, s;
  double LH;
  int i;
  
  r = q->back; 
  
  for(i = 0; i < tr->numBranches; i++)
    {
      qz[i] = q->z[i];
      pz[i] = p->z[i];
    }
   
  insertFast(tr, p, q, tr->numBranches);
      
  evaluateGeneric(tr, p->next->next);   

  addInsertion(q, tr->likelihood, ins);
  
  
  if(veryFast)
    if(tr->likelihood > tr->endLH)
      {                   
        tr->insertNode = q;
        tr->removeNode = p;   
        for(i = 0; i < tr->numBranches; i++)
          tr->currentZQR[i] = tr->zqr[i];      
        tr->endLH = tr->likelihood;                            
      }  
 
  LH = tr->likelihood;                  
              
  hookup(q, r, qz, tr->numBranches);
      
  p->next->next->back = p->next->back = (nodeptr) NULL;
  
  if(Thorough)
    {
      s = p->back;
      hookup(p, s, pz, tr->numBranches);          
    }
      
  return LH;
}

static double testInsertCandidates(tree *tr, nodeptr p, nodeptr q)
{
  double  qz[NUM_BRANCHES], pz[NUM_BRANCHES];
  nodeptr  r, s;
  double LH;
  int i;
  
  r = q->back; 
  
  for(i = 0; i < tr->numBranches; i++)
    {
      qz[i] = q->z[i];
      pz[i] = p->z[i];
    }
   
  insertFast(tr, p, q, tr->numBranches);
      
  evaluateGeneric(tr, p->next->next);       

  if(tr->likelihood > tr->endLH)
    {                     
      tr->insertNode = q;
      tr->removeNode = p;   
      for(i = 0; i < tr->numBranches; i++)
        tr->currentZQR[i] = tr->zqr[i];      
      tr->endLH = tr->likelihood;                            
    }  
 
  LH = tr->likelihood;                  
              
  hookup(q, r, qz, tr->numBranches);
      
  p->next->next->back = p->next->back = (nodeptr) NULL;
  
  if(Thorough)
    {
      s = p->back;
      hookup(p, s, pz, tr->numBranches);          
    }
      
  return LH;
}


/* 
   function below just computes all subtree roots, 
   we go to every inner node and store the three outgoing subtree 
   roots. Assumes evidently that nodeptr p that is passed to the first 
   instance of this recursion is not a tip ;-)
*/

static void getSubtreeRoots(nodeptr p, nodeptr *ptr, int *count, int mxtips)
{
  if(isTip(p->number, mxtips))
    return;
  
  ptr[*count] = p;
  *count = *count + 1;
  
  ptr[*count] = p->next;
  *count = *count + 1;

  ptr[*count] = p->next->next;
  *count = *count + 1;
 
  getSubtreeRoots(p->next->back, ptr, count, mxtips);
  getSubtreeRoots(p->next->next->back, ptr, count, mxtips);
}
  

/*
  conducts linear SPRs, computes the approximate likelihood score 
  fo an insertion of the subtree being rearranged into the left and 
  right branch. Depending on the score it then descends into 
  either the right or the left tree. 
  
  I also added a radius that limits the number of branches away from the original pruning position into which the tree 
  will be inserted.

  We actually need to test empirically what a good setting may be !
*/
  

static void insertBeyond(tree *tr, nodeptr p, nodeptr q, int radius, insertions *ins, boolean veryFast)
{
  if(radius > 0)
    {
      int count = 0;
      
      double 
        twoScores[2];
      
      twoScores[0] = unlikely;
      twoScores[1] = unlikely;
      
      if(isTip(q->next->back->number, tr->mxtips) && isTip(q->next->next->back->number, tr->mxtips))
        return;
      
      if(!isTip(q->next->back->number, tr->mxtips))    
        {
          twoScores[0] = testInsertFast(tr, p, q->next->back, ins, veryFast);    
          count++;
        }
      
      if(!isTip(q->next->next->back->number, tr->mxtips))    
        {
          twoScores[1] = testInsertFast(tr, p, q->next->next->back, ins, veryFast);
          count++;
        }          
      
      if(count == 2 && !veryFast)
        {
          double 
            weight;

          if(twoScores[0] > twoScores[1])           
            {
              weight = exp(twoScores[0] - twoScores[0]) / (exp(twoScores[0] - twoScores[0]) + exp(twoScores[1] - twoScores[0]));
              
