void mexFunction(int nlhs, mxArray *plhs[], int nrhs, const mxArray *prhs[])
{
int n, p ;
double *a, *l;
double *adiag, *ldiag;
int *acol_ptr, *arow_ind ;
int *lcol_ptr, *lrow_ind ;
/* Working arrays */
double *w1, *w2 ;
int *iw ;
/* Local variables */
int i, nnz ;
double alpha ;
/* Right hand side */
n = (int) mxGetScalar(prhs[0]);
a = mxGetPr(prhs[1]);
arow_ind = mxGetIr(prhs[1]);
acol_ptr = mxGetJc(prhs[1]);
adiag = mxGetPr(prhs[2]) ;
p = (int) mxGetScalar(prhs[3]) ;
nnz = acol_ptr[n] ;
/* Left-hand side */
if (plhs[0] != NULL) mxDestroyArray(plhs[0]) ;
plhs[0] = mxCreateSparse(n, n, nnz+n*p, mxREAL) ;
l = mxGetPr(plhs[0]);
lrow_ind = mxGetIr(plhs[0]);
lcol_ptr = mxGetJc(plhs[0]);
if (plhs[1] != NULL) mxDestroyArray(plhs[1]) ;
plhs[1] = mxCreateDoubleMatrix(n, 1, mxREAL) ;
ldiag = mxGetPr(plhs[1]) ;
/* Allocate working arrays */
if (( (w1=(double *)calloc(n, sizeof(double)))==(double *)NULL ) ||
( (w2=(double *)calloc(n, sizeof(double)))==(double *)NULL ) ||
( (iw=(int *)calloc(3*n, sizeof(int)))==(int *)NULL )) {
printf("Not enough memory\n") ;
}
/* Change a's i j into fortran index */
for (i = 0; i < nnz; i++) arow_ind[i] = arow_ind[i] + 1 ;
for (i = 0; i <= n; i++) acol_ptr[i] = acol_ptr[i] + 1 ;
/* Call icf fortran subroutine */
alpha = 0.0 ;
dicfs_(&n, &nnz, a, adiag, acol_ptr, arow_ind,
l, ldiag, lcol_ptr, lrow_ind, &p, &alpha,
iw, w1, w2) ;
/* Change a's i j into C index */
for (i = 0; i < nnz; i++)
arow_ind[i] = arow_ind[i] - 1 ;
for (i = 0; i < lcol_ptr[n]; i++)
lrow_ind[i] = lrow_ind[i] - 1 ;
for (i = 0; i <= n; i++) {
acol_ptr[i] = acol_ptr[i] - 1 ;
lcol_ptr[i] = lcol_ptr[i] - 1 ;
}
/* Free memory */
free(w1) ;
free(w2) ;
free(iw) ;
} /* mexFunction */
void mexFunction(int nlhs, mxArray *plhs[], int nrhs, const mxArray *prhs[])
{
/* counters */
mwIndex c, r, nv, oc, or, vc, tc;
mwSize m, n;
double iv;
/* pointers */
mwIndex *io, *jo;
const double *v, *i, *j;
double *vo;
/* input check */
if ((nrhs != 1) &&
(nrhs != 2) &&
(nrhs != 5))
mexErrMsgTxt("Requires 1, 2, or 5 input arguments.");
if ((mxGetClassID(*prhs) != mxDOUBLE_CLASS) ||
mxIsSparse(*prhs))
mexErrMsgTxt("First input argument must be of type full double.");
nv = mxGetNumberOfElements(*prhs);
if ((nrhs == 2) &&
((!mxIsSparse(prhs[1])) ||
(mxGetClassID(prhs[1]) != mxDOUBLE_CLASS) ||
(mxGetNzmax(prhs[1]) < nv)))
mexErrMsgTxt("For two-argument call, second input argument must be of type sparse double with at least V values.");
else if ((nrhs == 5) &&
((mxGetClassID(prhs[1]) != mxDOUBLE_CLASS) ||
(mxIsSparse(prhs[1])) ||
(mxGetNumberOfElements(prhs[1]) != 1) ||
(mxGetClassID(prhs[2]) != mxDOUBLE_CLASS) ||
(mxIsSparse(prhs[2])) ||
(mxGetNumberOfElements(prhs[2]) != 1) ||
(mxGetClassID(prhs[3]) != mxDOUBLE_CLASS) ||
(mxIsSparse(prhs[3])) ||
(mxGetNumberOfElements(prhs[3]) != nv) ||
(mxGetClassID(prhs[4]) != mxDOUBLE_CLASS) ||
(mxIsSparse(prhs[4])) ||
(mxGetNumberOfElements(prhs[4]) != nv)))
mexErrMsgTxt("For five-argument call, arguments must be Vx1 v, 1x1 m, 1x1 n, Vx1 i, and Vx1 j.");
/* get data */
v = (const double *) mxGetData(*prhs);
/* one argument only */
if (nrhs == 1) {
/* create output */
*plhs = mxCreateSparse(nv, nv, nv, mxREAL);
if (*plhs == NULL)
mexErrMsgTxt("Error creating sparse double matrix.");
vo = (double *) mxGetData(*plhs);
if (vo == NULL)
mexErrMsgTxt("Error retrieving double data pointer.");
io = (mwIndex *) mxGetIr(*plhs);
if (io == NULL)
mexErrMsgTxt("Error retrieving Ir index pointer.");
jo = (mwIndex *) mxGetJc(*plhs);
if (jo == NULL)
mexErrMsgTxt("Error retrieving Jc index pointer.");
/* loop */
for (vc = 0; vc < nv; ++vc) {
*vo++ = *v++;
*io++ = vc;
*jo++ = vc;
}
*jo = vc;
/* two arguments, overwrite existing sparse */
} else if (nrhs == 2) {
/* copy pointer */
*plhs = (mxArray *) ((void *) prhs[1]);
/* get data pointer */
vo = (double *) mxGetData(*plhs);
/* loop */
for (vc = nv; vc > 0; --vc)
*vo++ = *v++;
/* five arguments, create new sparse from complex data */
} else {
/* get size */
i = (const double *) mxGetData(prhs[1]);
m = (int) *i;
i = (const double *) mxGetData(prhs[2]);
n = (int) *i;
i = (const double *) mxGetData(prhs[3]);
j = (const double *) mxGetData(prhs[4]);
/* create sparse matrix */
*plhs = mxCreateSparse(m, n, nv, mxREAL);
if (*plhs == NULL)
mexErrMsgTxt("Error creating sparse double matrix.");
vo = (double *) mxGetData(*plhs);
if (vo == NULL)
mexErrMsgTxt("Error retrieving double data pointer.");
io = (mwIndex *) mxGetIr(*plhs);
if (io == NULL)
mexErrMsgTxt("Error retrieving Ir index pointer.");
jo = (mwIndex *) mxGetJc(*plhs);
if (jo == NULL)
mexErrMsgTxt("Error retrieving Jc index pointer.");
/* loop */
for (vc = nv, or = 0, oc = 0, tc = 0; vc > 0; --vc) {
/* get target index */
r = (mwIndex) *i++;
c = (mwIndex) *j++;
iv = *v++;
/* only allow if valid */
if ((c < oc) ||
((c == oc) && (r <= or)) ||
(r < 1) ||
(r > m) ||
(c < 1) ||
(c > n))
continue;
/* update column ? */
while (c > oc) {
*jo++ = tc;
++oc;
}
/* write into arrays */
*io++ = (r - 1);
*vo++ = iv;
/* keep track of written values */
or = r;
++tc;
}
/* extend column */
while (oc <= n) {
*jo++ = tc;
++oc;
}
}
}
void mexFunction
(
int nlhs,
mxArray *plhs [],
int nrhs,
const mxArray *prhs []
)
{
/* Compressed column form of L, x and b */
mwIndex *Lp, *Ap, *Ai, *Up, *pp, *pi ;
Int *Li, *Ui;
mwIndex *Lp1, *Li1, *Up1, *Ui1 ;
double *Lx, *Ax, *Ux, *px ;
double *Lx1, *Ux1 ;
double *t1, *t2 ;
mwIndex anrow ;
mwIndex ancol ;
mwIndex lnnz ;
mwIndex unnz ;
mwIndex i ;
mwIndex j ;
mwIndex app_xnnz ;
mwIndex memsize ;
mwIndex result ;
mwIndex *ws; /* workspace */
double *X ; /* space to scatter x */
mwIndex *pinv ; /* column permutation inverse */
if ( nlhs != 3 || nrhs < 3 )
{
//mexErrMsgTxt (" Incorrect number of arguments to sproductmex \n") ;
}
Ap = mxGetJc(prhs[0]) ;
Ai = mxGetIr(prhs[0]) ;
Ax = mxGetPr(prhs[0]) ;
anrow = mxGetM (prhs[0]) ;
ancol = mxGetN (prhs[0]) ;
t1 = mxGetPr (prhs[1]) ;
t2 = mxGetPr (prhs[2]) ;
lnnz = (mwIndex)*t1 ;
unnz = (mwIndex)*t2 ;
/*printf("lnnz=%d unnz=%d\n", lnnz, unnz);*/
