irstlm/src/lmtable.cpp

Go to the documentation of this file.

// $Id: lmtable.cpp 390 2010-06-11 06:32:48Z nicolabertoldi $

/******************************************************************************
IrstLM: IRST Language Model Toolkit
Copyright (C) 2006 Marcello Federico, ITC-irst Trento, Italy

This library is free software; you can redistribute it and/or
modify it under the terms of the GNU Lesser General Public
License as published by the Free Software Foundation; either
version 2.1 of the License, or (at your option) any later version.

This library 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
Lesser General Public License for more details.

You should have received a copy of the GNU Lesser General Public
License along with this library; if not, write to the Free Software
Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301 USA

******************************************************************************/
#include <stdio.h>
#include <cstdlib>
#include <stdlib.h>
#include <fcntl.h>
#include <iostream>
#include <fstream>
#include <stdexcept>
#include <string>
#include <set>
#include <cassert>
#include <limits>
#include "math.h"
#include "mempool.h"
#include "htable.h"
#include "ngramcache.h"
#include "dictionary.h"
#include "n_gram.h"
#include "lmtable.h"

#include "util.h"

#define DEBUG 0

//special value for pruned iprobs
#define NOPROB ((float)-1.329227995784915872903807060280344576e36)

using namespace std;

inline void error(const char* message){
  std::cerr << message << "\n";
  throw std::runtime_error(message);
}


//instantiate an empty lm table
lmtable::lmtable(){
        
  configure(1,false);
        
  dict=new dictionary((char *)NULL,1000000);
        
  memset(cursize, 0, sizeof(cursize));
  memset(tbltype, 0, sizeof(tbltype));
  memset(maxsize, 0, sizeof(maxsize));
  memset(info, 0, sizeof(info));
  memset(NumCenters, 0, sizeof(NumCenters));
        
  max_cache_lev=0;
  for (int i=0;i<=LMTMAXLEV+1;i++) lmtcache[i]=NULL;
        
  probcache=NULL;
  statecache=NULL;
  statesizecache=NULL;
  
  memmap=0;
  
  isPruned=false;
  isInverted=false;
        
  //statistics
  for (int i=0;i<=LMTMAXLEV+1;i++) totget[i]=totbsearch[i]=0;
        
  logOOVpenalty=0.0; //penalty for OOV words (default 0)
        
  // by default, it is a standard LM, i.e. queried for score
  setOrderQuery(false);
};

void lmtable::init_probcache(){
  assert(probcache==NULL);
  probcache=new ngramcache(maxlev,sizeof(double),400000);
#ifdef TRACE_CACHE
  cacheout=new std::fstream(get_temp_folder()++"tracecache",std::ios::out);
  sentence_id=0;
#endif
}

void lmtable::init_statecache(){
  assert(statecache==NULL);
  if(maxlev > 1){  // we don't need a state cache for a unigram model
    statecache=new ngramcache(maxlev-1,sizeof(char *),200000);
    statesizecache=new ngramcache(maxlev-1,sizeof(int),200000);
  }
}

void lmtable::init_lmtcaches(int uptolev){
  max_cache_lev=uptolev;
  for (int i=2;i<=max_cache_lev;i++){
    assert(lmtcache[i]==NULL);
    lmtcache[i]=new ngramcache(i,sizeof(char *),200000);
  }
}

void lmtable::check_cache_levels(){
  if (probcache && probcache->isfull()) probcache->reset(probcache->cursize());
  if (statecache && statecache->isfull()) { 
    statecache->reset(statecache->cursize());
    statesizecache->reset(statesizecache->cursize());
  }
  for (int i=2;i<=max_cache_lev;i++)
    if (lmtcache[i]->isfull()) lmtcache[i]->reset(lmtcache[i]->cursize());
}

void lmtable::reset_caches(){
  if (probcache) probcache->reset(MAX(probcache->cursize(),probcache->maxsize()));
  if (statecache){
    statecache->reset(MAX(statecache->cursize(),statecache->maxsize()));
    statesizecache->reset(MAX(statesizecache->cursize(),statesizecache->maxsize()));
  }
  for (int i=2;i<=max_cache_lev;i++)
    lmtcache[i]->reset(MAX(lmtcache[i]->cursize(),lmtcache[i]->maxsize()));
}

void lmtable::configure(int n,bool quantized){
  maxlev=n;
  if (n==1)
    tbltype[1]=(quantized?QLEAF:LEAF);
  else{
    for (int i=1;i<n;i++) tbltype[i]=(quantized?QINTERNAL:INTERNAL);
    tbltype[n]=(quantized?QLEAF:LEAF);
  }
}



//loadstd::istream& inp a lmtable from a lm file

void lmtable::load(istream& inp,const char* filename,const char* outfilename,int keep_on_disk,OUTFILE_TYPE outtype){
        
#ifdef WIN32
  if (keep_on_disk>0){
    std::cerr << "lmtable::load memory mapping not yet available under WIN32\n";
    keep_on_disk = 0;
  }
#endif
        
  //give a look at the header to select loading method
  char header[MAX_LINE];
  inp >> header; cerr << header << "\n";
        
  if (strncmp(header,"Qblmt",5)==0 || strncmp(header,"blmt",4)==0){
    loadbin(inp,header,filename,keep_on_disk);
  }
  else{ //input is in textual form
    
          if (keep_on_disk && outfilename==NULL) {
      cerr << "Load Error: inconsistent setting. Passed input file: textual. Memory map: yes. Outfilename: not specified.\n";
      exit(0);
    }
                
    loadtxt(inp,header,outfilename,keep_on_disk);
  }
        
  cerr << "OOV code is " << lmtable::getDict()->oovcode() << "\n";
}


//load language model on demand through a word-list file

int lmtable::reload(std::set<string> words){
        
        
        //build dictionary
        dictionary dict(NULL,(int)words.size()); dict.incflag(1);
        
        std::set<string>::iterator w;
        for (w = words.begin(); w != words.end(); ++w);
                dict.encode((*w).c_str());
                
                
                
                return 1;       
}

int parseWords(char *sentence, const char **words, int max)
{
  char *word;
  int i = 0;
        
  const char *const wordSeparators = " \t\r\n";
        
  for (word = strtok(sentence, wordSeparators);
       i < max && word != 0;
       i++, word = strtok(0, wordSeparators))
        {
                words[i] = word;
        }
        
  if (i < max){words[i] = 0;}
        
  return i;
}



//Load a LM as a text file. LM could have been generated either with the
//IRST LM toolkit or with the SRILM Toolkit. In the latter we are not
//sure that n-grams are lexically ordered (according to the 1-grams).
//However, we make the following assumption:
//"all successors of any prefix are sorted and written in contiguous lines!"
//This method also loads files processed with the quantization
//tool: qlm

int parseline(istream& inp, int Order,ngram& ng,float& prob,float& bow){
        
  const char* words[1+ LMTMAXLEV + 1 + 1];
  int howmany;
  char line[MAX_LINE];
        
  inp.getline(line,MAX_LINE);
  if (strlen(line)==MAX_LINE-1){
    cerr << "parseline: input line exceed MAXLINE ("
                << MAX_LINE << ") chars " << line << "\n";
    exit(1);
  }
        
  howmany = parseWords(line, words, Order + 3);
        
  if (!(howmany == (Order+ 1) || howmany == (Order + 2)))
                assert(howmany == (Order+ 1) || howmany == (Order + 2));
        
  //read words
  ng.size=0;
  for (int i=1;i<=Order;i++)
    ng.pushw(strcmp(words[i],"<unk>")?words[i]:ng.dict->OOV());
        
  //read logprob/code and logbow/code
  assert(sscanf(words[0],"%f",&prob));
  if (howmany==(Order+2))
    assert(sscanf(words[Order+1],"%f",&bow));
  else
    bow=0.0; //this is log10prob=0 for implicit backoff
        
  return 1;
}


void lmtable::loadcenters(istream& inp,int Order){
  char line[MAX_LINE];
        
  //first read the coodebook
  cerr << Order << " read code book ";
  inp >> NumCenters[Order];
  Pcenters[Order]=new float[NumCenters[Order]];
  Bcenters[Order]=(Order<maxlev?new float[NumCenters[Order]]:NULL);
        
  for (int c=0;c<NumCenters[Order];c++){
    inp >> Pcenters[Order][c];
    if (Order<maxlev) inp >> Bcenters[Order][c];
  };
  //empty the last line
  inp.getline((char*)line,MAX_LINE);
}

void lmtable::loadtxt(istream& inp,const char* header,const char* outfilename,int mmap){
  if (mmap>0)
    loadtxtmmap(inp,header,outfilename);
  else {
    loadtxt(inp,header);
    lmtable::getDict()->genoovcode();
  }
}

void lmtable::loadtxtmmap(istream& inp,const char* header,const char* outfilename){
        
        char nameNgrams[BUFSIZ];
        char nameHeader[BUFSIZ];
        