              if(weight >= POSTERIOR_THRESHOLD)
                insertBeyond(tr, p, q->next->back, radius - 1, ins, veryFast);
              else
                {
                  insertBeyond(tr, p, q->next->back, radius - 1, ins, veryFast);          
                  insertBeyond(tr, p, q->next->next->back, radius - 1, ins, veryFast);
                }
            }
          else
            {
              weight = exp(twoScores[1] - twoScores[1]) / (exp(twoScores[1] - twoScores[0]) + exp(twoScores[1] - twoScores[1]));
              
              if(weight >= POSTERIOR_THRESHOLD)
                insertBeyond(tr, p, q->next->next->back, radius - 1, ins, veryFast);
              else
                {
                  insertBeyond(tr, p, q->next->back, radius - 1, ins, veryFast);          
                  insertBeyond(tr, p, q->next->next->back, radius - 1, ins, veryFast);
                }
            }
        }
      else
        {
          if(twoScores[0] > twoScores[1])
            insertBeyond(tr, p, q->next->back, radius - 1, ins, veryFast);
          else
            insertBeyond(tr, p, q->next->next->back, radius - 1, ins, veryFast);
        }
    }

}



typedef struct 
{
  int direction;
  double likelihood;

} fourLikelihoods;






static int fourCompare(const void *p1, const void *p2)
{
  fourLikelihoods *rc1 = (fourLikelihoods *)p1;
  fourLikelihoods *rc2 = (fourLikelihoods *)p2;

  double i = rc1->likelihood;
  double j = rc2->likelihood;

  if (i > j)
    return (-1);
  if (i < j)
    return (1);
  return (0);
}

static int scoreCompare(const void *p1, const void *p2)
{
  scores *rc1 = (scores *)p1;
  scores *rc2 = (scores *)p2;

  double i = rc1->lh;
  double j = rc2->lh;

  if (i > j)
    return (-1);
  if (i < j)
    return (1);
  return (0);
}




static double linearSPRs(tree *tr, int radius, boolean veryFast)
{
  int 
    numberOfSubtrees = (tr->mxtips - 2) * 3,
    count = 0,
    k,
    i;

  double 
    fourScores[4];

  nodeptr 
    *ptr = (nodeptr *)rax_malloc(sizeof(nodeptr) * numberOfSubtrees);
  
  fourLikelihoods 
    *fourLi = (fourLikelihoods *)rax_malloc(sizeof(fourLikelihoods) * 4);

  insertions 
    *ins = (insertions*)rax_malloc(sizeof(insertions));

  

  ins->count = 0;
  ins->maxCount = 2048;

  ins->s = (scores *)rax_malloc(sizeof(scores) * ins->maxCount);


  /* recursively compute the roots of all subtrees in the current tree 
     and store them in ptr */

  getSubtreeRoots(tr->start->back, ptr, &count, tr->mxtips);

  assert(count == numberOfSubtrees);
 
  tr->startLH = tr->endLH = tr->likelihood;      

  /* loop over subtrees, i.e., execute a full SPR cycle */

  for(i = 0; i < numberOfSubtrees; i++)
    {           
      nodeptr 
        p = ptr[i],
        p1 = p->next->back,
        p2 = p->next->next->back;
      
      double   
        p1z[NUM_BRANCHES], 
        p2z[NUM_BRANCHES];
        
      ins->count = 0;

      /*printf("Node %d %d\n", p->number, i);*/

      tr->bestOfNode = unlikely;  

      for(k = 0; k < 4; k++)
        fourScores[k] = unlikely;
      
      assert(!isTip(p->number, tr->rdta->numsp));                   
      
      if(!isTip(p1->number, tr->rdta->numsp) || !isTip(p2->number, tr->rdta->numsp))
        {
          double 
            max = unlikely;

          int 
            maxInt = -1;          

          for(k = 0; k < tr->numBranches; k++)
            {
              p1z[k] = p1->z[k];
              p2z[k] = p2->z[k];                   
            }
          
          /* remove the current subtree */

          removeNodeBIG(tr, p,  tr->numBranches);  

          /* pre score with fast insertions */
          if(veryFast)
            Thorough = 1;
          else
            Thorough = 0;

          if (!isTip(p1->number, tr->rdta->numsp)) 
            {
              fourScores[0] = testInsertFast(tr, p, p1->next->back, ins, veryFast);
              fourScores[1] = testInsertFast(tr, p, p1->next->next->back, ins, veryFast);                       
            }
          
          if (!isTip(p2->number, tr->rdta->numsp)) 
            {
              fourScores[2] = testInsertFast(tr, p, p2->next->back, ins, veryFast);
              fourScores[3] = testInsertFast(tr, p, p2->next->next->back, ins, veryFast);                                 
            }
          
          if(veryFast)
            Thorough = 1;
          else
            Thorough = 0;