BaskerClassicNS::BaskerClassic <mwIndex, double> mybasker;
mybasker.factor(anrow, ancol, lnnz, Ap, Ai, Ax);
mybasker.returnL(&anrow, &lnnz, &Lp, &Li, &Lx);
mybasker.returnU(&anrow, &unnz, &Up, &Ui, &Ux);
mybasker.returnP(&pp);
plhs[0] = mxCreateSparse (anrow, ancol, lnnz+1, mxREAL) ;
Lp1 = mxGetJc (plhs[0]) ;
Li1 = mxGetIr (plhs[0]) ;
Lx1 = mxGetPr (plhs[0]) ;
plhs[1] = mxCreateSparse (anrow, ancol, unnz, mxREAL) ;
Up1 = mxGetJc (plhs[1]) ;
Ui1 = mxGetIr (plhs[1]) ;
Ux1 = mxGetPr (plhs[1]) ;
mwIndex *pp1, *pp2;
double *ppx;
plhs[2] = mxCreateSparse (ancol, ancol, ancol, mxREAL);
pp1 = mxGetJc (plhs[2]);
pp2 = mxGetIr (plhs[2]);
ppx = mxGetPr (plhs[2]);
Lp1[0] = Lp[0];
for ( i = 0 ; i < ancol ; i++)
{
Lp1[i+1] = Lp[i+1];
for ( j = Lp[i] ; j < Lp[i+1] ; j++ )
{
Li1[j] = Li[j];
Lx1[j] = Lx[j];
}
}
Up1[0] = Up[0];
for ( i = 0 ; i < ancol ; i++)
{
Up1[i+1] = Up[i+1];
for ( j = Up[i] ; j < Up[i+1] ; j++ )
{
Ui1[j] = Ui[j];
Ux1[j] = Ux[j];
}
}
//mexPrintf("Perm \n");
for ( i = 0; i < ancol; i++)
{
//mexPrintf("%d ", pp[i]);
pp1[i] = i;
//pp2[i] = i;
pp2[pp[i]] = i ;
ppx[i] = 1;
}
pp1[ancol] = ancol;
mxFree (pp) ;
mxFree (Lp) ;
mxFree (Li) ;
mxFree (Lx) ;
mxFree (Up) ;
mxFree (Ui) ;
mxFree (Ux) ;
}
const char *model_to_matlab_structure(mxArray *plhs[], int num_of_feature, struct svm_model *model)
{
int i, j, n;
double *ptr;
mxArray *return_model, **rhs;
int out_id = 0;
rhs = (mxArray **)mxMalloc(sizeof(mxArray *)*NUM_OF_RETURN_FIELD);
// Parameters
rhs[out_id] = mxCreateDoubleMatrix(5, 1, mxREAL);
ptr = mxGetPr(rhs[out_id]);
ptr[0] = model->param.svm_type;
ptr[1] = model->param.kernel_type;
ptr[2] = model->param.degree;
ptr[3] = model->param.gamma;
ptr[4] = model->param.coef0;
out_id++;
// nr_class
rhs[out_id] = mxCreateDoubleMatrix(1, 1, mxREAL);
ptr = mxGetPr(rhs[out_id]);
ptr[0] = model->nr_class;
out_id++;
// total SV
rhs[out_id] = mxCreateDoubleMatrix(1, 1, mxREAL);
ptr = mxGetPr(rhs[out_id]);
ptr[0] = model->l;
out_id++;
// rho
n = model->nr_class*(model->nr_class-1)/2;
rhs[out_id] = mxCreateDoubleMatrix(n, 1, mxREAL);
ptr = mxGetPr(rhs[out_id]);
for(i = 0; i < n; i++)
ptr[i] = model->rho[i];
out_id++;
// Label
if(model->label)
{
rhs[out_id] = mxCreateDoubleMatrix(model->nr_class, 1, mxREAL);
ptr = mxGetPr(rhs[out_id]);
for(i = 0; i < model->nr_class; i++)
ptr[i] = model->label[i];
}
else
rhs[out_id] = mxCreateDoubleMatrix(0, 0, mxREAL);
out_id++;
// probA, probB
if(model->probA != NULL && model->probB != NULL)
{
rhs[out_id] = mxCreateDoubleMatrix(n, 1, mxREAL);
ptr = mxGetPr(rhs[out_id]);
for(i = 0; i < n; i++)
ptr[i] = model->probA[i];
rhs[out_id+1] = mxCreateDoubleMatrix(n, 1, mxREAL);
ptr = mxGetPr(rhs[out_id+1]);
for(i = 0; i < n; i++)
ptr[i] = model->probB[i];
}
else
{
rhs[out_id] = mxCreateDoubleMatrix(0, 0, mxREAL);
rhs[out_id+1] = mxCreateDoubleMatrix(0, 0, mxREAL);
}
out_id+=2;
// nSV
if(model->nSV)
{
rhs[out_id] = mxCreateDoubleMatrix(model->nr_class, 1, mxREAL);
ptr = mxGetPr(rhs[out_id]);
for(i = 0; i < model->nr_class; i++)
ptr[i] = model->nSV[i];
}
else
rhs[out_id] = mxCreateDoubleMatrix(0, 0, mxREAL);
out_id++;
// sv_coef
rhs[out_id] = mxCreateDoubleMatrix(model->l, model->nr_class-1, mxREAL);
ptr = mxGetPr(rhs[out_id]);
for(i = 0; i < model->nr_class-1; i++)
for(j = 0; j < model->l; j++)
ptr[(i*(model->l))+j] = model->sv_coef[i][j];
out_id++;
// SVs
{
int ir_index, nonzero_element;
mwIndex *ir, *jc;
mxArray *pprhs[1], *pplhs[1];
if(model->param.kernel_type == PRECOMPUTED)
{
nonzero_element = model->l;
num_of_feature = 1;
}
else
{
nonzero_element = 0;
for(i = 0; i < model->l; i++) {
j = 0;
while(model->SV[i][j].index != -1) {
nonzero_element++;
j++;
}
}
}
// SV in column, easier accessing
rhs[out_id] = mxCreateSparse(num_of_feature, model->l, nonzero_element, mxREAL);
ir = mxGetIr(rhs[out_id]);
jc = mxGetJc(rhs[out_id]);
ptr = mxGetPr(rhs[out_id]);
ir_index = jc[0] = 0;
for(i = 0;i < model->l; i++)
{
if(model->param.kernel_type == PRECOMPUTED)
{
// make a (1 x model->l) matrix
ir[ir_index] = 0;
ptr[ir_index] = model->SV[i][0].value;
ir_index++;
jc[i+1] = jc[i] + 1;
}
else
{
int x_index = 0;
while (model->SV[i][x_index].index != -1)
{
ir[ir_index] = model->SV[i][x_index].index - 1;
ptr[ir_index] = model->SV[i][x_index].value;
ir_index++, x_index++;
}
jc[i+1] = jc[i] + x_index;
}
}
// transpose back to SV in row
pprhs[0] = rhs[out_id];
if(mexCallMATLAB(1, pplhs, 1, pprhs, "transpose"))
return "cannot transpose SV matrix";
rhs[out_id] = pplhs[0];
out_id++;
}
/* Create a struct matrix contains NUM_OF_RETURN_FIELD fields */
return_model = mxCreateStructMatrix(1, 1, NUM_OF_RETURN_FIELD, field_names);
/* Fill struct matrix with input arguments */
for(i = 0; i < NUM_OF_RETURN_FIELD; i++)
mxSetField(return_model,0,field_names[i],mxDuplicateArray(rhs[i]));
/* return */
plhs[0] = return_model;
mxFree(rhs);
return NULL;
}
void mexFunction(
int nargout,
mxArray *out[],
int nargin,
const mxArray *in[]
)
{
/* declare variables */
int nr, nc, np, total;
int i, j, k, ix, iy, jx, jy, ii, jj, iip1, jjp1, iip2, jjp2, step;
double sigma, di, dj, a, z, maxori, phase1, phase2, slope;
// int *ir, *jc;
mwIndex*ir,*jc;
unsigned int *pi, *pj;
double *emag, *ephase, *w;
/* check argument */
if (nargin<4) {
mexErrMsgTxt("Four input arguments required");
}
if (nargout>1) {
mexErrMsgTxt("Too many output arguments");
}
/* get edgel information */
nr = mxGetM(in[0]);
nc = mxGetN(in[0]);
if ( nr*nc ==0 || nr != mxGetM(in[1]) || nc != mxGetN(in[1]) ) {
mexErrMsgTxt("Edge magnitude and phase shall be of the same image size");
}
emag = mxGetPr(in[0]);
ephase = mxGetPr(in[1]);
np = nr * nc;
/* get new index pair */
if (!mxIsUint32(in[2]) | !mxIsUint32(in[3])) {
mexErrMsgTxt("Index pair shall be of type UINT32");
}
if (mxGetM(in[3]) * mxGetN(in[3]) != np + 1) {
mexErrMsgTxt("Wrong index representation");
}
pi = (unsigned int*)mxGetData(in[2]); /* pointer to w_i */
pj = (unsigned int*)mxGetData(in[3]); /* pointer to w_j */
/* create output */
out[0] = mxCreateSparse(np,np,pj[np],mxREAL);