        FILE *fd = NULL;
        table_pos_t filesize=0;
        
        int Order,n;
        
        int maxlevel_h;
        //char *SepString = " \t\n"; unused
        
        //open input stream and prepare an input string
        char line[MAX_LINE];
        
        //prepare word dictionary
        //dict=(dictionary*) new dictionary(NULL,1000000,NULL,NULL);
        lmtable::getDict()->incflag(1);
        
        //put here ngrams, log10 probabilities or their codes
        ngram ng(lmtable::getDict());
        ngram ing(lmtable::getDict());
        
        float pb,bow;;
        
        //check the header to decide if the LM is quantized or not
        isQtable=(strncmp(header,"qARPA",5)==0?true:false);
        
        //check the header to decide if the LM table is incomplete
        isItable=(strncmp(header,"iARPA",5)==0?true:false);
        
        if (isQtable){
                //check if header contains other infos
                inp >> line;
                if (!(maxlevel_h=atoi(line))){
                        cerr << "loadtxt with mmap requires new qARPA header. Please regenerate the file.\n";
                        exit(1);
                }
                
                for (n=1;n<=maxlevel_h;n++){
                        inp >> line;
                        if (!(NumCenters[n]=atoi(line))){
                                cerr << "loadtxt with mmap requires new qARPA header. Please regenerate the file.\n";
                                exit(0);
                        }
                }
        }
        
        //we will configure the table later we we know the maxlev;
        bool yetconfigured=false;
        
        cerr << "loadtxtmmap()\n";
        
        // READ ARPA Header
        
        while (inp.getline(line,MAX_LINE)){
                
                if (strlen(line)==MAX_LINE-1){
                        cerr << "lmtable::loadtxtmmap: input line exceed MAXLINE ("
                        << MAX_LINE << ") chars " << line << "\n";
                        exit(1);
                }
                
                bool backslash = (line[0] == '\\');
                
                if (sscanf(line, "ngram %d=%d", &Order, &n) == 2) {
                        maxsize[Order] = n; maxlev=Order; //upadte Order
                        cerr << "size[" << Order << "]=" << maxsize[Order] << "\n";
                }
                
                if (backslash && sscanf(line, "\\%d-grams", &Order) == 1) {
                        
                        //at this point we are sure about the size of the LM
                        if (!yetconfigured){
                                configure(maxlev,isQtable);
                                yetconfigured=true;
                                
                                //opening output file
                                strcpy(nameNgrams,outfilename);
                                strcat(nameNgrams, "-ngrams");
                                
                                cerr << "saving ngrams probs in " << nameNgrams << "\n";
                                
                                fd = fopen(nameNgrams, "w+");
                                
                                // compute the size of file (only for tables and - possibly - centroids; no header nor dictionary)
                                for (int l=1;l<=maxlev;l++){
                                        if (l<maxlev)
                                                filesize +=  (table_pos_t) maxsize[l] * nodesize(tbltype[l]) + 2 * NumCenters[l] * sizeof(float);
                                        else
                                                filesize +=  (table_pos_t) maxsize[l] * nodesize(tbltype[l]) + NumCenters[l] * sizeof(float);
                                }
                                
                                cerr << "global filesize = " << filesize << "\n";
                                // set the file to the proper size:
                                ftruncate(fileno(fd),filesize);
                                table[0]=(char *)(MMap(fileno(fd),PROT_READ|PROT_WRITE,0,filesize,&tableGaps[0]));
                                
                                //allocate space for tables into the file through mmap:
                                
                                if (maxlev>1)
                                        table[1]=table[0] + (table_pos_t) (2 * NumCenters[1] * sizeof(float));
                                else
                                        table[1]=table[0] + (table_pos_t) (NumCenters[1] * sizeof(float));
                                
                                for (int l=2;l<=maxlev;l++)
                                        if (l<maxlev)
                                                table[l]=(char *)(table[l-1] + (table_pos_t) maxsize[l-1]*nodesize(tbltype[l-1]) +
                                                                                                                        2 * NumCenters[l] * sizeof(float));
                                        else
                                                table[l]=(char *)(table[l-1] + (table_pos_t) maxsize[l-1]*nodesize(tbltype[l-1]) +
                                                                                                                        NumCenters[l] * sizeof(float));
                                
                                for (int l=2;l<=maxlev;l++){
                                        cerr << "table[" << l << "]-table[" << l-1 << "]="
                                        << (table_pos_t) table[l]-(table_pos_t) table[l-1] << " (nodesize=" << nodesize(tbltype[l-1]) << ")\n";
                                }
                        }
                        
                        cerr << Order << "-grams: reading ";
                        if (isQtable) {
                                loadcenters(inp,Order);
                                // writing centroids on disk
                                if (Order<maxlev){
                                        memcpy(table[Order] - 2 * NumCenters[Order] * sizeof(float),
                                                                 Pcenters[Order],
                                                                 NumCenters[Order] * sizeof(float));
                                        memcpy(table[Order] - NumCenters[Order] * sizeof(float),
                                                                 Bcenters[Order],
                                                                 NumCenters[Order] * sizeof(float));
                                } else
                                        memcpy(table[Order] - NumCenters[Order] * sizeof(float),
                                                                 Pcenters[Order],
                                                                 NumCenters[Order] * sizeof(float));
                        }
                        
                        //allocate support vector to manage badly ordered n-grams
                        if (maxlev>1 && Order<maxlev) {
                                startpos[Order]=new table_entry_pos_t[maxsize[Order]];
                                for (table_entry_pos_t c=0;c<maxsize[Order];c++) startpos[Order][c]=BOUND_EMPTY1;
                        }
                        //prepare to read the n-grams entries
                        cerr << maxsize[Order] << " entries\n";
                        
                        //WE ASSUME A WELL STRUCTURED FILE!!!
                        for (table_entry_pos_t c=0;c<maxsize[Order];c++){
                                
                                if (parseline(inp,Order,ng,pb,bow)){
                                        //if table is in incomplete ARPA format pb is just the
                                        //discounted frequency, so we need to add bow * Pr(n-1 gram)
                                        
                                        // if table is inverted then revert n-gram
                                        if (isInverted & Order>1){
                                                ing.invert(ng);
                                                ng=ing;
                                        }
                                        
                                        if (isItable && Order>1){
                                                //get bow of lower of context
                                                get(ng,ng.size,ng.size-1);
                                                float rbow=0.0;
                                                if (ng.lev==ng.size-1){ //found context
                                                        rbow=ng.bow;
                                                        //  int ibow=ng.bow; rbow=*((float *)&ibow);
                                                }
                                                
                                                int tmp=maxlev;
                                                maxlev=Order-1;
                                                pb= log(exp((double)pb * M_LN10) +  exp(((double)rbow + lprob(ng)) * M_LN10))/M_LN10;
                                                maxlev=tmp;
                                        }
                                        
                                        if (isQtable) add(ng, (qfloat_t)pb, (qfloat_t)bow);
                                        else add(ng, pb, bow);
                                        
                                }
                        }
                        // To avoid huge memory write concentrated at the end of the program
                        msync(table[0],filesize,MS_SYNC);
                        
                        // now we can fix table at level Order -1
                        // (not required if the input LM is in lexicographical order)
                        if (maxlev>1 && Order>1){
                                checkbounds(Order-1);
                                delete startpos[Order-1];
                        }
                }
        }
        
        cerr << "closing output file: " << nameNgrams << "\n";
        for (int i=1;i<=maxlev;i++)
                if (maxsize[i] != cursize[i]) {
                        for (int l=1;l<=maxlev;l++)
                                cerr << "Level " << l << ": starting ngrams=" << maxsize[l] << " - actual stored ngrams=" << cursize[l] << "\n";
                        break;
                }
                        
                        Munmap(table[0],filesize,MS_SYNC);
        for (int l=1;l<=maxlev;l++)
                table[l]=0; // to avoid wrong free in ~lmtable()
        cerr << "running fclose...\n";
        fclose(fd);
        cerr << "done\n";
        
        lmtable::getDict()->incflag(0);
        lmtable::getDict()->genoovcode();
        