          /* find the most promising direction */
          

          if(!veryFast)
            {
              int 
                j = 0,
                validEntries = 0;

              double 
                lmax = unlikely,
                posterior = 0.0;          

              for(k = 0; k < 4; k++)
                {
                  fourLi[k].direction = k;
                  fourLi[k].likelihood = fourScores[k];
                }

              qsort(fourLi, 4, sizeof(fourLikelihoods), fourCompare);
              
              for(k = 0; k < 4; k++)
                if(fourLi[k].likelihood > unlikely)
                  validEntries++;
              
              lmax = fourLi[0].likelihood;                           

              while(posterior <= POSTERIOR_THRESHOLD && j < validEntries)         
                {                     
                  double 
                    all = 0.0,
                    prob = 0.0;

                  for(k =  0; k < validEntries; k++)       
                    all += exp(fourLi[k].likelihood - lmax);         
              
                  posterior += (prob = (exp(fourLi[j].likelihood - lmax) / all));                 

                  switch(fourLi[j].direction)
                    { 
                    case 0:
                      insertBeyond(tr, p, p1->next->back, radius, ins, veryFast);
                      break;
                    case 1:
                      insertBeyond(tr, p, p1->next->next->back, radius, ins, veryFast);
                      break;
                    case 2:
                      insertBeyond(tr, p, p2->next->back, radius, ins, veryFast);
                      break;
                    case 3:
                      insertBeyond(tr, p, p2->next->next->back, radius, ins, veryFast);
                      break;
                    default:
                      assert(0);
                    }
                                      
                  j++;
                }                   

              qsort(ins->s, ins->count, sizeof(scores), scoreCompare);       

              Thorough = 1;

              for(k = 0; k < MIN(ins->count, 20); k++)
                testInsertCandidates(tr, p, ins->s[k].p);                     
            }
          else
            {
              Thorough = 1;
              

              for(k = 0; k < 4; k++)
                {
                  if(max < fourScores[k])
                    {
                      max = fourScores[k];
                      maxInt = k;
                    }
                }
              
              /* descend into this direction and re-insert subtree there */
              
              if(maxInt >= 0)
                {
                  switch(maxInt)
                    { 
                    case 0:
                      insertBeyond(tr, p, p1->next->back, radius, ins, veryFast);
                      break;
                    case 1:
                      insertBeyond(tr, p, p1->next->next->back, radius, ins, veryFast);
                      break;
                    case 2:
                      insertBeyond(tr, p, p2->next->back, radius, ins, veryFast);
                      break;
                    case 3:
                      insertBeyond(tr, p, p2->next->next->back, radius, ins, veryFast);
                      break;
                    default:
                      assert(0);
                    }
                }
            }         
          
          /* repair branch and reconnect subtree to its original position from which it was pruned */

          hookup(p->next,       p1, p1z, tr->numBranches); 
          hookup(p->next->next, p2, p2z, tr->numBranches);               
          
          /* repair likelihood vectors */

          newviewGeneric(tr, p);

          /* if the rearrangement of subtree rooted at p yielded a better likelihood score 
             restore the altered topology and use it from now on 
          */

          if(tr->endLH > tr->startLH)                   
            {                                
              restoreTreeFast(tr);               
              tr->startLH = tr->endLH = tr->likelihood;                
            }
                        
        }             
    }
 
  return tr->startLH;     
}




static boolean allSmoothed(tree *tr)
{
  int i;
  boolean result = TRUE;
  
  for(i = 0; i < tr->numBranches; i++)
    {
      if(tr->partitionSmoothed[i] == FALSE)
        result = FALSE;
      else
        tr->partitionConverged[i] = TRUE;
    }

  return result;
}

void nniSmooth(tree *tr, nodeptr p, int maxtimes)
{
  int
    i;

  for(i = 0; i < tr->numBranches; i++)  
    tr->partitionConverged[i] = FALSE;  

 

  while (--maxtimes >= 0) 
    {     
      

      for(i = 0; i < tr->numBranches; i++)      
        tr->partitionSmoothed[i] = TRUE;
      
      

      assert(!isTip(p->number, tr->mxtips));    

     

      assert(!isTip(p->back->number, tr->mxtips));  
      
      update(tr, p);
     
      update(tr, p->next);
     
      update(tr, p->next->next);
      
      update(tr, p->back->next);
      
      update(tr, p->back->next->next);           
     
      if (allSmoothed(tr)) 
        break;
      
    }

  

  for(i = 0; i < tr->numBranches; i++)
    {
      tr->partitionSmoothed[i] = FALSE; 
      tr->partitionConverged[i] = FALSE;
    }