if (out[0]==NULL) {
mexErrMsgTxt("Not enough memory for the output matrix");
}
w = mxGetPr(out[0]);
ir = mxGetIr(out[0]); /* */
jc = mxGetJc(out[0]);
/* find my sigma */
if (nargin<5) {
sigma = 0;
for (k=0; k<np; k++) {
if (emag[k]>sigma) { sigma = emag[k]; }
}
sigma = sigma / 6;
printf("sigma = %6.5f",sigma);
} else {
sigma = mxGetScalar(in[4]);
}
a = 0.5 / (sigma * sigma);
/* computation */
total = 0;
for (j=0; j<np; j++) {
jc[j] = total;
jx = j / nr; /* col */
jy = j % nr; /* row */
for (k=pj[j]; k<pj[j+1]; k++) {
i = pi[k];
if (i==j) {
maxori = 1;
} else {
ix = i / nr;
iy = i % nr;
/* scan */
di = (double) (iy - jy);
dj = (double) (ix - jx);
maxori = 0.;
phase1 = ephase[j];
/* sample in i direction */
if (abs(di) >= abs(dj)) {
slope = dj / di;
step = (iy>=jy) ? 1 : -1;
iip1 = jy;
jjp1 = jx;
for (ii=0;ii<abs(di);ii++){
iip2 = iip1 + step;
jjp2 = (int)(0.5 + slope*(iip2-jy) + jx);
phase2 = ephase[iip2+jjp2*nr];
if (phase1 != phase2) {
z = (emag[iip1+jjp1*nr] + emag[iip2+jjp2*nr]);
if (z > maxori){
maxori = z;
}
}
iip1 = iip2;
jjp1 = jjp2;
phase1 = phase2;
}
/* sample in j direction */
} else {
slope = di / dj;
step = (ix>=jx) ? 1: -1;
jjp1 = jx;
iip1 = jy;
for (jj=0;jj<abs(dj);jj++){
jjp2 = jjp1 + step;
iip2 = (int)(0.5+ slope*(jjp2-jx) + jy);
phase2 = ephase[iip2+jjp2*nr];
if (phase1 != phase2){
z = (emag[iip1+jjp1*nr] + emag[iip2+jjp2*nr]);
if (z > maxori){
maxori = z;
}
}
iip1 = iip2;
jjp1 = jjp2;
phase1 = phase2;
}
}
maxori = 0.5 * maxori;
//maxori = exp(-maxori * maxori * a); //segmentation
maxori = 1-exp(-maxori * maxori * a); //symmetry
}
ir[total] = i;
if (maxori>0.2) w[total] = maxori; else w[total]=0.2; //symmetry
//w[total] = maxori; //segmentation
total = total + 1;
} /* i */
} /* j */
jc[np] = total;
}
void mexFunction
(
/* === Parameters ======================================================= */
int nlhs, /* number of left-hand sides */
mxArray *plhs [], /* left-hand side matrices */
int nrhs, /* number of right--hand sides */
const mxArray *prhs [] /* right-hand side matrices */
)
{
Zmat A;
char *fname, rhstyp[4], title[81], key[9], type[4];
int i,j,k,l, mynrhs, status;
integer nrhsix, m, *rhsptr=NULL, *rhsind=NULL, ierr, nr,nc,nz, tmp,tmp0,tmp2,tmp3,ncolumns;
size_t mrows, ncols, sizebuf;
mwSize nnz, buflen;
mwIndex *irs, *jcs;
double *pr, *pi, *sr, *si;
doublecomplex *sol, *rhs=NULL, *rhsval=NULL;
mxArray *fout, *f_input;
FILE *fp;
if (nrhs!=1)
mexErrMsgTxt("Only one input argument required.");
else if (nlhs!=3)
mexErrMsgTxt("Three output arguments are required.");
else if (mxGetClassID(prhs[0])!=mxCHAR_CLASS)
mexErrMsgTxt("Input must be a string.");
/* get filename */
f_input = (mxArray *) prhs[0];
mrows = mxGetM (f_input) ;
ncols = mxGetN (f_input) ;
/* Get the length of the input string. */
buflen = (mrows*ncols) + 1;
/* Allocate memory for input and output strings. */
fname = (char *) mxCalloc((size_t)buflen, (size_t)sizeof(char));
/* Copy the string data from tmp into a C string pdata */
status = mxGetString(f_input, fname, buflen);
ncolumns = 0;
tmp0 = 0;
if ((fp=fopen(fname,"r"))==NULL) {
mexPrintf(" file %s",fname);
mexErrMsgTxt(" not found");
return;
}
fclose(fp);
rhstyp[0]=' ';
rhstyp[1]=' ';
rhstyp[2]=' ';
A.nc=0;
Zreadmtc(&tmp0,&tmp0,&tmp0,fname,A.a,A.ja,A.ia,
rhs,&ncolumns,rhstyp,&nr,&nc,&nz,title,key,type,
&nrhsix,rhsptr,rhsind,rhsval,&ierr,strlen(fname),2,72,8,3);
ncols=ncolumns;
/* if a right hand side is given, then use these */
if (ierr) {
mexPrintf(" ierr = %d\n",ierr);
mexErrMsgTxt(" error in reading the matrix, stop.\n");
switch(ierr) {
case 1:
mexErrMsgTxt("too many columns\n");
break;
case 2:
mexErrMsgTxt("too many nonzeros\n");
break;
case 3:
mexErrMsgTxt("too many columns and nonzeros\n");
break;
case 4:
mexErrMsgTxt("right hand side has incompatible type\n");
break;
case 5:
mexErrMsgTxt("too many right hand side entries\n");
break;
case 6:
mexErrMsgTxt("wrong type (real/complex)\n");
break;
}
exit(ierr);
}
if (ncols>0) {
m=1;
if (rhstyp[1]=='G' || rhstyp[1]=='g') {
m++;
}
if (rhstyp[2]=='X' || rhstyp[2]=='x') {
m++;
}
}
else
m=0;
m*=nr*ncols;
rhsptr=NULL;
rhsind=NULL;
rhsval=NULL;
if (ncols!=0 && (rhstyp[0]=='M' || rhstyp[0]=='m')) {
rhsptr=(integer *) MAlloc((size_t)(ncols+1)*sizeof(integer),"Zloadhbo:rhsptr");
rhsind=(integer *) MAlloc((size_t)nrhsix*sizeof(integer), "Zloadhbo:rhsind");
rhsval=(doublecomplex *) MAlloc((size_t)nrhsix*sizeof(doublecomplex), "Zloadhbo:rhsval");
}
A.ia=(integer *) MAlloc((size_t)(nc+1)*sizeof(integer), "Zloadhbo:A.ia");
A.ja=(integer *) MAlloc((size_t)nz*sizeof(integer), "Zloadhbo:A.ja");
A.a =(doublecomplex*)MAlloc((size_t)nz*sizeof(doublecomplex),"Zloadhbo:A.a");
A.nr=nr;
A.nc=nc;
rhs =(doublecomplex *)MAlloc((size_t)m*sizeof(doublecomplex),"Zloadhbo:rhs");
/* advance pointer to reserve space when uncompressing the right hand side */
if (ncols!=0 && (rhstyp[0]=='M' || rhstyp[0]=='m'))
rhs+=nr*ncols;
tmp = 3;
tmp2 = nc;
tmp3 = nz;
if (ncols!=0 && (rhstyp[0]=='M' || rhstyp[0]=='m'))
m-=nr*ncols;
Zreadmtc(&tmp2,&tmp3,&tmp,fname,A.a,A.ja,A.ia,
rhs,&m,rhstyp,&nr,&nc,&nz,title,key,type,
&nrhsix,rhsptr,rhsind,rhsval,&ierr,strlen(fname),2,72,8,3);
if (ncols!=0 && (rhstyp[0]=='M' || rhstyp[0]=='m'))
m+=nr*ncols;
if (ierr) {
mexPrintf(" ierr = %d\n",ierr);
mexErrMsgTxt(" error in reading the matrix, stop.\n");
return;
}
/* set output parameters */
nlhs=3;
fout =mxCreateSparse((mwSize)nr,(mwSize)nc, (mwSize)nz, mxCOMPLEX);
plhs[0]=fout;
sr = (double *) mxGetPr(fout);
si = (double *) mxGetPi(fout);
irs = (mwIndex *) mxGetIr(fout);
jcs = (mwIndex *) mxGetJc(fout);
jcs[0]=0;
for (i=0; i<nc; i++) {
jcs[i+1]=A.ia[i+1]-1;
for (j=A.ia[i]-1; j<A.ia[i+1]-1; j++) {
irs[j]=A.ja[j]-1;
sr[j] =A.a[j].r;
si[j] =A.a[j].i;
}
}
if (ncols>0) {
m=1;
if (rhstyp[1]=='G' || rhstyp[1]=='g') {
m++;
}
if (rhstyp[2]=='X' || rhstyp[2]=='x') {
m++;
}
}
else
m=0;
fout=mxCreateDoubleMatrix((mwSize)nr,(mwSize)m*ncols, mxCOMPLEX);
plhs[1]=fout;
if (ncols!=0 && (rhstyp[0]=='M' || rhstyp[0]=='m')) {
}
else {