        // saving header + dictionary
        
        strcpy(nameHeader,outfilename);
        strcat(nameHeader, "-header");
        cerr << "saving header+dictionary in " << nameHeader << "\n";
        fstream out(nameHeader,ios::out);
        
        // print header
        if (isQtable){
                out << "Qblmt" << (isInverted?"I ":" ") << maxlev;
                for (int i=1;i<=maxlev;i++) out << " " << maxsize[i];  // not cursize[i] because the file was already allocated
                out << "\nNumCenters";
                for (int i=1;i<=maxlev;i++)  out << " " << NumCenters[i];
                out << "\n";
                
        }else{
                out << "blmt" << (isInverted?"I ":" ") << maxlev;
                for (int i=1;i<=maxlev;i++) out << " " << maxsize[i];  // not cursize[i] because the file was already allocated
                out << "\n";
        }
        
        lmtable::getDict()->save(out);
        
        out.close();
        cerr << "done\n";
        
        // cat header+dictionary and n-grams files:
        
        char cmd[MAX_LINE];
        sprintf(cmd,"cat %s >> %s", nameNgrams, nameHeader);
        cerr << "run cmd <" << cmd << ">\n";
        system(cmd);
        
        sprintf(cmd,"mv %s %s", nameHeader, outfilename);
        cerr << "run cmd <" << cmd << ">\n";
        system(cmd);
        
        sprintf(cmd,"rm %s", nameNgrams);
        cerr << "run cmd <" << cmd << ">\n";
        system(cmd);
        
        //no more operations are available, the file must be saved!
        exit(0);
        return;
}

void lmtable::loadtxt(istream& inp,const char* header){
        
        
        //open input stream and prepare an input string
        char line[MAX_LINE];
        
        //prepare word dictionary
        //dict=(dictionary*) new di
        dictionary(NULL,1000000);
        lmtable::getDict()->incflag(1);
        
        //put here ngrams, log10 probabilities or their codes
        ngram ng(lmtable::getDict());   
        ngram ing(lmtable::getDict()); //support n-gram 
        
        float prob,bow;
        
        //check the header to decide if the LM is quantized or not
        isQtable=(strncmp(header,"qARPA",5)==0?true:false);
        
        //check the header to decide if the LM table is incomplete
        isItable=(strncmp(header,"iARPA",5)==0?true:false);
        
        //we will configure the table later we we know the maxlev;
        bool yetconfigured=false;
        
        cerr << "loadtxt()\n";
        
        // READ ARPA Header
        int Order,n;
        
        while (inp.getline(line,MAX_LINE)){
                
                if (strlen(line)==MAX_LINE-1){
                        cerr << "lmtable::loadtxt: input line exceed MAXLINE ("
                        << MAX_LINE << ") chars " << line << "\n";
                        exit(1);
                }
                
                bool backslash = (line[0] == '\\');
                
                if (sscanf(line, "ngram %d=%d", &Order, &n) == 2) {
                        maxsize[Order] = n; maxlev=Order; //upadte Order
                        
                }
                
                if (backslash && sscanf(line, "\\%d-grams", &Order) == 1) {
                        
                        //at this point we are sure about the size of the LM
                        if (!yetconfigured){
                                configure(maxlev,isQtable);yetconfigured=true;
                                //allocate space for loading the table of this level
                                for (int i=1;i<=maxlev;i++)
                                        table[i] = new char[(table_pos_t) maxsize[i] * nodesize(tbltype[i])];
                        }
                        
                        cerr << Order << "-grams: reading ";
                        
                        if (isQtable) loadcenters(inp,Order);
                        
                        //allocate support vector to manage badly ordered n-grams
                        if (maxlev>1 && Order<maxlev) {
                                startpos[Order]=new table_entry_pos_t[maxsize[Order]];
                                for (table_entry_pos_t c=0;c<maxsize[Order];c++){
                                        startpos[Order][c]=BOUND_EMPTY1;
                                }
                        }
                        
                        //prepare to read the n-grams entries
                        cerr << maxsize[Order] << " entries\n";
                        
                        //WE ASSUME A WELL STRUCTURED FILE!!!
                        
                        for (table_entry_pos_t c=0;c<maxsize[Order];c++){
                                
                                if (parseline(inp,Order,ng,prob,bow)){
                                        
                                        // if table is inverted then revert n-gram
                                        if (isInverted & Order>1){
                                                ing.invert(ng);
                                                ng=ing;
                                        }
                                        
                                        //if table is in incomplete ARPA format prob is just the
                                        //discounted frequency, so we need to add bow * Pr(n-1 gram)
                                        
                                        //cerr << "ng: " << ng << " prob: " << prob << " bow: " << bow << "\n";
                                        
                                        if (isItable && Order>1) {
                                                //get bow of lower context
                                                get(ng,ng.size,ng.size-1);
                                                float rbow=0.0;
                                                if (ng.lev==ng.size-1){ //found context
                                                        rbow=ng.bow;
                                                        //              int ibow=ng.bow; rbow=*((float *)&ibow);
                                                }
                                                
                                                int tmp=maxlev;
                                                maxlev=Order-1;
                                                //            cerr << ng << "rbow: " << rbow << "prob: " << prob << "low-prob: " << lprob(ng) << "\n";
                                                prob= log(exp((double)prob * M_LN10) +  exp(((double)rbow + lprob(ng)) * M_LN10))/M_LN10;
                                                //            cerr << "new prob: " << prob << "\n";
                                                
                                                maxlev=tmp;
                                        }
                                        
                                        if (isQtable) add(ng, (qfloat_t)prob, (qfloat_t)bow);
                                        else add(ng, prob, bow);
                                        
                                        /*
                                         add(ng,
                                                         (int)(isQtable?prob:*((int *)&prob)),
                                                         (int)(isQtable?bow:*((int *)&bow)));
                                         */
                                }
                        }
                        // now we can fix table at level Order -1
                        if (maxlev>1 && Order>1) checkbounds(Order-1);
                }
        }
        
        lmtable::getDict()->incflag(0);
        cerr << "done\n";
        
}


void lmtable::printTable(int level) {
  char*  tbl=table[level];
  LMT_TYPE ndt=tbltype[level];
  int ndsz=nodesize(ndt);
  table_entry_pos_t printEntryN=1000;
  if (cursize[level]>0)
    printEntryN=(printEntryN<cursize[level])?printEntryN:cursize[level];
        
  cout << "level = " << level << "\n";
        
        //TOCHECK: Nicola, 18 dicembre 2009
  float p;
  for (table_entry_pos_t c=0;c<printEntryN;c++){
    p=prob(tbl,ndt);
    cout << p << " " << word(tbl) << "\n";
    //cout << *(float *)&p << " " << word(tbl) << "\n";
    tbl+=ndsz;
  }
  return;
}

//Checkbound with sorting of n-gram table on disk

void lmtable::checkbounds(int level){
        
        char*  tbl=table[level];
        char*  succtbl=table[level+1];
        
        LMT_TYPE ndt=tbltype[level], succndt=tbltype[level+1];
        int ndsz=nodesize(ndt), succndsz=nodesize(succndt);
        
        //re-order table at level+1 on disk
        //generate random filename to avoid collisions
        ofstream out;string filePath;
        createtempfile(out,filePath,ios::out|ios::binary);
        
        table_entry_pos_t start,end,newstart;
        
        //re-order table at level l+1
        newstart=0;
        //cout << "BOUND_EMPTY1:" << BOUND_EMPTY1 << " BOUND_EMPTY2:" << BOUND_EMPTY2 << std::endl;
        for (table_entry_pos_t c=0;c<cursize[level];c++){
                start=startpos[level][c]; end=bound(tbl+ (table_pos_t) c*ndsz,ndt);
                //    start=startpos[level][c]; end=bound(tbl+c*ndsz,ndt);
                
                //if start==BOUND_EMPTY1 there are no successors for this entry and end==BOUND_EMPTY2
                if (start==BOUND_EMPTY1) end=BOUND_EMPTY2;
                if (end==BOUND_EMPTY2) end=start;
                
                assert(start<=end);
                assert(newstart+(end-start)<=cursize[level+1]);
                assert(end == BOUND_EMPTY1 || end<=cursize[level+1]);
                
                
                if (start<end){
                        out.write((char*)(succtbl + (table_pos_t) start * succndsz),(table_pos_t) (end-start) * succndsz);
                        if (!out.good()){
                                std::cerr << " Something went wrong while writing temporary file " << filePath
                                << " Maybe there is not enough space on this filesystem\n";
                                
                                out.close();
                                removefile(filePath);
                        }
                }
                
                bound(tbl+(table_pos_t) c*ndsz,ndt,newstart+(end-start));
                newstart+=(end-start);
        }
        
        out.close();
        
        fstream inp(filePath.c_str(),ios::in|ios::binary);
        
        inp.read(succtbl,(table_pos_t) cursize[level+1]*succndsz);
        //  inp.read(succtbl,cursize[level+1]*succndsz);
        inp.close();
        
        
}

//Add method inserts n-grams in the table structure. It is ONLY used during
//loading of LMs in text format. It searches for the prefix, then it adds the
//suffix to the last level and updates the start-end positions.