  
}


static void storeBranches(tree *tr, nodeptr p, double *pqz, double *pz1, double *pz2, double *qz1, double *qz2)
{
  int 
    i;
  
  nodeptr 
    q = p->back;

  for(i = 0; i < tr->numBranches; i++)
    {
      pqz[i] = p->z[i];
      pz1[i] = p->next->z[i];
      pz2[i] = p->next->next->z[i];
      qz1[i] = q->next->z[i];
      qz2[i] = q->next->next->z[i];
    }  
}


static int SHSupport(int nPos, int nBootstrap, int *col, double loglk[3], double *siteloglk[3], int lower, int upper, boolean perPartition) 
{
  double
    resampleDelta,
    resample1,
    resample2,  
    support = 0.0,
    delta1 = loglk[0] - loglk[1],
    delta2 = loglk[0] - loglk[2],
    delta = delta1 < delta2 ? delta1 : delta2,
    diff;
  
  int
    iBest,
    i,
    j,
    nSupport = 0,
    iBoot;

  boolean
    shittySplit = FALSE;

  
  if(loglk[1] >= loglk[0])
    {
      diff = ABS(loglk[1] - loglk[0]);
      /*printf("%1.80f %1.80f\n", loglk[0], loglk[1]);*/
      shittySplit = TRUE;
      if(!perPartition)
        assert(diff < 0.1);
    }

  if(loglk[2] >= loglk[0])
    {
      diff = ABS(loglk[2] - loglk[0]);
      /*printf("%1.80f %1.80f\n", loglk[0], loglk[2]);*/
      shittySplit = TRUE; 
      if(!perPartition)
        assert(diff < 0.1);
    }

  if(loglk[0] > loglk[2] && loglk[0] > loglk[1])
    {
      if(loglk[2] > loglk[1])
        diff = ABS(loglk[2] - loglk[0]);
      else
        diff = ABS(loglk[1] - loglk[0]);
      if(diff < 0.1)
        shittySplit = TRUE;
    }
      
  if(shittySplit)
    return 0;
    

  /*
    printf("%f %f %f\n", loglk[0], loglk[1], loglk[2]);
    assert(loglk[0] >= loglk[1] && loglk[0] >= loglk[2]);
  */
 
  for(iBoot = 0; iBoot < nBootstrap; iBoot++) 
    {
      double resampled[3];
      
      for (i = 0; i < 3; i++)
        resampled[i] = -loglk[i];
      
      for (j = lower; j < upper; j++) 
        {
          int 
            pos = col[iBoot * nPos + j];
          
          for (i = 0; i < 3; i++)
            resampled[i] += pos * siteloglk[i][j];
        }

      iBest = 0;
      
      /*printf("%d %f %f %f\n", iBoot, resampled[0], resampled[1], resampled[2]);*/

      for (i = 1; i < 3; i++)
        if (resampled[i] > resampled[iBest])
          iBest = i;
      
      resample1 = resampled[iBest] - resampled[(iBest+1)%3];
      resample2 = resampled[iBest] - resampled[(iBest+2)%3];
      resampleDelta = resample1 < resample2 ? resample1 : resample2;
      
      if(resampleDelta < delta)
        nSupport++;
    }
  
  support = (nSupport/(double)nBootstrap);
    
  return ((int)((support * 100.0) + 0.5));
}

static void setupBranchInfo(nodeptr p, tree *tr, int *counter)
{
  if(!isTip(p->number, tr->mxtips))       
    {
      nodeptr q;          

      if(!(isTip(p->back->number, tr->mxtips)))
        {                 
          p->bInf = p->back->bInf = &(tr->bInf[*counter]);                                                

          p->bInf->oP = p;
          p->bInf->oQ = p->back;
          
          *counter = *counter + 1;
        }
      
      q = p->next;

      while(q != p)
        {
          setupBranchInfo(q->back, tr, counter);        
          q = q->next;
        }
        
      return;
    }
}




static void doNNIs(tree *tr, nodeptr p, double *lhVectors[3], boolean shSupport, int *interchanges, int *innerBranches,
                   double *pqz_0, double *pz1_0, double *pz2_0, double *qz1_0, double *qz2_0, double *pqz_1, double *pz1_1, double *pz2_1,
                   double *qz1_1, double *qz2_1, double *pqz_2, double *pz1_2, double *pz2_2, double *qz1_2, double *qz2_2)
{     
  nodeptr 
    q = p->back,     
    pb1 = p->next->back,
    pb2 = p->next->next->back;

  assert(!isTip(p->number, tr->mxtips));     
  
  if(!isTip(q->number, tr->mxtips))
    {   
      int 
        model,
        whichNNI = 0;
      
      nodeptr    
        qb1 = q->next->back,
        qb2 = q->next->next->back;  
      
      double             
        lh[3],
        *lhv[3];         

      lhv[0] = (double*)rax_malloc(sizeof(double) * tr->NumberOfModels);
      lhv[1] = (double*)rax_malloc(sizeof(double) * tr->NumberOfModels);
      lhv[2] = (double*)rax_malloc(sizeof(double) * tr->NumberOfModels);