pr = mxGetPr(fout);
pi = mxGetPi(fout);
for (i=0; i<nr*m*ncols; i++) {
pr[i]=rhs[i].r;
pi[i]=rhs[i].i;
}
}
rhstyp[3]='\0';
plhs[2] = mxCreateString(rhstyp);
free(A.a);
free(A.ia);
free(A.ja);
if (ncols>0)
free(rhs);
if (ncols!=0 && (rhstyp[0]=='M' || rhstyp[0]=='m')) {
free(rhsptr);
free(rhsind);
free(rhsval);
}
return;
}
Need ompcore(double D[], double x[], double DtX[], double XtX[], double G[], mwSize n, mwSize m, mwSize L,
int T, double eps, int gamma_mode, int profile, double msg_delta, int erroromp)
{
profdata pd;
/* mxArray *Gamma;*/
mwIndex i, j, signum, pos, *ind, *gammaIr, *gammaJc, gamma_count;
mwSize allocated_coefs, allocated_cols;
int DtX_specified, XtX_specified, batchomp, standardomp, *selected_atoms,*times_atoms ;
double *alpha, *r, *Lchol, *c, *Gsub, *Dsub, sum, *gammaPr, *tempvec1, *tempvec2;
double eps2, resnorm, delta, deltaprev, secs_remain;
int mins_remain, hrs_remain;
clock_t lastprint_time, starttime;
Need my;
/*** status flags ***/
DtX_specified = (DtX!=0); /* indicates whether D'*x was provided */
XtX_specified = (XtX!=0); /* indicates whether sum(x.*x) was provided */
standardomp = (G==0); /* batch-omp or standard omp are selected depending on availability of G */
batchomp = !standardomp;
/*** allocate output matrix ***/
if (gamma_mode == FULL_GAMMA) {
/* allocate full matrix of size m X L */
Gamma = mxCreateDoubleMatrix(m, L, mxREAL);
gammaPr = mxGetPr(Gamma);
gammaIr = 0;
gammaJc = 0;
}
else {
/* allocate sparse matrix with room for allocated_coefs nonzeros */
/* for error-omp, begin with L*sqrt(n)/2 allocated nonzeros, otherwise allocate L*T nonzeros */
allocated_coefs = erroromp ? (mwSize)(ceil(L*sqrt((double)n)/2.0) + 1.01) : L*T;
Gamma = mxCreateSparse(m, L, allocated_coefs, mxREAL);
gammaPr = mxGetPr(Gamma);
gammaIr = mxGetIr(Gamma);
gammaJc = mxGetJc(Gamma);
gamma_count = 0;
gammaJc[0] = 0;
}
/*** helper arrays ***/
alpha = (double*)mxMalloc(m*sizeof(double)); /* contains D'*residual */
ind = (mwIndex*)mxMalloc(n*sizeof(mwIndex)); /* indices of selected atoms */
selected_atoms = (int*)mxMalloc(m*sizeof(int)); /* binary array with 1's for selected atoms */
times_atoms = (int*)mxMalloc(m*sizeof(int));
c = (double*)mxMalloc(n*sizeof(double)); /* orthogonal projection result */
/* current number of columns in Dsub / Gsub / Lchol */
allocated_cols = erroromp ? (mwSize)(ceil(sqrt((double)n)/2.0) + 1.01) : T;
/* Cholesky decomposition of D_I'*D_I */
Lchol = (double*)mxMalloc(n*allocated_cols*sizeof(double));
/* temporary vectors for various computations */
tempvec1 = (double*)mxMalloc(m*sizeof(double));
tempvec2 = (double*)mxMalloc(m*sizeof(double));
if (batchomp) {
/* matrix containing G(:,ind) - the columns of G corresponding to the selected atoms, in order of selection */
Gsub = (double*)mxMalloc(m*allocated_cols*sizeof(double));
}
else {
/* matrix containing D(:,ind) - the selected atoms from D, in order of selection */
Dsub = (double*)mxMalloc(n*allocated_cols*sizeof(double));
/* stores the residual */
r = (double*)mxMalloc(n*sizeof(double));
}
if (!DtX_specified) {
/* contains D'*x for the current signal */
DtX = (double*)mxMalloc(m*sizeof(double));
}
/*** initializations for error omp ***/
if (erroromp) {
eps2 = eps*eps; /* compute eps^2 */
if (T<0 || T>n) { /* unspecified max atom num - set max atoms to n */
T = n;
}
}
/*** initialize timers ***/
initprofdata(&pd); /* initialize profiling counters */
starttime = clock(); /* record starting time for eta computations */
lastprint_time = starttime; /* time of last status display */
/********************** perform omp for each signal **********************/
for (signum=0; signum<L; ++signum) {
/* initialize residual norm and deltaprev for error-omp */
if (erroromp) {
if (XtX_specified) {
resnorm = XtX[signum];
}
else {
resnorm = dotprod(x+n*signum, x+n*signum, n);
addproftime(&pd, XtX_TIME);
}
deltaprev = 0; /* delta tracks the value of gamma'*G*gamma */
}
else {
/* ignore residual norm stopping criterion */
eps2 = 0;
resnorm = 1;
}
if (resnorm>eps2 && T>0) {
/* compute DtX */
if (!DtX_specified) {
matT_vec(1, D, x+n*signum, DtX, n, m);
addproftime(&pd, DtX_TIME);
}
/* initialize alpha := DtX */
memcpy(alpha, DtX + m*signum*DtX_specified, m*sizeof(double));
/* mark all atoms as unselected */
for (i=0; i<m; ++i) {
selected_atoms[i] = 0;
}
for (i=0; i<m; ++i) {
times_atoms[i] = 0;
}
}
/* main loop */
i=0;
while (resnorm>eps2 && i<T) {
/* index of next atom */
pos = maxabs(alpha, m);
addproftime(&pd, MAXABS_TIME);
/* stop criterion: selected same atom twice, or inner product too small */
if (selected_atoms[pos] || alpha[pos]*alpha[pos]<1e-14) {
break;
}
/* mark selected atom */
ind[i] = pos;
selected_atoms[pos] = 1;
times_atoms[pos]++;
/* matrix reallocation */
if (erroromp && i>=allocated_cols) {
allocated_cols = (mwSize)(ceil(allocated_cols*MAT_INC_FACTOR) + 1.01);
Lchol = (double*)mxRealloc(Lchol,n*allocated_cols*sizeof(double));
batchomp ? (Gsub = (double*)mxRealloc(Gsub,m*allocated_cols*sizeof(double))) :
(Dsub = (double*)mxRealloc(Dsub,n*allocated_cols*sizeof(double))) ;
}
/* append column to Gsub or Dsub */
if (batchomp) {
memcpy(Gsub+i*m, G+pos*m, m*sizeof(double));
}
else {
memcpy(Dsub+i*n, D+pos*n, n*sizeof(double));
}
/*** Cholesky update ***/
if (i==0) {
*Lchol = 1;
}
else {
/* incremental Cholesky decomposition: compute next row of Lchol */
if (standardomp) {
matT_vec(1, Dsub, D+n*pos, tempvec1, n, i); /* compute tempvec1 := Dsub'*d where d is new atom */
addproftime(&pd, DtD_TIME);
}
else {
vec_assign(tempvec1, Gsub+i*m, ind, i); /* extract tempvec1 := Gsub(ind,i) */
}
backsubst('L', Lchol, tempvec1, tempvec2, n, i); /* compute tempvec2 = Lchol \ tempvec1 */
for (j=0; j<i; ++j) { /* write tempvec2 to end of Lchol */
Lchol[j*n+i] = tempvec2[j];
}
/* compute Lchol(i,i) */
sum = 0;
for (j=0; j<i; ++j) { /* compute sum of squares of last row without Lchol(i,i) */
sum += SQR(Lchol[j*n+i]);
}
if ( (1-sum) <= 1e-14 ) { /* Lchol(i,i) is zero => selected atoms are dependent */
break;
}