//int lmtable::add(ngram& ng,int iprob,int ibow){
template<typename TA, typename TB> 
int lmtable::add(ngram& ng, TA iprob,TB ibow){
        
  char *found;
        LMT_TYPE ndt=tbltype[1]; //default initialization
        int ndsz=nodesize(ndt); //default initialization
        static int no_more_msg = 0;
        
  if (ng.size>1){
                
    // find the prefix starting from the first level
    table_entry_pos_t start=0, end=cursize[1];
                
          
    for (int l=1;l<ng.size;l++){
                        
      ndt=tbltype[l]; ndsz=nodesize(ndt);
                        
      if (search(l,start,(end-start),ndsz,
                 ng.wordp(ng.size-l+1),LMT_FIND, &found)){
                                
        //update start-end positions for next step
        if (l< (ng.size-1)){
          //set start position
          if (found==table[l]) start=0; //first pos in table
          else start=bound(found - ndsz,ndt); //end of previous entry
                                        
          //set end position
          end=bound(found,ndt);
        }
      }
      else {
                                if (!no_more_msg)
                                        cerr << "warning: missing back-off (at level " << l << ") for ngram " << ng << " (and possibly for others)\n";
                                
                                no_more_msg++;
                                if (!(no_more_msg % 5000000))
                                        cerr << "!";
                                
                                return 0;
      }
    }
                
    // update book keeping information about level ng-size -1.
    // if this is the first successor update start position
    table_entry_pos_t position=(table_entry_pos_t) (((table_pos_t) found-(table_pos_t) table[ng.size-1])/ndsz);
    //table_entry_pos_t position=((table_entry_pos_t)((table_pos_t) found-(table_pos_t) table[ng.size-1])/ndsz);
                
    if (startpos[ng.size-1][position]==BOUND_EMPTY1)
      startpos[ng.size-1][position]=cursize[ng.size];
                
    //always update ending position
    bound(found,ndt,cursize[ng.size]+1);
                
  }
        
  // just add at the end of table[ng.size]
        
  assert(cursize[ng.size]< maxsize[ng.size]); // is there enough space?
  ndt=tbltype[ng.size];ndsz=nodesize(ndt);
        
  found=table[ng.size] + ((table_pos_t) cursize[ng.size] * ndsz);
  //found=table[ng.size] + (cursize[ng.size] * ndsz);
  word(found,*ng.wordp(1));
  prob(found,ndt,iprob);
  if (ng.size<maxlev){bow(found,ndt,ibow);bound(found,ndt,BOUND_EMPTY2);}
        
  cursize[ng.size]++;
        
  if (!(cursize[ng.size]%5000000))
    cerr << ".";
        
  return 1;
        
}


void *lmtable::search(int lev,
                      table_entry_pos_t offs,
                      table_entry_pos_t n,
                      int sz,
                      int *ngp,
                      LMT_ACTION action,
                      char **found){
        
  /***
        if (n >=2)
        cout << "searching entry for codeword: " << ngp[0] << "...";
  ***/
        
  //assume 1-grams is a 1-1 map of the vocabulary
  //CHECK: explicit cast of n into float because table_pos_t could be unsigned and larger than MAXINT
  if (lev==1) return *found=(*ngp < (float) n ? table[1] + (table_pos_t)*ngp * sz:NULL);
        
        
  //prepare table to be searched with mybserach
  char* tb;
  tb=table[lev] + (table_pos_t) offs * sz;
  //prepare search pattern
  char w[LMTCODESIZE];putmem(w,ngp[0],0,LMTCODESIZE);
        
  table_entry_pos_t idx=0; // index returned by mybsearch
  *found=NULL;  //initialize output variable
        
  totbsearch[lev]++;
  switch(action){
                case LMT_FIND:
                        //    if (!tb || !mybsearch(tb,n,sz,(unsigned char *)w,&idx)) return NULL;
                        if (!tb || !mybsearch(tb,n,sz,w,&idx)) return NULL;
                        else
                                //      return *found=tb + (idx * sz);
                                return *found=tb + ((table_pos_t)idx * sz);
                default:
                        error((char*)"lmtable::search: this option is available");
  };
        
  return NULL;
}


//int lmtable::mybsearch(char *ar, table_pos_t n, int size, unsigned char *key, table_pos_t *idx)
int lmtable::mybsearch(char *ar, table_entry_pos_t n, int size, char *key, table_entry_pos_t *idx)
{
  register table_entry_pos_t low, high;
  register char *p;
  register int result=0;
        /*
         register unsigned char *p;
         register table_pos_t result=0;
         register int i;
         */
        
  /* return idx with the first position equal or greater than key */
        
  /*   Warning("start bsearch \n"); */
        
  low = 0;high = n; *idx=0;
  while (low < high)
        {
                
                *idx = (low + high) / 2;
                
                p = (char *) (ar + ((table_pos_t)*idx * size));
                //      p = (unsigned char *) (ar + (*idx * size));
                
                //comparison
                /*
                 for (i=(LMTCODESIZE-1);i>=0;i--){
                         result=key[i]-p[i];
                         if (result) break;
                 }
                 */
                
                result=word(key)-word(p);
                
                if (result < 0)
                        high = *idx;
                else if (result > 0)
                        low = *idx + 1;
                else
                        return 1;
        }
        
  *idx=low;
        
  return 0;
        
}


// generates a LM copy for a smaller dictionary

lmtable* lmtable::cpsublm(dictionary* subdict,bool keepunigr){
        
        //keepunigr=false;              
        //create new lmtable that inherits all features of this lmtable
        
        lmtable* slmt=new lmtable();            
        slmt->configure(maxlev,isQtable);
        slmt->dict=new dictionary((keepunigr?dict:subdict),false);
        std::cerr << "subdict size: " << slmt->dict->size() << "\n";
        
        if (isQtable){
                for (int i=1;i<=maxlev;i++)  {
                        slmt->NumCenters[i]=NumCenters[i];
                        slmt->Pcenters[i]=new float [NumCenters[i]];
                        memcpy(slmt->Pcenters[i],Pcenters[i],NumCenters[i] * sizeof(float));
                        slmt->Bcenters[i]=new float [NumCenters[i]];
                        memcpy(slmt->Bcenters[i],Bcenters[i],NumCenters[i] * sizeof(float));            
                }
        }
        
        //mange dictionary information
        
        //generate OOV codes and build dictionary lookup table 
        dict->genoovcode(); slmt->dict->genoovcode(); subdict->genoovcode();
        std::cerr << "subdict size: " << slmt->dict->size() << "\n";
        int* lookup;lookup=new int [dict->size()];
        for (int c=0;c<dict->size();c++){
                lookup[c]=subdict->encode(dict->decode(c));
                if (c != dict->oovcode() && lookup[c] == subdict->oovcode())
                        lookup[c]=-1; // words of this->dict that are not in slmt->dict
        }
        
        //variables useful to navigate in the lmtable structure
        LMT_TYPE ndt,pndt; int ndsz,pndsz; 
        char *entry, *newentry; 
        table_entry_pos_t start, end, origin;
        
        for (int l=1;l<=maxlev;l++){
                
                slmt->cursize[l]=0;
                slmt->table[l]=NULL;
                
                if (l==1){ //1-gram level 
                        
                        ndt=tbltype[l]; ndsz=nodesize(ndt);
                        
                        for (table_entry_pos_t p=0;p<cursize[l];p++){
                                
                                entry=table[l] + (table_pos_t) p * ndsz; 
                                if (lookup[word(entry)]!=-1 || keepunigr){
                                        
                                        if ((slmt->cursize[l] % slmt->dict->size()) ==0)
                                                slmt->table[l]=(char *)realloc(slmt->table[l],((table_pos_t) slmt->cursize[l] + (table_pos_t) slmt->dict->size()) * ndsz);
                                        
