/* Generate H for  LDPC*/
/* Based on sparse.c by Matt Davey <mcdavey@mrao.cam.ac.uk> with his permission*/
#include <math.h>
#include "mex.h"

/* Input Arguments: parameters*/
#define	M_IN	prhs[0]          /* number of parity checks */
#define	N_IN	prhs[1]          /* blocklength */ 
#define	T_IN	prhs[2]          /* mean column weight */
#define	Q_IN	prhs[3]          /* GF base  */
#define	SEED_IN	prhs[4]          /* seed for random generator */


/* Output Arguments: matrices*/
#define	H_OUT	plhs[0]


void mexFunction(
                 int nlhs,       mxArray *plhs[],
                 int nrhs, const mxArray *prhs[]
		 )
{
  short	**M_list, **N_list, *M_target;
  double *pp,*sr,*s,*ss;
  int N,M,q,i,j,k,nzmax,*irs,*jcs,l,
	  tr,tm,tc,done,redo,tmp,regime,tr_max,t_max, m_low;
  float t;
  long seed;
  void adjust(int *, int *, int *, int *);
  unsigned int K2,M2;
  char c;
  mxArray *arg_in[2], *arg_out[1]; /* to call rand generator of Matlab*/

  /* Check for proper number of arguments */
  if (nrhs != 5) {
    mexErrMsgTxt("GENERATE requires five input arguments.");
  } else if (nlhs > 1) {
    mexErrMsgTxt("GENERATE requires one output argument.");
  }

  pp = mxGetPr(N_IN);    N = (int) (*pp);
  pp = mxGetPr(M_IN);    M = (int) (*pp);
  pp = mxGetPr(T_IN);    t = (float) (*pp);
  pp = mxGetPr(Q_IN);    q = (int) (*pp);
  pp = mxGetPr(SEED_IN); seed = (int) (*pp);
  

  arg_in[0] = mxCreateString("state");
  arg_in[1] = mxCreateDoubleMatrix(1, 1, mxREAL);
  s = mxGetPr(arg_in[1]);
  s[0] = seed; /* this will be used to call rand*/

  /* initialize random generator */
  mexCallMATLAB(0, NULL, 2, arg_in, "rand"); /* rand('state',seed) */
  s[0] = 1; /* from now on use s to store "1"*/


/* I have no idea about the details of the following - please
ask the author - :).  igor*/





/* Generate some sparse matrices for error-correcting codes.  Supply
 * the following parameters:
 * 	N: Blocklength
 * 	M: Number of parity checks
 * 	t: Mean column weight.
 *
 * If t<3, then we generate weight 2 columns systematically in the
 * form of blocks of identity matrices to reduce probability of
 * getting short cycle lengths.  Can't have more than M weight 2
 * columns, though.
 *
 * Having generated any weight 2 columns, we fill the rest.  Calculate
 * number of columns to fill with floor(t) and number with ceiling(t).
 * Find mean row weight and calculate number of rows to fill with
 * floor(r) and ceiling(r).
 *
 * We fill rows as follows:
 * (Regime 0): Fill so that rows contain <= tr ones until M-tm rows
 * 	    contain tr ones.
 * (Regime 1): Fill so that rows contain <= (tr+1) ones until tm rows
 * 	    contain (tr+1) ones.
 * (Regime 2): Fill remaining rows containing < tr ones until done.
 */

 
  /* N=6112;M=4512;t=2.3;*/

     
  t_max=(int)ceil(t);
  M_list=(short **)mxMalloc(N*sizeof(short *));
  M_target=(short *)mxMalloc(M*sizeof(short *));
  N_list=(short **)mxMalloc(M*sizeof(short *));
  for(i=0;i<N;i++){
    M_list[i]=(short *)mxMalloc((t_max+1)*sizeof(short));
  }
  i=0;
  /* Do we have any weight 2 columns?
   *
   * If so, do this first.  Remember that M mightn't have a large
   * power of two as a divisor, so might need to find some M'<M to use
   * as unit length.
   */
  K2=0;
  if(t<3){
    K2=ceil((double)N*(3-t));
	if(K2>M)
		mexErrMsgTxt("GENERATE: Can't have more than M weight 2 columns.");
    j=2;
    done=0;
    for(i=0;!done;i++){
      M2=floor((double)M/(double)j);
      if((M2*(j-1))>=K2) done=1;
      j*=2;
    }
    M2*=(j/4);
  }
  /* 
   * i contains number of identity blocks we'll need.... */
  tr=((short)floor((double)(t*N)/(double)M));
  /* Now we want `tr' to be final minimum row weight, `tr_max' to be
   * final maximum row weight, `tm' to be number of rows which will
   * have weight greater than `tr'.  'tc' will be a running count of
   * how many rows we still have to fill up to weight `tr'.  Once we
   * hit this many, we can start overfilling rows.
   */
  if (i>tr){
    tr_max=i;
    /* If identity blocks make overheavy rows, we need to calculate
     * the minimum row weight.
     */
    done=0;
    k=1;
    j=floor((double)t*N)-2*K2; /* Number of ones left to distribute */
    for(i=0;!done;i++){
      /* (M-2*M2) rows will be empty after identity blocks */
      j-=((M-2*M2)+(2*M2*(k-1))/k);
      if(j<0) {
	done=1;
      }
      else {
	k*=2;
      }
    }
    tr=i-1;
    tm=M+j;
  }
  else {
    /* This is easier! */
    tr_max=tr+1;
    tm=(int)floor((((double)t*N)/(double)M-tr)*M +0.5);
  }
  tc=M-tm;
  for(i=0;i<M;i++){
    N_list[i]=(short *)mxMalloc((tr_max+1)*sizeof(short));  
  }
  for(i=0;i<M;i++){
    N_list[i][0]=0;
  }