Lchol[i*n+i] = sqrt(1-sum);
}
addproftime(&pd, LCHOL_TIME);
i++;
/* perform orthogonal projection and compute sparse coefficients */
vec_assign(tempvec1, DtX + m*signum*DtX_specified, ind, i); /* extract tempvec1 = DtX(ind) */
cholsolve('L', Lchol, tempvec1, c, n, i); /* solve LL'c = tempvec1 for c */
addproftime(&pd, COMPCOEF_TIME);
/* update alpha = D'*residual */
if (standardomp) {
mat_vec(-1, Dsub, c, r, n, i); /* compute r := -Dsub*c */
vec_sum(1, x+n*signum, r, n); /* compute r := x+r */
/*memcpy(r, x+n*signum, n*sizeof(double)); /* assign r := x */
/*mat_vec1(-1, Dsub, c, 1, r, n, i); /* compute r := r-Dsub*c */
addproftime(&pd, COMPRES_TIME);
matT_vec(1, D, r, alpha, n, m); /* compute alpha := D'*r */
addproftime(&pd, DtR_TIME);
/* update residual norm */
if (erroromp) {
resnorm = dotprod(r, r, n);
addproftime(&pd, UPDATE_RESNORM_TIME);
}
}
else {
mat_vec(1, Gsub, c, tempvec1, m, i); /* compute tempvec1 := Gsub*c */
memcpy(alpha, DtX + m*signum*DtX_specified, m*sizeof(double)); /* set alpha = D'*x */
vec_sum(-1, tempvec1, alpha, m); /* compute alpha := alpha - tempvec1 */
addproftime(&pd, UPDATE_DtR_TIME);
/* update residual norm */
if (erroromp) {
vec_assign(tempvec2, tempvec1, ind, i); /* assign tempvec2 := tempvec1(ind) */
delta = dotprod(c,tempvec2,i); /* compute c'*tempvec2 */
resnorm = resnorm - delta + deltaprev; /* residual norm update */
deltaprev = delta;
addproftime(&pd, UPDATE_RESNORM_TIME);
}
}
}
/*** generate output vector gamma ***/
if (gamma_mode == FULL_GAMMA) { /* write the coefs in c to their correct positions in gamma */
for (j=0; j<i; ++j) {
gammaPr[m*signum + ind[j]] = c[j];
}
}
else {
/* sort the coefs by index before writing them to gamma */
quicksort(ind,c,i);
addproftime(&pd, INDEXSORT_TIME);
/* gamma is full - reallocate */
if (gamma_count+i >= allocated_coefs) {
while(gamma_count+i >= allocated_coefs) {
allocated_coefs = (mwSize)(ceil(GAMMA_INC_FACTOR*allocated_coefs) + 1.01);
}
mxSetNzmax(Gamma, allocated_coefs);
mxSetPr(Gamma, mxRealloc(gammaPr, allocated_coefs*sizeof(double)));
mxSetIr(Gamma, mxRealloc(gammaIr, allocated_coefs*sizeof(mwIndex)));
gammaPr = mxGetPr(Gamma);
gammaIr = mxGetIr(Gamma);
}
/* append coefs to gamma and update the indices */
for (j=0; j<i; ++j) {
gammaPr[gamma_count] = c[j];
gammaIr[gamma_count] = ind[j];
gamma_count++;
}
gammaJc[signum+1] = gammaJc[signum] + i;
}
/*** display status messages ***/
if (msg_delta>0 && (clock()-lastprint_time)/(double)CLOCKS_PER_SEC >= msg_delta)
{
lastprint_time = clock();
/* estimated remainig time */
secs2hms( ((L-signum-1)/(double)(signum+1)) * ((lastprint_time-starttime)/(double)CLOCKS_PER_SEC) ,
&hrs_remain, &mins_remain, &secs_remain);
mexPrintf("omp: signal %d / %d, estimated remaining time: %02d:%02d:%05.2f\n",
signum+1, L, hrs_remain, mins_remain, secs_remain);
mexEvalString("drawnow;");
}
}
/* end omp */
/*** print final messages ***/
if (msg_delta>0) {
mexPrintf("omp: signal %d / %d\n", signum, L);
}
if (profile) {
printprofinfo(&pd, erroromp, batchomp, L);
}
/* free memory */
if (!DtX_specified) {
mxFree(DtX);
}
if (standardomp) {
mxFree(r);
mxFree(Dsub);
}
else {
mxFree(Gsub);
}
mxFree(tempvec2);
mxFree(tempvec1);
mxFree(Lchol);
mxFree(c);
mxFree(selected_atoms);
mxFree(ind);
mxFree(alpha);
my.qGamma=Gamma;
my.qtimes__atoms=times__atoms;
/*return Gamma;*/
return my;
}
/* read in a problem (in svmlight format) */
void read_problem(const char *filename, mxArray *plhs[])
{
int elements, max_index, min_index, i, k;
FILE *fp = fopen(filename,"r");
int l = 0;
mwIndex *ir, *jc;
/* double *labels, *samples; */
double *samples;
if(fp == NULL)
{
mexPrintf("can't open input file %s\n",filename);
fake_answer(plhs);
return;
}
max_index = 0;
min_index = 1; /* our index starts from 1 */
elements = 0;
int index;
double value;
while(1)
{
/* label */
/* int index;
double value;
fscanf(fp,"%lf",&value); */
/* features */
while(1)
{
int c;
do {
c = getc(fp);
if(c=='\n') goto out;
if(c==EOF) goto eof;
} while(isspace(c));
ungetc(c,fp);
fscanf(fp,"%d:%lf",&index, &value);
if (index < min_index)
min_index = index;
elements++;
}
out:
if(index > max_index)
max_index = index;
l++;
}
eof:
rewind(fp);
/* y */
/* plhs[0] = mxCreateDoubleMatrix(l, 1, mxREAL); */
/* x^T */
if (min_index <= 0)
/* plhs[1] = mxCreateSparse(max_index-min_index+1, l, elements, mxREAL); */
plhs[0] = mxCreateSparse(max_index-min_index+1, l, elements, mxREAL);
else
/* plhs[1] = mxCreateSparse(max_index, l, elements, mxREAL); */
plhs[0] = mxCreateSparse(max_index, l, elements, mxREAL);
/* labels = mxGetPr(plhs[0]); */
/* samples = mxGetPr(plhs[1]); */
samples = mxGetPr(plhs[0]);
/*
ir = mxGetIr(plhs[1]);
jc = mxGetJc(plhs[1]); */
ir = mxGetIr(plhs[0]);
jc = mxGetJc(plhs[0]);
k=0;
for(i=0;i<l;i++)
{
jc[i] = k;
/* fscanf(fp,"%lf",&labels[i]); */
while(1)
{
int c, index;
do {
c = getc(fp);
if(c=='\n') goto out2;
} while(isspace(c));
ungetc(c,fp);
fscanf(fp,"%d:%lf",&index,&samples[k]);
ir[k] = index - min_index; /* precomputed kernel has <index> start from 0 */
++k;
}
out2:
;
}
jc[l] = k;
fclose(fp);
/*
{
mxArray *rhs[1], *lhs[1];
rhs[0] = plhs[0];
if(mexCallMATLAB(1, lhs, 1, rhs, "transpose"))
{
mexPrintf("Error: cannot transpose problem\n");
return;
}
plhs[0] = lhs[0];
}
*/
}
/* ************************************************************
PROCEDURE mexFunction - Entry for Matlab
************************************************************ */
void mexFunction(const int nlhs, mxArray *plhs[],
const int nrhs, const mxArray *prhs[])
{
const mxArray *L_FIELD;
mxArray *myplhs[NPAROUT];
int m, i, j, inz, iwsiz, nsuper, tmpsiz, fwsiz, nskip, nadd, m1;
double *fwork, *d, *skipPr, *orgd;
const double *permPr,*xsuperPr,*Ppr,*absd;
int *perm, *snode, *xsuper, *iwork, *xlindx, *skip, *skipJc;
const int *LINir, *Pjc, *Pir;
double canceltol, maxu, abstol;
jcir L;
char useAbsd, useDelay;
/* ------------------------------------------------------------
Check for proper number of arguments
blkchol(L,P, pars,absd) with nparinmin=2.