                                        newentry=slmt->table[l] + (table_pos_t) slmt->cursize[l] * ndsz; 
                                        memcpy(newentry,entry,ndsz);
                                        if (!keepunigr) //do not change encoding if keepunigr is true
                                                slmt->word(newentry,lookup[word(entry)]);
                                        
                                        if (l<maxlev) 
                                                slmt->bound(newentry,ndt,p); //store in bound the entry itself (**) !!!!
                                        slmt->cursize[l]++;                     
                                }
                        }
                }
                
                else{ //n-grams n>1: scan lower order table
                        
                        pndt=tbltype[l-1]; pndsz=nodesize(pndt);
                        ndt=tbltype[l]; ndsz=nodesize(ndt);
                        
                        for (table_entry_pos_t p=0; p<slmt->cursize[l-1]; p++){
                                
                                //determine start and end of successors of this entry
                                origin=slmt->bound(slmt->table[l-1] + (table_pos_t)p * pndsz,pndt); //position of n-1 gram in this table (**)
                                if (origin == 0) start=0;                              //succ start at first pos in table[l]
                                else start=bound(table[l-1] + (table_pos_t)(origin-1) * pndsz,pndt);//succ start after end of previous entry
                                        end=bound(table[l-1] + (table_pos_t)origin * pndsz,pndt);           //succ end where indicated 
                                        
                                        if (!keepunigr || lookup[word(table[l-1] + (table_pos_t)origin * pndsz)]!=-1){
                                                while (start < end){
                                                        
                                                        entry=table[l] + (table_pos_t) start * ndsz;
                                                        
                                                        if (lookup[word(entry)]!=-1){
                                                                
                                                                if ((slmt->cursize[l] % slmt->dict->size()) ==0)
                                                                        slmt->table[l]=(char *)realloc(slmt->table[l],(table_pos_t) (slmt->cursize[l]+slmt->dict->size()) * ndsz);
                                                                
                                                                newentry=slmt->table[l] + (table_pos_t) slmt->cursize[l] * ndsz; 
                                                                memcpy(newentry,entry,ndsz);
                                                                if (!keepunigr) //do not change encoding if keepunigr is true
                                                                        slmt->word(newentry,lookup[word(entry)]);
                                                                if (l<maxlev)
                                                                        slmt->bound(newentry,ndt,start); //store in bound the entry itself!!!!
                                                                slmt->cursize[l]++;
                                                        } 
                                                        
                                                        start++;
                                                        
                                                }
                                        }
                                        
                                        //updated bound information of incoming entry
                                        slmt->bound(slmt->table[l-1] + (table_pos_t) p * pndsz, pndt,slmt->cursize[l]);
                                        
                        }                                               
                        
                }
                
        }
        
        
        return slmt;
}



// saves a LM table in text format

void lmtable::savetxt(const char *filename){
        
  fstream out(filename,ios::out);
  table_entry_pos_t cnt[1+MAX_NGRAM];
  int l;
        
  out.precision(7);
        
  if (isQtable){
    out << "qARPA " << maxlev;
    for (l=1;l<=maxlev;l++)
      out << " " << NumCenters[l];
    out << endl;
  }
        
  ngram ng(lmtable::getDict(),0);
        
  cerr << "savetxt: " << filename << "\n";
        
  if (isPruned) ngcnt(cnt); //check size of table by considering pruned n-grams
        
  out << "\n\\data\\\n";
  for (l=1;l<=maxlev;l++){
    out << "ngram " << l << "= " << (isPruned?cnt[l]:cursize[l]) << "\n";
  }
        
  for (l=1;l<=maxlev;l++){
                
    out << "\n\\" << l << "-grams:\n";
    cerr << "save: " << (isPruned?cnt[l]:cursize[l]) << " " << l << "-grams\n";
    if (isQtable){
      out << NumCenters[l] << "\n";
      for (int c=0;c<NumCenters[l];c++){
        out << Pcenters[l][c];
        if (l<maxlev) out << " " << Bcenters[l][c];
        out << "\n";
      }
    }
                
    ng.size=0;
    dumplm(out,ng,1,l,0,cursize[1]);
                
  }
        
  out << "\\end\\\n";
  cerr << "done\n";
}


void lmtable::savebin(const char *filename){
        
  if (isPruned){
    cerr << "savebin: pruned LM cannot be saved in binary form\n";
    exit(0);
  }
                
  fstream out(filename,ios::out);
  cerr << "savebin: " << filename << "\n";
        
  // print header
        if (isQtable){
                out << "Qblmt" << (isInverted?"I":"") << " " << maxlev;
                for (int i=1;i<=maxlev;i++) out << " " << cursize[i];
                out << "\nNumCenters";
                for (int i=1;i<=maxlev;i++)  out << " " << NumCenters[i];
                out << "\n";
                
        }else{
    out << "blmt" << (isInverted?"I":"") << " " << maxlev;
    for (int i=1;i<=maxlev;i++) out << " " << cursize[i] ;
    out << "\n";
  }
        
  lmtable::getDict()->save(out);
        
  for (int i=1;i<=maxlev;i++){
    cerr << "saving " << cursize[i] << " " << i << "-grams\n";
    if (isQtable){
      out.write((char*)Pcenters[i],NumCenters[i] * sizeof(float));
      if (i<maxlev)
        out.write((char *)Bcenters[i],NumCenters[i] * sizeof(float));
    }
    out.write(table[i],(table_pos_t) cursize[i]*nodesize(tbltype[i]));
  }
        
  cerr << "done\n";
}


//manages the long header of a bin file
//and allocates table for each n-gram level

void lmtable::loadbinheader(istream& inp,const char* header){
        
        // read rest of header
        inp >> maxlev;
        
        if (strncmp(header,"Qblmt",5)==0){ 
                isQtable=1;
                if (strncmp(header,"QblmtI",6)==0)
                        isInverted=1;
        }
        else if(strncmp(header,"blmt",4)==0){ 
                isQtable=0;
                if (strncmp(header,"blmtI",5)==0)
                        isInverted=1;
        }
        else error((char*)"loadbin: LM file is not in binary format");
        
        configure(maxlev,isQtable);
        
        for (int l=1;l<=maxlev;l++){
                inp >> cursize[l]; maxsize[l]=cursize[l];
        }
        
        if (isQtable){
                char header2[100];
                inp >> header2;
                for (int i=1;i<=maxlev;i++){
                        inp >> NumCenters[i];
                        cerr << "reading  " << NumCenters[i] << " centers\n";
                }
        }
}

//load codebook of level l

void lmtable::loadbincodebook(istream& inp,int l){
        
  Pcenters[l]=new float [NumCenters[l]];
  inp.read((char*)Pcenters[l],NumCenters[l] * sizeof(float));
  if (l<maxlev){
    Bcenters[l]=new float [NumCenters[l]];
    inp.read((char *)Bcenters[l],NumCenters[l]*sizeof(float));
  }
        
}


//load a binary lmfile

void lmtable::loadbin(istream& inp, const char* header,const char* filename,int mmap){
        
  cerr << "loadbin()\n";
  loadbinheader(inp,header);
  lmtable::getDict()->load(inp);
        
  //if MMAP is used, then open the file
  if (filename && mmap>0){
                
#ifdef WIN32
    error("lmtable::loadbin mmap facility not yet supported under WIN32\n");
#else
                
    if (mmap <= maxlev) memmap=mmap;
    else error((char*)"keep_on_disk value is out of range\n");
                
    if ((diskid=open(filename, O_RDONLY))<0){
      std::cerr << "cannot open " << filename << "\n";
      error((char*)"dying");
    }
                
    //check that the LM is uncompressed
    char miniheader[4];
    read(diskid,miniheader,4);
    if (strncmp(miniheader,"Qblm",4) && strncmp(miniheader,"blmt",4))
      error((char*)"mmap functionality does not work with compressed binary LMs\n");
#endif
  }
        
  for (int l=1;l<=maxlev;l++){
    if (isQtable) loadbincodebook(inp,l);
    if ((memmap == 0) || (l < memmap)){
      cerr << "loading " << cursize[l] << " " << l << "-grams\n";
      table[l]=new char[(table_pos_t) cursize[l] * nodesize(tbltype[l])];
      inp.read(table[l],(table_pos_t) cursize[l] * nodesize(tbltype[l]));
    }
    else{
                        