      *innerBranches = *innerBranches + 1;
          
      nniSmooth(tr, p, 16);     
      
      if(shSupport)
        {          
          evaluateGenericVector(tr, p);
          memcpy(lhVectors[0], tr->perSiteLL, sizeof(double) * tr->cdta->endsite);
        }
      else
        evaluateGeneric(tr, p);
      
      lh[0] = tr->likelihood;

      for(model = 0; model < tr->NumberOfModels; model++)
        lhv[0][model] = tr->perPartitionLH[model];
            
      storeBranches(tr, p, pqz_0, pz1_0, pz2_0, qz1_0, qz2_0);
                
      /*******************************************/
      
      hookup(p, q, pqz_0, tr->numBranches); 
      
      hookup(p->next,       qb1, qz1_0, tr->numBranches); 
      hookup(p->next->next, pb2, pz2_0, tr->numBranches); 
      
      hookup(q->next,       pb1, pz1_0, tr->numBranches);       
      hookup(q->next->next, qb2, qz2_0, tr->numBranches); 
      
      newviewGeneric(tr, p);
      newviewGeneric(tr, p->back);
            
      nniSmooth(tr, p, 16);
      
      if(shSupport)
        {
          evaluateGenericVector(tr, p);
          memcpy(lhVectors[1], tr->perSiteLL, sizeof(double) * tr->cdta->endsite);
        }
      else
        evaluateGeneric(tr, p);
      
      lh[1] = tr->likelihood;   
        
      for(model = 0; model < tr->NumberOfModels; model++)
        lhv[1][model] = tr->perPartitionLH[model];
      
      storeBranches(tr, p, pqz_1, pz1_1, pz2_1, qz1_1, qz2_1);
      
      if(lh[1] > lh[0])
        whichNNI = 1;
      
      /*******************************************/
      
      hookup(p, q, pqz_0, tr->numBranches); 
      
      hookup(p->next,       qb1, qz1_0, tr->numBranches); 
      hookup(p->next->next, pb1, pz1_0, tr->numBranches); 
        
      hookup(q->next,       pb2, pz2_0, tr->numBranches);               
      hookup(q->next->next, qb2, qz2_0, tr->numBranches); 
      
      newviewGeneric(tr, p);
      newviewGeneric(tr, p->back);
           
      nniSmooth(tr, p, 16);
      
      if(shSupport)
        {
          evaluateGenericVector(tr, p);
          memcpy(lhVectors[2], tr->perSiteLL, sizeof(double) * tr->cdta->endsite);
        }
      else
        evaluateGeneric(tr, p);
      
      lh[2] = tr->likelihood;            
      
      for(model = 0; model < tr->NumberOfModels; model++)
        lhv[2][model] = tr->perPartitionLH[model];

      storeBranches(tr, p, pqz_2, pz1_2, pz2_2, qz1_2, qz2_2);   
      
      if(lh[2] > lh[0] && lh[2] > lh[1])
        whichNNI = 2;
      
      /*******************************************/
      
      if(shSupport)
        whichNNI = 0;

      switch(whichNNI)
        {
        case 0:  
          hookup(p, q, pqz_0, tr->numBranches);           
          
          hookup(p->next,       pb1, pz1_0, tr->numBranches); 
          hookup(p->next->next, pb2, pz2_0, tr->numBranches); 
          
          hookup(q->next,       qb1, qz1_0, tr->numBranches);   
          hookup(q->next->next, qb2, qz2_0, tr->numBranches);
          break;
        case 1:  
          hookup(p, q, pqz_1, tr->numBranches);             
          
          hookup(p->next,       qb1, pz1_1, tr->numBranches); 
          hookup(p->next->next, pb2, pz2_1, tr->numBranches); 
          
          hookup(q->next,       pb1, qz1_1, tr->numBranches);   
          hookup(q->next->next, qb2, qz2_1, tr->numBranches); 
          break;
        case 2:   
          hookup(p, q, pqz_2, tr->numBranches); 
          
          hookup(p->next,       qb1, pz1_2, tr->numBranches); 
          hookup(p->next->next, pb1, pz2_2, tr->numBranches); 
          
          hookup(q->next,       pb2, qz1_2, tr->numBranches);           
          hookup(q->next->next, qb2, qz2_2, tr->numBranches); 
          break;
        default:
          assert(0);
        }       
      
      newviewGeneric(tr, p);
      newviewGeneric(tr, q);     
                 
      if(whichNNI > 0)
        *interchanges = *interchanges + 1;
      
      if(shSupport)       
        {        
          p->bInf->support = SHSupport(tr->cdta->endsite, 1000, tr->resample, lh, lhVectors, 0, tr->cdta->endsite, FALSE);         