------------------------------------------------------------ */
if(nrhs < NPARINMIN)
mexErrMsgTxt("blkchol requires more input arguments");
if(nlhs > NPAROUT)
mexErrMsgTxt("blkchol produces less output arguments");
/* ------------------------------------------------------------
Get input matrix P to be factored.
------------------------------------------------------------ */
if( (m = mxGetM(P_IN)) != mxGetN(P_IN))
mexErrMsgTxt("P must be square");
if(!mxIsSparse(P_IN))
mexErrMsgTxt("P must be sparse");
Pjc = mxGetJc(P_IN);
Pir = mxGetIr(P_IN);
Ppr = mxGetPr(P_IN);
/* ------------------------------------------------------------
Disassemble block Cholesky structure L
------------------------------------------------------------ */
if(!mxIsStruct(L_IN))
mexErrMsgTxt("Parameter `L' should be a structure.");
if( (L_FIELD = mxGetField(L_IN,0,"perm")) == NULL) /* L.perm */
mexErrMsgTxt("Missing field L.perm.");
if(m != mxGetM(L_FIELD) * mxGetN(L_FIELD))
mexErrMsgTxt("perm size mismatch");
permPr = mxGetPr(L_FIELD);
if( (L_FIELD = mxGetField(L_IN,0,"L")) == NULL) /* L.L */
mexErrMsgTxt("Missing field L.L.");
if( m != mxGetM(L_FIELD) || m != mxGetN(L_FIELD) )
mexErrMsgTxt("Size L.L mismatch.");
if(!mxIsSparse(L_FIELD))
mexErrMsgTxt("L.L should be sparse.");
L.jc = mxGetJc(L_FIELD);
LINir = mxGetIr(L_FIELD);
if( (L_FIELD = mxGetField(L_IN,0,"xsuper")) == NULL) /* L.xsuper */
mexErrMsgTxt("Missing field L.xsuper.");
nsuper = mxGetM(L_FIELD) * mxGetN(L_FIELD) - 1;
if( nsuper > m )
mexErrMsgTxt("Size L.xsuper mismatch.");
xsuperPr = mxGetPr(L_FIELD);
if( (L_FIELD = mxGetField(L_IN,0,"tmpsiz")) == NULL) /* L.tmpsiz */
mexErrMsgTxt("Missing field L.tmpsiz.");
tmpsiz = mxGetScalar(L_FIELD);
/* ------------------------------------------------------------
Disassemble pars structure: canceltol, maxu
------------------------------------------------------------ */
canceltol = 1E-15; /* supply with defaults */
maxu = 5E5;
abstol = 1E-20;
useAbsd = 0;
useDelay = 0;
if(nrhs >= NPARINMIN + 1){ /* 3rd argument = pars */
if(!mxIsStruct(PARS_IN))
mexErrMsgTxt("Parameter `pars' should be a structure.");
if( (L_FIELD = mxGetField(PARS_IN,0,"canceltol")) != NULL)
canceltol = mxGetScalar(L_FIELD); /* pars.canceltol */
if( (L_FIELD = mxGetField(PARS_IN,0,"maxu")) != NULL)
maxu = mxGetScalar(L_FIELD); /* pars.maxu */
if( (L_FIELD = mxGetField(PARS_IN,0,"abstol")) != NULL){
abstol = mxGetScalar(L_FIELD); /* pars.abstol */
abstol = MAX(abstol, 0.0);
}
if( (L_FIELD = mxGetField(PARS_IN,0,"delay")) != NULL)
useDelay = mxGetScalar(L_FIELD); /* pars.delay */
/* ------------------------------------------------------------
Get optional vector absd
------------------------------------------------------------ */
if(nrhs >= NPARIN){
useAbsd = 1;
absd = mxGetPr(ABSD_IN);
if(m != mxGetM(ABSD_IN) * mxGetN(ABSD_IN))
mexErrMsgTxt("absd size mismatch");
}
}
/* ------------------------------------------------------------
Create sparse output matrix L(m x m).
------------------------------------------------------------ */
L_OUT = mxCreateSparse(m,m, L.jc[m],mxREAL);
L.ir = mxGetIr(L_OUT);
L.pr = mxGetPr(L_OUT);
memcpy(mxGetJc(L_OUT), L.jc, (m+1) * sizeof(int));
memcpy(L.ir, LINir, L.jc[m] * sizeof(int));
/* ------------------------------------------------------------
Create ouput vector d(m).
------------------------------------------------------------ */
D_OUT = mxCreateDoubleMatrix(m,1,mxREAL);
d = mxGetPr(D_OUT);
/* ------------------------------------------------------------
Compute required sizes of working arrays:
iwsiz = 2*(m + nsuper).
fwsiz = tmpsiz.
------------------------------------------------------------ */
iwsiz = MAX(2*(m+nsuper), 1);
fwsiz = MAX(tmpsiz, 1);
/* ------------------------------------------------------------
Allocate working arrays:
integer: perm(m), snode(m), xsuper(nsuper+1),
iwork(iwsiz), xlindx(m+1), skip(m),
double: orgd(m), fwork(fwsiz).
------------------------------------------------------------ */
m1 = MAX(m,1); /* avoid alloc to 0 */
perm = (int *) mxCalloc(m1,sizeof(int));
snode = (int *) mxCalloc(m1,sizeof(int));
xsuper = (int *) mxCalloc(nsuper+1,sizeof(int));
iwork = (int *) mxCalloc(iwsiz,sizeof(int));
xlindx = (int *) mxCalloc(m+1,sizeof(int));
skip = (int *) mxCalloc(m1, sizeof(int));
orgd = (double *) mxCalloc(m1,sizeof(double));
fwork = (double *) mxCalloc(fwsiz,sizeof(double));
/* ------------------------------------------------------------
Convert PERM, XSUPER to integer and C-Style
------------------------------------------------------------ */
for(i = 0; i < m; i++){
j = permPr[i];
perm[i] = --j;
}
for(i = 0; i <= nsuper; i++){
j = xsuperPr[i];
xsuper[i] = --j;
}
/* ------------------------------------------------------------
Let L = tril(P(PERM,PERM)), uses orgd(m) as temp working storage.
------------------------------------------------------------ */
permuteP(L.jc,L.ir,L.pr, Pjc,Pir,Ppr, perm, orgd, m);
/* ------------------------------------------------------------
If no orgd has been supplied, take orgd = diag(L on input)
Otherwise, let orgd = absd(perm).
------------------------------------------------------------ */
if(useAbsd)
for(j = 0; j < m; j++)
orgd[j] = absd[perm[j]];
else
for(j = 0; j < m; j++)
orgd[j] = L.pr[L.jc[j]];
/* ------------------------------------------------------------
Create "snode" and "xlindx"; change L.ir to the compact subscript
array (with xlindx), and do BLOCK SPARSE CHOLESKY.
------------------------------------------------------------ */
nskip = spchol(m, nsuper, xsuper, snode, xlindx,
L.ir, orgd, L.jc, L.pr, d, perm, abstol,
canceltol, maxu, skip, &nadd, iwsiz, iwork, fwsiz, fwork);
if(nskip < 0)
mexErrMsgTxt("Insufficient workspace in pblkchol");
/* ------------------------------------------------------------
Copy original row-indices from LINir to L.ir.
------------------------------------------------------------ */
memcpy(L.ir, LINir, L.jc[m] * sizeof(int));
/* ------------------------------------------------------------
Create output matrices skip = sparse([],[],[],m,1,nskip),
diagadd = sparse([],[],[],m,1,nadd),
------------------------------------------------------------ */
SKIP_OUT = mxCreateSparse(m,1, MAX(1,nskip),mxREAL);
memcpy(mxGetIr(SKIP_OUT), skip, nskip * sizeof(int));
skipJc = mxGetJc(SKIP_OUT);
skipJc[0] = 0; skipJc[1] = nskip;
skipPr = mxGetPr(SKIP_OUT);
/* ------------------------------------------------------------
useDelay = 1 then L(:,i) is i-th column before ith pivot; useful
for pivot-delaying strategy. (Fwslv(L, L(:,i)) still required.)
------------------------------------------------------------ */
if(useDelay == 1)
for(j = 0; j < nskip; j++)
skipPr[j] = 1.0;
else
for(j = 0; j < nskip; j++){
i = skip[j];
skipPr[j] = L.pr[L.jc[i]]; /* Set skipped l(:,i)=ei. */
L.pr[L.jc[i]] = 1.0;
fzeros(L.pr+L.jc[i]+1,L.jc[i+1]-L.jc[i]-1);
}
DIAGADD_OUT = mxCreateSparse(m,1, MAX(1,nadd),mxREAL);
memcpy(mxGetIr(DIAGADD_OUT), iwork, nadd * sizeof(int));
skipJc = mxGetJc(DIAGADD_OUT);
skipJc[0] = 0; skipJc[1] = nadd;
skipPr = mxGetPr(DIAGADD_OUT);
for(j = 0; j < nadd; j++)
skipPr[j] = orgd[iwork[j]];
/* ------------------------------------------------------------
Release working arrays.