#ifdef WIN32
      error((char*)"mmap not available under WIN32\n");
#else
      cerr << "mapping " << cursize[l] << " " << l << "-grams\n";
      tableOffs[l]=inp.tellg();
      table[l]=(char *)MMap(diskid,PROT_READ,
                            tableOffs[l], (table_pos_t) cursize[l]*nodesize(tbltype[l]),
                                                                                                                &tableGaps[l]);
      table[l]+=(table_pos_t) tableGaps[l];
      inp.seekg((table_pos_t) cursize[l]*nodesize(tbltype[l]),ios_base::cur);
#endif
                        
    }
  };
        
  cerr << "done\n";
        
}



int lmtable::get(ngram& ng,int n,int lev){
        
  /***
        cout << "cerco:" << ng << "\n";
  ***/
  totget[lev]++;
        
  if (lev > maxlev) error((char*)"get: lev exceeds maxlevel");
  if (n < lev) error((char*)"get: ngram is too small");
        
  //set boudaries for 1-gram
  table_entry_pos_t offset=0,limit=cursize[1];
        
  //information of table entries
  table_entry_pos_t hit;
  char* found; LMT_TYPE ndt;
  ng.link=NULL;
  ng.lev=0;
        
  for (int l=1;l<=lev;l++){
                
    //initialize entry information
    hit = 0 ; found = NULL; ndt=tbltype[l];
                
    if (lmtcache[l] && lmtcache[l]->get(ng.wordp(n),(char *)&found))
      hit=1;
    else
      search(l,
             offset,
             (limit-offset),
             nodesize(ndt),
             ng.wordp(n-l+1),
             LMT_FIND,
             &found);
                
    //insert both found and not found items!!!
    if (lmtcache[l] && hit==0)
      lmtcache[l]->add(ng.wordp(n),(char *)&found);
                
    if (!found) return 0;
    if (prob(found,ndt)==NOPROB) return 0; //pruned n-gram
                ng.path[l]=found; //store path of found entries
    ng.bow=(l<maxlev?bow(found,ndt):0);
    ng.prob=prob(found,ndt);
    ng.link=found;
    ng.info=ndt;
    ng.lev=l;
                
    if (l<maxlev){ //set start/end point for next search
                        
      //if current offset is at the bottom also that of successors will be
      if (offset+1==cursize[l]) limit=cursize[l+1];
      else limit=bound(found,ndt);
                        
      //if current start is at the begin, then also that of successors will be
      if (found==table[l]) offset=0;
      else offset=bound((found - nodesize(ndt)),ndt);
                        
      assert(offset!=BOUND_EMPTY1); assert(limit!=BOUND_EMPTY1);
    }
  }
        
  //put information inside ng
  ng.size=n;  ng.freq=0;
  ng.succ=(lev<maxlev?limit-offset:0);
        
#ifdef TRACE_CACHE
  if (ng.size==maxlev && sentence_id>0){
    *cacheout << sentence_id << " miss " << ng << " " << (unsigned int) ng.link << "\n";
  }
#endif
        
  return 1;
}


//recursively prints the language model table

void lmtable::dumplm(fstream& out,ngram ng, int ilev, int elev, table_entry_pos_t ipos,table_entry_pos_t epos){
        
        LMT_TYPE ndt=tbltype[ilev];
        ngram ing(ng.dict);
        int ndsz=nodesize(ndt);
        
        assert(ng.size==ilev-1);
        assert(ipos>=0 && epos<=cursize[ilev] && ipos<epos);
        ng.pushc(0);
        
        for (table_entry_pos_t i=ipos;i<epos;i++){
                *ng.wordp(1)=word(table[ilev]+(table_pos_t) i*ndsz);
                float ipr=prob(table[ilev]+(table_pos_t) i*ndsz,ndt);
                //int ipr=prob(table[ilev]+i*ndsz,ndt);
                
                //skip pruned n-grams
                if(isPruned && ipr==NOPROB) continue;
                
                if (ilev<elev){
                        //get first and last successor position
                        table_entry_pos_t isucc=(i>0?bound(table[ilev]+ (table_pos_t) (i-1) * ndsz,ndt):0);
                        table_entry_pos_t esucc=bound(table[ilev]+ (table_pos_t) i * ndsz,ndt);
                        if (isucc < esucc) //there are successors!
                                dumplm(out,ng,ilev+1,elev,isucc,esucc);
                        //else
                        //cout << "no successors for " << ng << "\n";
                }
                else{
                        //out << i << " "; //this was just to count printed n-grams
                        out << ipr <<"\t";
                        //out << (isQtable?ipr:*(float *)&ipr) <<"\t";
                        
                        // if table is inverted then revert n-gram
                        if (isInverted & ng.size>1){
                                ing.invert(ng);
                                ng=ing;
                        }
                        
                        for (int k=ng.size;k>=1;k--){
                                if (k<ng.size) out << " ";
                                out << lmtable::getDict()->decode(*ng.wordp(k));
                        }
                        
                        if (ilev<maxlev){
                                float ibo=bow(table[ilev]+ (table_pos_t)i * ndsz,ndt);
                                if (isQtable) out << "\t" << ibo;
                                else if (ibo!=0.0) out << "\t" << ibo;
                                /*
                                 int ibo=bow(table[ilev]+ i * ndsz,ndt);
                                 if (isQtable) out << "\t" << ibo;
                                 else
                                 if (*((float *)&ibo)!=0.0)
                                 out << "\t" << *((float *)&ibo);
                                 */
                        }
                        out << "\n";
                }
        }
}

//succscan iteratively returns all successors of an ngram h for which
//get(h,h.size,h.size) returned true.


int lmtable::succscan(ngram& h,ngram& ng,LMT_ACTION action,int lev){
  assert(lev==h.lev+1 && h.size==lev && lev<=maxlev);
        
  LMT_TYPE ndt=tbltype[h.lev];
  int ndsz=nodesize(ndt);
        
  table_entry_pos_t offset;
  switch (action){
                
                case LMT_INIT:
                        //reset ngram local indexes
                        
                        ng.size=lev;
                        ng.trans(h);
                        //get number of successors of h
                        ng.midx[lev]=0;
                        offset=(h.link>table[h.lev]?bound(h.link-ndsz,ndt):0);
                        h.succ=bound(h.link,ndt)-offset;
                        h.succlink=table[lev]+(table_pos_t) offset * nodesize(tbltype[lev]);
                        return 1; 
                        
                case LMT_CONT:
      
                        if (ng.midx[lev] < h.succ)
      {
        //put current word into ng
        *ng.wordp(1)=word(h.succlink+(table_pos_t) ng.midx[lev]*nodesize(tbltype[lev]));
        ng.midx[lev]++;
        return 1;
      }
                        else
                                return 0;
                        
                default:
                        cerr << "succscan: only permitted options are LMT_INIT and LMT_CONT\n";
                        exit(0);
  }
        
}

//maxsuffptr returns the largest suffix of an n-gram that is contained
//in the LM table. This can be used as a compact representation of the
//(n-1)-gram state of a n-gram LM. if the input k-gram has k>=n then it
//is trimmed to its n-1 suffix.

const char *lmtable::maxsuffptr(ngram ong, unsigned int* size){
  //cerr << "lmtable::maxsuffptr\n";
  //cerr << "ong: " << ong
  //    << " -> ong.size: " << ong.size << "\n";
        
  if (ong.size==0){
    if (size!=NULL) *size=0;
    return (char*) NULL;
  }
        
  if (ong.size>=maxlev) ong.size=maxlev-1;
        
  if (size!=NULL) *size=ong.size; //will return the largest found ong.size
        
  ngram ng=ong;
        