  regime=0;
  /* Generate weight 2 columns.  First create two identity matrices on
   * top of each other, then two 1/2 size ones in the lower rows of
   * the matrix, and so on.
   *
   * j:   length of current identity block
   * k:   base of this block
   * i:   current column position
   */
  j=M2;
  k=0;
  i=0;
  while(i<K2){
    for(;(i-k)<j && i<K2;i++){
      M_list[i][0]=2;
      M_list[i][1]=i;
      M_list[i][2]=i+j;
      N_list[i][0]++;
      if(N_list[i][0]==tr) adjust(&tm,&tr,&tc,&regime);
      N_list[i][N_list[i][0]]=i;
      N_list[i+j][0]++;
      if(N_list[i+j][0]==tr) adjust(&tm,&tr,&tc,&regime);
      N_list[i+j][N_list[i+j][0]]=i;
    }
    k=i;
    j/=2;
  }
 
 /* Now fill the unsystematic columns, ensuring weight per row as even as poss. */
  i=K2;
  if(K2==0){
    /* Fill low weight columns */
    for(i=0;i<(int)(N*(t_max-t)+0.5);i++){
      for(k=1;k<=(int)floor(t);k++){
	done=0;
	do {
      mexCallMATLAB(1, arg_out,1 , &arg_in[1], "rand"); /* ss = rand(1) */
	  ss = mxGetPr(arg_out[0]);
	  j=(short)floor(M*ss[0]);
      mexCallMATLAB(1, arg_out,1 , &arg_in[1], "rand");
	  ss = mxGetPr(arg_out[0]);
	  if((ss[0])<(1-(double)N_list[j][0]/(double)tr)) {
	    done=1;
	    for(l=1;l<k;l++) if(j==M_list[i][l]) done=0;
	  }
	} while(!done);
	N_list[j][0]++;
	N_list[j][N_list[j][0]]=i;
	if(N_list[j][0]==tr) adjust(&tm,&tr,&tc,&regime);
	M_list[i][k]=j;
      }
      M_list[i][0]=k-1;
    }
  }
  redo=1;
  for(;i<N;i++){
    fprintf(stderr,"%d\r",i);
    for(k=1;k<=t_max;k++){
      done=0;
      do {
	/* find the lowest weight rows, and fill one of them */
	if(redo){
	  l=tr_max;
	  for(j=0;j<M;j++) if (N_list[j][0]<l) l=N_list[j][0];
	  m_low=0;
	  for(j=0;j<M;j++) if (N_list[j][0]==l) {M_target[m_low]=j; m_low++;}
	}
	mexCallMATLAB(1, arg_out,1 , &arg_in[1], "rand");
	ss = mxGetPr(arg_out[0]);
	j=M_target[tmp=(short)floor(m_low*ss[0])];
	/*	if(ss[0]<(1-(double)N_list[j][0]/(double)tr)) {*/
	  done=1;
	  for(l=1;l<k;l++) if(j==M_list[i][l]) done=0;
	  if(done==1){
	    if(m_low==1) redo=1;
	    else {
	      for(;tmp<(m_low-1);tmp++) M_target[tmp]=M_target[tmp+1];
	      m_low--;
	      redo=0;
	    }
	  }
	  /*	}*/
	  
      } while(!done);
      N_list[j][0]++;
      N_list[j][N_list[j][0]]=i;
      if(N_list[j][0]==tr) adjust(&tm,&tr,&tc,&regime);
      M_list[i][k]=j;
    }
    M_list[i][0]=k-1;
  }
  tr=((short)ceil((double)(3*N-K2)/(double)M));
  

  for(i=0;i<M;i++){
    mxFree(N_list[i]);
  }
  mxFree(N_list);


/* done generating H matrix */




  
    /* Allocate space for sparse matrix */
    nzmax=0; for(j=0 ; j<N ; j++) nzmax +=M_list[j][0];
	mexPrintf("%d \n",nzmax);
    /* NOTE: The maximum number of non-zero elements cannot be less
       than the number of columns in the matrix. */
    if (N>nzmax){
	nzmax=N;
    }
    plhs[0] = mxCreateSparse(M,N,nzmax,mxREAL);
    sr  = mxGetPr(plhs[0]);
    irs = mxGetIr(plhs[0]);  /* row */
    jcs = mxGetJc(plhs[0]);  /* column */
    
    /* Copy nonzeros */
    k = 0; 
    for (j=0; (j<N ); j++) {
	jcs[j] = k;
	for (i=1; (i<=M_list[j][0] ); i++) {
		sr[k] = 1;
		irs[k] = M_list[j][i];
		k++;	    
	}
    }
    jcs[N] = k;



  for(i=0;i<N;i++){
    mxFree(M_list[i]);
  }
  mxFree(M_list);

  return;
}



void adjust(int *tm, int *tr, int *tc, int *regime){
  switch(*regime){
  case 0:
    (*tc)--;
    if((*tc)==0){
      *regime=1;
      (*tr)++;
    }
    break;
  case 1:
    (*tm)--;
    if((*tm)==0){
      *regime=2;
      (*tr)--;
    }
    break;
  }
}