------------------------------------------------------------ */
mxFree(fwork);
mxFree(orgd);
mxFree(skip);
mxFree(xlindx);
mxFree(iwork);
mxFree(xsuper);
mxFree(snode);
mxFree(perm);
/* ------------------------------------------------------------
Copy requested output parameters (at least 1), release others.
------------------------------------------------------------ */
i = MAX(nlhs, 1);
memcpy(plhs,myplhs, i * sizeof(mxArray *));
for(; i < NPAROUT; i++)
mxDestroyArray(myplhs[i]);
}
void mexFunction(int nlhs, mxArray *plhs[], int nrhs, const mxArray*prhs[]) {
double *Pw_d, *pr_x, *Pw_z, *Pz_d;
double beta;
int *ir_x, prev_ir_x,*ir_Pw_d, *jc_x, *jc_Pw_d, *bad_index;
int element_count, nTopics;
int nn,ii,jj;
unsigned int nWords, nDocs, indc, indx, indy, next_indx;
bool temperedEM;
/* sanity check of the parameters */
if ((!(mxIsSparse(prhs[0])))
||(!((nrhs==3)||(nrhs==4)))) {
printf("usage : C = mex_Estep(X,Pw_z,Pz_d)\n");
printf(" or : C = mex_Estep(X,Pw_z,Pz_d,beta)\n");
printf(" computes the normalization oonstant 'C' only for those values\n");
printf(" which are needed (nonzeros of X). If no 'beta' is given this function\n");
printf(" does plain EM, TEM otherwise\n");
return;
}
nWords = mxGetM(prhs[0]);
nDocs = mxGetN(prhs[0]);
nTopics = mxGetM(prhs[2]);
jc_x = mxGetJc(prhs[0]);
ir_x = mxGetIr(prhs[0]);
pr_x = mxGetPr(prhs[0]);
Pw_z = mxGetPr(prhs[1]);
Pz_d = mxGetPr(prhs[2]);
if (nrhs == 4)
beta = *mxGetPr(prhs[3]);
else
beta = 1;
temperedEM = beta!=1;
/* sanity check of the parameters */
if ((! ( mxGetM(prhs[1]) == nWords))|| (!(mxGetN(prhs[1]) == nTopics) ) ){
printf("Dimensions mismatch of Pw_z, should be of size %d x %d\n",nWords,nTopics);
return;
}
if ((!(mxGetM(prhs[2])==nTopics))||(!(mxGetN(prhs[2])==nDocs))){
printf("Dimensions mismatch of Pz_d, should be of size %d x %d\n",nTopics,nDocs);
return;
}
/* create sparse output matrix */
plhs[0] = mxCreateSparse(nWords, nDocs, numnonzeros(prhs[0]), mxREAL);
jc_Pw_d = mxGetJc(plhs[0]);
ir_Pw_d = mxGetIr(plhs[0]);
Pw_d = mxGetPr(plhs[0]);
element_count = 0;
jc_Pw_d[0] = 0;
indx = 0;
if (!(temperedEM)){ /* plain EM */
for (indc = 0; indc < 10; indc++) {
jc_Pw_d[indc+1] = jc_Pw_d[indc];
next_indx = jc_x[indc]; /* next column index */
while (indx < next_indx) { /* while nonzero entries in the matrix */
jc_Pw_d[indc]++; /* point to next index */
/* copy element of sparse matrix */
ir_Pw_d[element_count] = ir_x[indx];
ii = indc;
jj = ir_Pw_d[element_count];
Pw_d[element_count] = 0;
/* sum over all topics */
for (nn=0;nn<nTopics;nn++)
Pw_d[element_count] += (Pz_d[ii*nTopics+nn] * Pw_z[jj+nWords*nn]);
printf("%d %d %d %d %f\n",jc_x[indc], ir_x[indx], jc_Pw_d[indc], ir_Pw_d[element_count],Pw_d[element_count]);
indx++;
element_count++;
}
}
}
else { /* tempered EM - see comments in plain EM */
for (indc = 0; indc < nDocs; indc++) {
jc_Pw_d[indc+1] = jc_Pw_d[indc];
next_indx = jc_x[indc+1];
while (indx < next_indx) {
jc_Pw_d[indc+1]++;
ir_Pw_d[element_count] = ir_x[indx];
ii = indc;
jj = ir_Pw_d[element_count];
Pw_d[element_count] = 0;
for (nn=0;nn<nTopics;nn++)
Pw_d[element_count] += pow(Pz_d[ii*nTopics+nn] * Pw_z[jj+nWords*nn],beta);
indx++;
element_count++;
}
}
}
}
void mexFunction(
int nargout,
mxArray *out[],
int nargin,
const mxArray *in[]
)
{
/* declare variables */
int nr, nc, np, total;
int i, j, k, ix, iy, jx, jy;
int *ir, *jc;
int squareDistance;
/* unsigned long *pi, *pj; */
unsigned int *pi, *pj;
double *w;
double temp,a1,a2,wij;
double rMin;
double sigmaX, sigmaIntensite,valeurMinW;
double *image;
/* check argument */
if (nargin<7) {
mexErrMsgTxt("Four input arguments required");
}
if (nargout>1) {
mexErrMsgTxt("Too many output arguments");
}
/* get edgel information */
nr = mxGetM(in[0]);
nc = mxGetN(in[0]);
np = nr * nc;
image = mxGetPr(in[0]);
/* get new index pair */
if (!mxIsUint32(in[1]) | !mxIsUint32(in[2])) {
mexErrMsgTxt("Index pair shall be of type UINT32");
}
if (mxGetM(in[2]) * mxGetN(in[2]) != np + 1) {
mexErrMsgTxt("Wrong index representation");
}
pi = mxGetData(in[1]);
pj = mxGetData(in[2]);
/* create output */
out[0] = mxCreateSparse(np,np,pj[np],mxREAL);
if (out[0]==NULL) {
mexErrMsgTxt("Not enough memory for the output matrix");
}
w = mxGetPr(out[0]);
ir = mxGetIr(out[0]);
jc = mxGetJc(out[0]);
rMin = mxGetScalar(in[3]);
sigmaX = mxGetScalar(in[4]);
sigmaIntensite= mxGetScalar(in[5]);
valeurMinW = mxGetScalar(in[6]);
a1 = 1.0/ (sigmaX*sigmaX);
a2 = 1.0 / (sigmaIntensite*sigmaIntensite );
/* computation */
total = 0;
for (j=0; j<np; j++) {
jc[j] = total;
jx = j / nr; /* col */
jy = j % nr; /* row */
for (k=pj[j]; k<pj[j+1]; k++) {
i = pi[k];
if (i==j) {
wij= 1; /*voir*/
} else {
ix = i / nr;
iy = i % nr;
squareDistance = (ix-jx)*(ix-jx)+(iy-jy)*(iy-jy);
temp = image[i]-image[j];
wij = exp(- squareDistance * a1 - temp*temp * a2 );
/*if(wij < valeurMinW)
wij = -0.1;*/
}
ir[total] = i;
w[total] = wij;
total = total + 1;
} /* i */
} /* j */
jc[np] = total;
}
void mexFunction
(
int nlhs,
mxArray *plhs [],
int nrhs,
const mxArray *prhs []
)
{
/* Compressed column form of L, x and b */
mwIndex *Lp, *Ap, *Ai, *Up, *pp, *pi ;
Int *Li, *Ui;
mwIndex *Lp1, *Li1, *Up1, *Ui1 ;
double *Lx, *Ax, *Ux, *px ;
double *Lx1, *Ux1 ;
double *t1, *t2 ;
mwIndex anrow ;
mwIndex ancol ;
mwIndex lnnz ;
mwIndex unnz ;
mwIndex i ;
mwIndex j ;
mwIndex app_xnnz ;
mwIndex memsize ;
mwIndex result ;
mwIndex *ws; /* workspace */
double *X ; /* space to scatter x */
mwIndex *pinv ; /* column permutation inverse */
if ( nlhs != 3 || nrhs < 3 )
{