  //ngram ng(lmtable::getDict()); //eventually use the <unk> word
  //ng.trans(ong);
        
  if (get(ng,ng.size,ng.size)){
    if (ng.succ==0) (*size)--;
    return ng.link;
  }
  else{
    ong.size--;
    return maxsuffptr(ong,size);
  }
}


const char *lmtable::cmaxsuffptr(ngram ong, unsigned int* size){
  //cerr << "lmtable::CMAXsuffptr\n";
  //cerr << "ong: " << ong
  //    << " -> ong.size: " << ong.size << "\n";
        
  if (size!=NULL) *size=ong.size; //will return the largest found ong.size
  if (ong.size==0) return (char*) NULL;
  if (ong.size>=maxlev) ong.size=maxlev-1;
        
  char* found;
  unsigned int isize; //internal state size variable
        
  if (statecache && (ong.size==maxlev-1) && statecache->get(ong.wordp(maxlev-1),(char *)&found)){
    if (size!=NULL) statesizecache->get(ong.wordp(maxlev-1),(char *)size);
    return found;
  }
  
  found=(char *)maxsuffptr(ong,&isize);
        
  if (statecache && ong.size==maxlev-1){
    //if (statecache->isfull()) statecache->reset();
    statecache->add(ong.wordp(maxlev-1),(char *)&found);
    statesizecache->add(ong.wordp(maxlev-1),(char *)&isize);
  };
        
  if (size!=NULL) *size=isize;
  
  return found;
}



//returns log10prob of n-gram
//bow: backoff weight
//bol: backoff level

/* double lmtable::lprob(ngram ong,double* bow, int* bol,int internalcall){
//double lmtable::lprob(ngram ong){

if (ong.size==0) return 0.0;
if (ong.size>maxlev) ong.size=maxlev;

if (internalcall==0){ //first call to lprob
                if (bow) *bow=0;
                if (bol) *bol=0;
}

ngram ng=ong;
//ngram ng(lmtable::getDict()); //avoid dictionary transfer
//ng.trans(ong);

double rbow,lpr=0;
float ibow,iprob;
//int ibow,iprob;

if (get(ng,ng.size,ng.size)){
                iprob=ng.prob;
                lpr = (double)(isQtable?Pcenters[ng.size][(qfloat_t)iprob]:iprob);
                //lpr = (double)(isQtable?Pcenters[ng.size][iprob]:*((float *)&iprob));
                if (*ng.wordp(1)==dict->oovcode()) lpr-=logOOVpenalty;
                return (double)lpr;
}
else{
                if (ng.size==1) //means an OOV word
                        return -log(UNIGRAM_RESOLUTION)/M_LN10;
                else{ //compute backoff
                                        //set backoff state, shift n-gram, set default bow prob
                        rbow=0.0;
                        if (bol) (*bol)++; //increase backoff level
                        if ((ng.lev==(ng.size-1)) && (*ng.wordp(2)!=dict->oovcode())){
                                //found history in table: use its bo weight
                                //avoid wrong quantization of bow of <unk>
                                ibow=ng.bow;
                                rbow= (double) (isQtable?Bcenters[ng.lev][(qfloat_t)ibow]:ibow);
                                //rbow= (double) (isQtable?Bcenters[ng.lev][ibow]:*((float *)&ibow));
                        }
                        
                        if (bow) (*bow)+=rbow;
                        
                        //prepare recursion step
                        ong.size--;
                        return rbow + lmtable::lprob(ong,bow,bol,internalcall=1);
                }
}
}
*/


//non recursive version, also includes maxsuffptr 

double lmtable::lprob(ngram ong,double* bow, int* bol, char** maxsuffptr,unsigned int* statesize){
        
        if (ong.size==0) return 0.0; //sanity check
        if (ong.size>maxlev) ong.size=maxlev; //adjust n-gram level to table size
        
        if (bow) *bow=0; //initialize back-off weight
        if (bol) *bol=0; //initialize bock-off level    
        
        
        double rbow=0,lpr=0; //output back-off weight and logprob 
        float ibow,iprob;  //internal back-off weight and logprob
        
        
        if (isInverted){
                ngram ing=ong; //inverted ngram
                
                ing.invert(ong);
                //cout << "ngram:" << ing << "\n";      
                get(ing,ing.size,ing.size); // dig in the trie
                if (ing.lev >0){ //found something?
                        iprob=ing.prob;
                        lpr = (double)(isQtable?Pcenters[ing.size][(qfloat_t)iprob]:iprob);
                        if (*ong.wordp(1)==dict->oovcode()) lpr-=logOOVpenalty; //add OOV penalty
                        if (statesize)  *statesize=MIN(ing.lev,(ing.size-1)); //find largest n-1 gram suffix 
                        if (maxsuffptr) *maxsuffptr=ing.path[MIN(ing.lev,(ing.size-1))];                        
                }else{ // means a real unknown word!    
                        lpr=-log(UNIGRAM_RESOLUTION)/M_LN10;
                        if (statesize)  *statesize=0;     //default statesize for zero-gram! 
                        if (maxsuffptr) *maxsuffptr=NULL; //default stateptr for zero-gram! 
                }
                
                if (ing.lev < ing.size){ //compute backoff weight
                        int depth=(ing.lev>0?ing.lev:1); //ing.lev=0 (real unknown word) is still a 1-gram 
                        if (bol) *bol=ing.size-depth;
                        ing.size--; //get n-gram context
                        get(ing,ing.size,ing.size); // dig in the trie
                        if (ing.lev>0){//found something?
                                                                                 //collect back-off weights
                                for (int l=depth;l<=ing.lev;l++){
                                        //start from first back-off level
                                        assert(ing.path[l]!=NULL); //check consistency of table
                                        ibow=this->bow(ing.path[l],tbltype[l]);
                                        rbow+= (double) (isQtable?Bcenters[l][(qfloat_t)ibow]:ibow);
                                        //avoids bad quantization of bow of <unk>
                                        if (isQtable && (*ing.wordp(1)==dict->oovcode())) 
                                                rbow-=(double)Bcenters[l][(qfloat_t)ibow];
                                }
                        }
                }
                
                if (bow) (*bow)=rbow;
                return rbow + lpr;
        }
        else{
                
                for (ngram ng=ong;ng.size>0;ng.size--){
                        
                        if (get(ng,ng.size,ng.size)){
                                iprob=ng.prob;
                                lpr = (double)(isQtable?Pcenters[ng.size][(qfloat_t)iprob]:iprob);
                                if (*ng.wordp(1)==dict->oovcode()) lpr-=logOOVpenalty; //add OOV penalty
                                if (maxsuffptr || statesize){ //one extra step is needed if ng.size=ong.size
                                        if (ong.size==ng.size){
                                                ng.size--;
                                                get(ng,ng.size,ng.size);
                                        }
                                        if (statesize)  *statesize=ng.size;
                                        if (maxsuffptr) *maxsuffptr=ng.link; //we should check ng.link != NULL   
                                }
                                return rbow+lpr; 
                        }else{
                                if (ng.size==1){ //means a real unknow word!
                                        if (maxsuffptr) *maxsuffptr=NULL; //default stateptr for zero-gram! 
                                        if (statesize)  *statesize=0;
                                        return rbow -log(UNIGRAM_RESOLUTION)/M_LN10;
                                }
                                else{ //compute backoff
                                        if (bol) (*bol)++; //increase backoff level
                                        if (ng.lev==(ng.size-1)){ //if search stopped at previous level
                                                ibow=ng.bow;
                                                rbow+= (double) (isQtable?Bcenters[ng.lev][(qfloat_t)ibow]:ibow);
                                                //avoids bad quantization of bow of <unk>
                                                if (isQtable && (*ng.wordp(2)==dict->oovcode())) 
                                                        rbow-=(double)Bcenters[ng.lev][(qfloat_t)ibow];
                                        }
                                        if (bow) (*bow)=rbow;
                                }
                                
                        }
                        
                }
        }
        assert(0); //never pass here!!!
        return 1.0;
}       

//return log10 probsL use cache memory

double lmtable::clprob(ngram ong){
        
  if (ong.size==0) return 0.0;
        
  if (ong.size>maxlev) ong.size=maxlev;
        
  double logpr;
        
#ifdef TRACE_CACHE
  if (probcache && ong.size==maxlev && sentence_id>0){
    *cacheout << sentence_id << " " << ong << "\n";
  }
#endif
        
  //cache hit
  if (probcache && ong.size==maxlev && probcache->get(ong.wordp(maxlev),(char *)&logpr)){
    return logpr;
  }
        
        
  //cache miss
  logpr=lmtable::lprob(ong);
        
  if (probcache && ong.size==maxlev){
    probcache->add(ong.wordp(maxlev),(char *)&logpr);
  };
        
  return logpr;
        
};



void lmtable::stat(int level){
  table_pos_t totmem=0,memory;
  float mega=1024 * 1024;
        
  cout.precision(2);
        
  cout << "lmtable class statistics\n";
        
  cout << "levels " << maxlev << "\n";
  for (int l=1;l<=maxlev;l++){
    memory=(table_pos_t) cursize[l] * nodesize(tbltype[l]);
    cout << "lev " << l
                        << " entries "<< cursize[l]
                        << " used mem " << memory/mega << "Mb\n";
    totmem+=memory;
  }
        
  cout << "total allocated mem " << totmem/mega << "Mb\n";
        
  cout << "total number of get and binary search calls\n";
  for (int l=1;l<=maxlev;l++){
    cout << "level " << l << " get: " << totget[l] << " bsearch: " << totbsearch[l] << "\n";
  }
        
  if (level >1 ) lmtable::getDict()->stat();
        