          //printf("ALL: %f %f %f ", lh[0], lh[1], lh[2]);

          for(model = 0; model < tr->NumberOfModels; model++)
            {        
              double
                perPartitionLH[3];

              perPartitionLH[0] = lhv[0][model];
              perPartitionLH[1] = lhv[1][model];
              perPartitionLH[2] = lhv[2][model];

              p->bInf->supports[model] = SHSupport(tr->cdta->endsite, 1000, tr->resample, perPartitionLH, lhVectors, tr->partitionData[model].lower, tr->partitionData[model].upper, TRUE);

              //printf(" p%d %f %f %f ", model, perPartitionLH[0], perPartitionLH[1], perPartitionLH[2]);
            }

          //printf("\n");       
        }

      rax_free(lhv[0]);
      rax_free(lhv[1]);
      rax_free(lhv[2]);        
    }     

  
  if(!isTip(pb1->number, tr->mxtips))
    doNNIs(tr, pb1, lhVectors, shSupport, interchanges, innerBranches,
           pqz_0, pz1_0, pz2_0, qz1_0, qz2_0, pqz_1, pz1_1, pz2_1,
           qz1_1, qz2_1, pqz_2, pz1_2, pz2_2, qz1_2, qz2_2);

  if(!isTip(pb2->number, tr->mxtips))
    doNNIs(tr, pb2, lhVectors, shSupport, interchanges,  innerBranches,
           pqz_0, pz1_0, pz2_0, qz1_0, qz2_0, pqz_1, pz1_1, pz2_1,
           qz1_1, qz2_1, pqz_2, pz1_2, pz2_2, qz1_2, qz2_2);             
  

  return;
}


static int encapsulateNNIs(tree *tr, double *lhVectors[3], boolean shSupport)
{
  int 
    innerBranches = 0,
    interchanges  = 0;

  double
    pqz_0[NUM_BRANCHES],
    pz1_0[NUM_BRANCHES],
    pz2_0[NUM_BRANCHES],
    qz1_0[NUM_BRANCHES],
    qz2_0[NUM_BRANCHES],
    pqz_1[NUM_BRANCHES],
    pz1_1[NUM_BRANCHES],
    pz2_1[NUM_BRANCHES],
    qz1_1[NUM_BRANCHES],
    qz2_1[NUM_BRANCHES],
    pqz_2[NUM_BRANCHES],
    pz1_2[NUM_BRANCHES],
    pz2_2[NUM_BRANCHES],
    qz1_2[NUM_BRANCHES],
    qz2_2[NUM_BRANCHES];   

  
  doNNIs(tr, tr->start->back, lhVectors, shSupport, &interchanges, &innerBranches,
         pqz_0, pz1_0, pz2_0, qz1_0, qz2_0, pqz_1, pz1_1, pz2_1,
         qz1_1, qz2_1, pqz_2, pz1_2, pz2_2, qz1_2, qz2_2);

  assert(innerBranches == (tr->mxtips - 3));

  return interchanges;
}

int *permutationSH(tree *tr, int nBootstrap, long _randomSeed) 
{
  int 
    replicate,
    model,
    maxNonZero = 0,
    *weightBuffer,   
    *col = (int*)rax_calloc(((size_t)tr->cdta->endsite) * ((size_t)nBootstrap), sizeof(int)),
    *nonzero = (int*)rax_calloc(tr->NumberOfModels, sizeof(int));
  
  long 
    randomSeed = _randomSeed;
  
  size_t 
    bufferSize;
  
  for(model = 0; model < tr->NumberOfModels; model++)
    { 
      int 
        j;
                       
      for (j = 0; j < tr->cdta->endsite; j++)  
        {
          if(tr->originalModel[j] == model)
            nonzero[model] += tr->originalWeights[j];
        }               
      
      if(nonzero[model] > maxNonZero)
        maxNonZero = nonzero[model];      
    }
  
  bufferSize = ((size_t)maxNonZero) * sizeof(int);
  weightBuffer = (int*)rax_malloc(bufferSize);

  for(replicate = 0; replicate < nBootstrap; replicate++)
    {      
      int
        j,       
        *wgt = &col[((size_t)tr->cdta->endsite) * ((size_t)replicate)];

      for(model = 0; model < tr->NumberOfModels; model++)
        {
          int
            pos,            
            nonz = nonzero[model];
     
          memset(weightBuffer, 0, bufferSize);

          for(j = 0; j < nonz; j++)
            weightBuffer[(int) (nonz * randum(&randomSeed))]++;          

          for(j = 0, pos = 0; j < tr->cdta->endsite; j++) 
            {
              if(model == tr->originalModel[j])
                {
                  int 
                    w;
                  
                  for(w = 0; w < tr->originalWeights[j]; w++)                    
                    {
                      wgt[j] += weightBuffer[pos];
                      pos++;                  
                    }                              
                }
            }     
        }
    }


  rax_free(weightBuffer);
  rax_free(nonzero);

  return col;
}







void fastSearch(tree *tr, analdef *adef, rawdata *rdta, cruncheddata *cdta)
{
  double    
    likelihood, 
    startLikelihood,
    *lhVectors[3];

  char 
    bestTreeFileName[1024];
  
  FILE 
    *f;

  int
    model;

 
  
  lhVectors[0] = (double *)NULL;
  lhVectors[1] = (double *)NULL;
  lhVectors[2] = (double *)NULL;  
      
  /* initialize model parameters with standard starting values */

  initModel(tr, rdta, cdta, adef);      

  printBothOpen("Time after init : %f\n", gettime() - masterTime);

  /* 
     compute starting tree, either by reading in a tree specified via -t 
     or by building one 
  */

  getStartingTree(tr, adef);
 
  printBothOpen("Time after init and starting tree: %f\n", gettime() - masterTime);
  
  /* 
     rough model parameter optimization, the log likelihood epsilon should 
     actually be determined based on the initial tree score and not be hard-coded 
  */
  
 if(adef->useBinaryModelFile)
    {
      readBinaryModel(tr);
      evaluateGenericInitrav(tr, tr->start);
      treeEvaluate(tr, 2);
    }
 else
   modOpt(tr, adef, FALSE, 10.0);
  
  printBothOpen("Time after init, starting tree, mod opt: %f\n", gettime() - masterTime);

  /* print out the number of rate categories used for the CAT model, one should 
     use less then the default, e.g., -c 16 works quite well */

  for(model = 0; model < tr->NumberOfModels; model++)
    printBothOpen("Partion %d number of Cats: %d\n", model, tr->partitionData[model].numberOfCategories);

  /* 
     means that we are going to do thorough insertions 
     with real newton-raphson based br-len opt at the three branches 
     adjactent to every insertion point 
  */

  Thorough = 1;

  
  /*
    loop over SPR cycles until the likelihood difference 
     before and after the SPR cycle is <= 0.5 log likelihood units.
     Rather than being hard-coded this should also be determined based on the 
     actual likelihood of the tree 
  */
 
  do
    {      
      startLikelihood = tr->likelihood;
   
      /* conduct a cycle of linear SPRs */

    

      likelihood = linearSPRs(tr, 20, adef->veryFast);          
           
      evaluateGeneric(tr, tr->start); 
      
      /* the NNIs also optimize br-lens of resulting topology a bit */
      encapsulateNNIs(tr, lhVectors, FALSE);                    
 
      printBothOpen("LH after SPRs %f, after NNI %f\n", likelihood, tr->likelihood);
    }
  while(ABS(tr->likelihood - startLikelihood) > 0.5);

 
  
 

  
  /* print out the resulting tree to the RAxML_bestTree. file. 
     note that boosttrapping or doing multiple inferences won't work.
     This thing computes a single tree and that's it */
      
  strcpy(bestTreeFileName, workdir); 
  strcat(bestTreeFileName, "RAxML_fastTree.");
  strcat(bestTreeFileName,         run_id);

 

  Tree2String(tr->tree_string, tr, tr->start->back, FALSE, TRUE, FALSE, FALSE, FALSE, adef, SUMMARIZE_LH, FALSE, FALSE, FALSE, FALSE);
    
  f = myfopen(bestTreeFileName, "wb");
  fprintf(f, "%s", tr->tree_string);
  fclose(f);  

 
    
  printBothOpen("RAxML fast tree written to file: %s\n", bestTreeFileName);
  
  writeBinaryModel(tr);
  
  printBothOpen("Total execution time: %f\n", gettime() - masterTime);

  printBothOpen("Good bye ... \n");
}


void shSupports(tree *tr, analdef *adef, rawdata *rdta, cruncheddata *cdta)
{
  double 
    diff,
    *lhVectors[3];

  char 
    bestTreeFileName[1024],
    shSupportFileName[1024];
  