}

void lmtable::reset_mmap(){
#ifndef WIN32
  if (memmap>0 and memmap<=maxlev)
    for (int l=memmap;l<=maxlev;l++){
      //std::cerr << "resetting mmap at level:" << l << "\n";
      Munmap(table[l]-tableGaps[l],(table_pos_t) cursize[l]*nodesize(tbltype[l])+tableGaps[l],0);
      table[l]=(char *)MMap(diskid,PROT_READ,
                            tableOffs[l], (table_pos_t)cursize[l]*nodesize(tbltype[l]),
                            &tableGaps[l]);
      table[l]+=(table_pos_t)tableGaps[l];
    }
#endif
}

// ng: input n-gram

// *lk: prob of n-(*bol) gram
// *boff: backoff weight vector
// *bol:  backoff level

double lmtable::lprobx(ngram    ong,
                       double   *lkp,
                       double   *bop,
                       int      *bol)
{
  double bo, lbo, pr;
  float         ipr;
  //int         ipr;
  ngram         ng(dict), ctx(dict);
        
  if(bol) *bol=0;
  if(ong.size==0) {
    if(lkp) *lkp=0;
    return 0;   // lprob ritorna 0, prima lprobx usava LOGZERO
  }
  if(ong.size>maxlev) ong.size=maxlev;
  ctx = ng = ong;
  bo=0;
  ctx.shift();
  while(!get(ng)) { // back-off
                
    //OOV not included in dictionary
    if(ng.size==1) {
      pr = -log(UNIGRAM_RESOLUTION)/M_LN10;
      if(lkp) *lkp=pr; // this is the innermost probability
      pr += bo; //add all the accumulated back-off probability
      return pr;
    }
    // backoff-probability
    lbo = 0.0; //local back-off: default is logprob 0
    if(get(ctx)){ //this can be replaced with (ng.lev==(ng.size-1))
      ipr = ctx.bow;
      lbo = isQtable?Bcenters[ng.size][(qfloat_t)ipr]:ipr;
      //lbo = isQtable?Bcenters[ng.size][ipr]:*(float*)&ipr;
    }
    if(bop) *bop++=lbo;
    if(bol) ++*bol;
    bo += lbo;
    ng.size--;
    ctx.size--;
  }
  ipr = ng.prob;
  pr = isQtable?Pcenters[ng.size][(qfloat_t)ipr]:ipr;
  //pr = isQtable?Pcenters[ng.size][ipr]:*((float*)&ipr);
  if(lkp) *lkp=pr;
  pr += bo;
  return pr;
}


// FABIO
table_entry_pos_t lmtable::wdprune(float        *thr,
                                                                                                                                         int    aflag)
{
  int   l;
  ngram ng(lmtable::getDict(),0);
        
  isPruned=true;  //the table now might contain pruned n-grams
        
  ng.size=0;
  for(l=2; l<=maxlev; l++) wdprune(thr, aflag, ng, 1, l, 0, cursize[1]);
  return 0;
}

// FABIO: LM pruning method

table_entry_pos_t lmtable::wdprune(float *thr, int aflag, ngram ng, int ilev, int elev, table_entry_pos_t ipos, table_entry_pos_t       epos, double    tlk,
                                                                                                                                         double bo, double      *ts, double     *tbs)
{
  LMT_TYPE      ndt=tbltype[ilev];
  int              ndsz=nodesize(ndt);
  char           *ndp;
  float          lk;
  float ipr, ibo;
  //int ipr, ibo;
  table_entry_pos_t i, k, nk;
        
  assert(ng.size==ilev-1);
  assert(ipos>=0 && epos<=cursize[ilev] && ipos<epos);
        
  ng.pushc(0); //increase size of n-gram
        
  for(i=ipos, nk=0; i<epos; i++) {
                
    //scan table at next level ilev from position ipos
    ndp = table[ilev]+(table_pos_t)i*ndsz;
    *ng.wordp(1) = word(ndp);
                
    //get probability
    ipr = prob(ndp, ndt);
    if(ipr==NOPROB) continue;   // Has it been already pruned ??
    lk = ipr;
    //lk = *(float*)&ipr;
                
    if(ilev<elev) { //there is an higher order
                        
      //get backoff-weight for next level
      ibo = bow(ndp, ndt);
      bo = ibo;
      //bo = *(float*)&ibo;
                        
      //get table boundaries for next level
      table_entry_pos_t isucc = i>0 ? bound(ndp-ndsz, ndt) : 0;
      table_entry_pos_t  esucc = bound(ndp, ndt);
      if(isucc>=esucc) continue; // no successors
                        
      //look for n-grams to be pruned with this context (see
      //back-off weight)
prune:  double ts=0, tbs=0;
      k = wdprune(thr, aflag, ng, ilev+1, elev, isucc, esucc,
                                                                        tlk+lk, bo, &ts, &tbs);
      //k  is the number of pruned n-grams with this context
      if(ilev!=elev-1) continue;
      if(ts>=1 || tbs>=1) {
                                cerr << "ng: " << ng
                                <<" ts=" << ts
                                <<" tbs=" << tbs
                                <<" k=" << k
                                <<" ns=" << esucc-isucc
                                << "\n";
                                if(ts>=1) {
                                        pscale(ilev+1, isucc, esucc,
                                                                 0.999999/ts);
                                        goto prune;
                                }
      }
      // adjusts backoff:
      // 1-sum_succ(pr(w|ng)) / 1-sum_succ(pr(w|bng))
      bo = log((1-ts)/(1-tbs))/M_LN10;
      ibo=(float)bo;
      //*(float*)&ibo=bo;
      bow(ndp, ndt, ibo);
    } else { //we are at the highest level
                        
      //get probability of lower order n-gram
      ngram     bng = ng; --bng.size;
      double blk = lprob(bng);
                        
      double wd = pow(10., tlk+lk) * (lk-bo-blk);
      if(aflag&&wd<0) wd=-wd;
      if(wd > thr[elev-1]) {    // kept
                                *ts += pow(10., lk);
                                *tbs += pow(10., blk);
      } else {          // discarded
                                ++nk;
                                prob(ndp, ndt, NOPROB);
      }
    }
  }
  return nk;
}

int lmtable::pscale(int lev, table_entry_pos_t ipos, table_entry_pos_t epos, double s)
{
  LMT_TYPE        ndt=tbltype[lev];
  int             ndsz=nodesize(ndt);
  char            *ndp;
  float             ipr;
  //int             ipr;
        
  s=log(s)/M_LN10;
  ndp = table[lev]+ (table_pos_t) ipos*ndsz;
  for(table_entry_pos_t i=ipos; i<epos; ndp+=ndsz,i++) {
    ipr = prob(ndp, ndt);
    if(ipr==NOPROB) continue;
    ipr+=(float) s;
    //*(float*)&ipr+=s;
    prob(ndp, ndt, ipr);
  }
  return 0;
}

//recompute table size by excluding pruned n-grams
table_entry_pos_t lmtable::ngcnt(table_entry_pos_t      *cnt)
{
  ngram ng(lmtable::getDict(),0);
  memset(cnt, 0, (maxlev+1)*sizeof(*cnt));
  ngcnt(cnt, ng, 1, 0, cursize[1]);
  return 0;
}

//recursively compute size
table_entry_pos_t lmtable::ngcnt(table_entry_pos_t *cnt, ngram  ng, int l, table_entry_pos_t ipos, table_entry_pos_t    epos){
        
  table_entry_pos_t     i, isucc, esucc;
  float ipr;
  //int ipr;
  char          *ndp;
  LMT_TYPE      ndt=tbltype[l];
  int           ndsz=nodesize(ndt);
        
  ng.pushc(0);
  for(i=ipos; i<epos; i++) {
    ndp = table[l]+(table_pos_t) i*ndsz;
    *ng.wordp(1)=word(ndp);
    ipr=prob(ndp, ndt);
    if(ipr==NOPROB) continue;
    ++cnt[l];
    if(l==maxlev) continue;
    isucc = (i>0)?bound(ndp-ndsz, ndt):0;
    esucc = bound(ndp, ndt);
    if(isucc < esucc) ngcnt(cnt, ng, l+1, isucc, esucc);
  }
  return 0;
}

---