  FILE 
    *f;

  int
    i,
    interchanges = 0,
    counter = 0;

  assert(adef->restart);
    
  tr->resample = permutationSH(tr, 1000, 12345);
  
  //a tiny bit dirty here, all allocs should be cleaned up!
  
  lhVectors[0] = (double *)rax_malloc(sizeof(double) * tr->cdta->endsite);
  lhVectors[1] = (double *)rax_malloc(sizeof(double) * tr->cdta->endsite);
  lhVectors[2] = (double *)rax_malloc(sizeof(double) * tr->cdta->endsite);
  tr->bInf = (branchInfo*)rax_malloc(sizeof(branchInfo) * (tr->mxtips - 3)); 
  
  for(i = 0; i < tr->mxtips - 3; i++)
    tr->bInf[i].supports = (int*)rax_malloc(sizeof(int) * tr->NumberOfModels);

  initModel(tr, rdta, cdta, adef);        
 
 
  getStartingTree(tr, adef);
  
  if(adef->useBinaryModelFile)
    {
      readBinaryModel(tr);
      evaluateGenericInitrav(tr, tr->start);
      treeEvaluate(tr, 2);
    }
  else
    {
      evaluateGenericInitrav(tr, tr->start);
      modOpt(tr, adef, FALSE, 1.0);
    }
  
  printBothOpen("Time after model optimization: %f\n", gettime() - masterTime);
  
  printBothOpen("Initial Likelihood %f\n\n", tr->likelihood);   
    
  do
    {   
      double 
        lh1,
        lh2;
   
      lh1 = tr->likelihood;

      interchanges = encapsulateNNIs(tr, lhVectors, FALSE);       

      evaluateGeneric(tr, tr->start);           

      lh2 = tr->likelihood;

      diff = ABS(lh1 - lh2);

      printBothOpen("NNI interchanges %d Likelihood %f\n", interchanges, tr->likelihood);
    }
  while(diff > 0.01);

  printBothOpen("\nFinal Likelihood of NNI-optimized tree: %f\n\n", tr->likelihood);

  setupBranchInfo(tr->start->back, tr, &counter);
  assert(counter == tr->mxtips - 3);
 
  interchanges = encapsulateNNIs(tr, lhVectors, TRUE);
              
  strcpy(bestTreeFileName, workdir); 
  strcat(bestTreeFileName, "RAxML_fastTree.");
  strcat(bestTreeFileName,         run_id); 

  Tree2String(tr->tree_string, tr, tr->start->back, FALSE, TRUE, FALSE, FALSE, FALSE, adef, SUMMARIZE_LH, FALSE, FALSE, FALSE, FALSE);
    
  f = myfopen(bestTreeFileName, "wb");
  fprintf(f, "%s", tr->tree_string);
  fclose(f);  

  
  strcpy(shSupportFileName, workdir); 
  strcat(shSupportFileName, "RAxML_fastTreeSH_Support.");
  strcat(shSupportFileName,         run_id);
  
  Tree2String(tr->tree_string, tr, tr->start->back, TRUE, TRUE, FALSE, FALSE, FALSE, adef, SUMMARIZE_LH, FALSE, TRUE, FALSE, FALSE);
  
  f = myfopen(shSupportFileName, "wb");
  fprintf(f, "%s", tr->tree_string);
  fclose(f);  
      
  printBothOpen("RAxML NNI-optimized tree written to file: %s\n", bestTreeFileName);
  
  printBothOpen("\nSame tree with SH-like supports written to file: %s\n", shSupportFileName);

  if(tr->NumberOfModels > 1)
    {
      char     
        shSupportPerPartitionFileName[1024];
      
      strcpy(shSupportPerPartitionFileName, workdir); 
      strcat(shSupportPerPartitionFileName, "RAxML_fastTree_perPartition_SH_Support.");
      strcat(shSupportPerPartitionFileName,         run_id);
  
      Tree2String(tr->tree_string, tr, tr->start->back, TRUE, TRUE, FALSE, FALSE, FALSE, adef, SUMMARIZE_LH, FALSE, FALSE, FALSE, TRUE);
      
      f = myfopen(shSupportPerPartitionFileName, "wb");
      fprintf(f, "%s", tr->tree_string);
      fclose(f);  
      
      printBothOpen("\nSame tree with SH-like support for each partition written to file: %s\n", shSupportPerPartitionFileName);
    }
  
  printBothOpen("\nTotal execution time: %f\n", gettime() - masterTime);

  exit(0);
}