mexErrMsgTxt (" Incorrect number of arguments to sproductmex \n") ;
}
Ap = mxGetJc(prhs[0]) ;
Ai = mxGetIr(prhs[0]) ;
Ax = mxGetPr(prhs[0]) ;
anrow = mxGetM (prhs[0]) ;
ancol = mxGetN (prhs[0]) ;
t1 = mxGetPr (prhs[1]) ;
t2 = mxGetPr (prhs[2]) ;
lnnz = (mwIndex)*t1 ;
unnz = (mwIndex)*t2 ;
/*printf("lnnz=%d unnz=%d\n", lnnz, unnz);*/
/* O(n) initialization */
ws = (mwIndex *) mxCalloc ( (ancol)+(4*anrow), sizeof(mwIndex)) ;
X = (double *) mxCalloc ( 2*anrow, sizeof(double)) ;
pinv = (mwIndex *) mxCalloc ( ancol, sizeof(mwIndex)) ;
Lp = (mwIndex *) mxCalloc ( ancol+1, sizeof(mwIndex)) ;
Li = (mwIndex *) mxCalloc ( lnnz, sizeof(Int)) ;
Lx = (double *) mxCalloc ( lnnz, sizeof(double)) ;
Up = (mwIndex *) mxCalloc ( ancol+1, sizeof(mwIndex)) ;
Ui = (mwIndex *) mxCalloc ( unnz, sizeof(Int)) ;
Ux = (double *) mxCalloc ( unnz, sizeof(double)) ;
PRINT(("Calling basker \n"));
result = basker_basker_l(Ap, Ai, Ax, anrow, ancol, ws , X,
Lp, &Li, &Lx, Up, &Ui, &Ux, &lnnz, &unnz, pinv) ;
PRINT(("Back in mex function %d \n",Lp[ancol]));
if (result)
{
mxFree (X) ;
mxFree (ws) ;
mxFree (Lp) ;
mxFree (Li) ;
mxFree (Lx) ;
mxFree (Up) ;
mxFree (Ui) ;
mxFree (Ux) ;
mexErrMsgTxt (" basker failed \n") ;
}
plhs[0] = mxCreateSparse (anrow, ancol, Lp[ancol], mxREAL) ;
Lp1 = mxGetJc (plhs[0]) ;
Li1 = mxGetIr (plhs[0]) ;
Lx1 = mxGetPr (plhs[0]) ;
plhs[1] = mxCreateSparse (anrow, ancol, Up[ancol], mxREAL) ;
Up1 = mxGetJc (plhs[1]) ;
Ui1 = mxGetIr (plhs[1]) ;
Ux1 = mxGetPr (plhs[1]) ;
plhs[2] = mxCreateSparse (ancol, ancol, ancol, mxREAL) ;
pp = mxGetJc (plhs[2]) ;
pi = mxGetIr (plhs[2]) ;
px = mxGetPr (plhs[2]) ;
Lp1[0] = Lp[0];
for ( i = 0 ; i < ancol ; i++)
{
Lp1[i+1] = Lp[i+1];
for ( j = Lp[i] ; j < Lp[i+1] ; j++ )
{
Li1[j] = Li[j];
Lx1[j] = Lx[j];
}
}
Up1[0] = Up[0];
for ( i = 0 ; i < ancol ; i++)
{
Up1[i+1] = Up[i+1];
for ( j = Up[i] ; j < Up[i+1] ; j++ )
{
Ui1[j] = Ui[j];
Ux1[j] = Ux[j];
}
}
for ( i = 0 ; i < ancol ; i++)
{
pp[i] = i ;
pi[ pinv[i] ] = i ;
px[i] = 1 ;
}
pp[anrow] = ancol ;
/*for ( i = 0 ; i < ancol ; i++)
printf("pinv[%d] = %d\n", i, pinv[i]);*/
for ( i = 0 ; i < ancol ; i++)
{
for ( j = Lp1[i] ; j < Lp1[i+1] ; j++ )
{
Li1[j] = pinv[Li1[j]] ;
}
}
mxFree (X) ;
mxFree (ws) ;
mxFree (pinv) ;
mxFree (Lp) ;
mxFree (Li) ;
mxFree (Lx) ;
mxFree (Up) ;
mxFree (Ui) ;
mxFree (Ux) ;
}
void mexFunction(int nlhs, mxArray *plhs[], int nrhs,
const mxArray *prhs[])
{
double *srww_old, *srww_new,*srwp_old,*srdp_new, *probs, *DS_NEW, *WS_NEW, *WS_OLD, *SI_NEW, *WC_OLD, *PROBA, *C, *Z;
double ALPHA,BETA, GAMMA0, GAMMA1, DELTA;
int *irww_old, *irww_new, *jcww_old, *jcww_new, *irwp_old, *jcwp_old, *irdp_new, *jcdp_new;
int *z,*d,*w, *s, *c, *sc, *order, *wp_old, *dp_new, *dp_new2, *ztot_old, *wtot_new, *wc_new;
int W_NEW, W_OLD,T_NEW, T_OLD,D_NEW,NITER,SEED,OUTPUT, nzmax_new, nzmax_old, nzmaxdp_new, ntokens_new, ntokens_old;
int i,j,cc,nt,n,wi,wipre,di, startindex, endindex, currentrow, count;
int NSAMPLE, updatecols;
/* Check for proper number of arguments. */
if (nrhs != 19) {
mexErrMsgTxt("19 input arguments required");
} else if (nlhs != 4) {
mexErrMsgTxt("4 output arguments required");
}
// Syntax 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
// [ DP_NEW,C,Z,PROB ] = NewDocumentsLDACOL( WS_NEW , DS_NEW , SI_NEW , WW_NEW , T , N , NSAMPLE , ALPHA , BETA , GAMMA0, GAMMA1 , DELTA , SEED , OUTPUT , WW_OLD , WP_OLD , WC_OLD , WS_OLD, UPDATECOLS );
/* process the input arguments */
if (mxIsDouble( prhs[ 0 ] ) != 1)
mexErrMsgTxt("WS_NEW must be double precision");
WS_NEW = mxGetPr(prhs[0]);
ntokens_new = mxGetM( prhs[ 0 ] ) * mxGetN( prhs[ 0 ] );
if (mxIsDouble( prhs[ 1 ] ) != 1)
mexErrMsgTxt("DS_NEW must be double precision");
DS_NEW = mxGetPr(prhs[1]);
if (mxIsDouble( prhs[ 2 ] ) != 1)
mexErrMsgTxt("SI_NEW must be double precision");
SI_NEW = mxGetPr(prhs[2]);
if ((mxIsSparse( prhs[ 3 ] ) != 1) || (mxIsDouble( prhs[ 3 ] ) != 1))
mexErrMsgTxt("WW_NEW collocation matrix must be a sparse double precision matrix");
/* dealing with sparse array WW_NEW */
srww_new = mxGetPr(prhs[3]);
irww_new = mxGetIr(prhs[3]);
jcww_new = mxGetJc(prhs[3]);
nzmax_new = mxGetNzmax(prhs[3]);
W_NEW = mxGetM( prhs[3] );
if (mxGetN( prhs[3] ) != W_NEW) mexErrMsgTxt("WW_NEW matrix should be square");
D_NEW = 0;
for (i=0; i<ntokens_new; i++) {
if (DS_NEW[ i ] > D_NEW) D_NEW = (int) DS_NEW[ i ];
if (WS_NEW[ i ] > W_NEW) mexErrMsgTxt("Some word tokens in WS_NEW stream exceed number of word types in WW_NEW matrix");
}
T_NEW = (int) mxGetScalar(prhs[4]);
if (T_NEW<=0) mexErrMsgTxt("Number of topics must be greater than zero");
NITER = (int) mxGetScalar(prhs[5]);
if (NITER<0) mexErrMsgTxt("Number of iterations must be greater than zero");
NSAMPLE = (int) mxGetScalar(prhs[6]);
if (NSAMPLE<0) mexErrMsgTxt("Number of samples must be greater than zero");
ALPHA = (double) mxGetScalar(prhs[7]);
if (ALPHA<=0) mexErrMsgTxt("ALPHA must be greater than zero");
BETA = (double) mxGetScalar(prhs[8]);
if (BETA<=0) mexErrMsgTxt("BETA must be greater than zero");
GAMMA0 = (double) mxGetScalar(prhs[9]);
if (GAMMA0<=0) mexErrMsgTxt("GAMMA0 must be greater than zero");
GAMMA1 = (double) mxGetScalar(prhs[10]);
if (GAMMA1<=0) mexErrMsgTxt("GAMMA1 must be greater than zero");
DELTA = (double) mxGetScalar(prhs[11]);
if (DELTA<=0) mexErrMsgTxt("DELTA must be greater than zero");
SEED = (int) mxGetScalar(prhs[12]);
// set the seed of the random number generator
OUTPUT = (int) mxGetScalar(prhs[13]);
// dealing with sparse array WW_OLD
if ((mxIsSparse( prhs[ 14 ] ) != 1) || (mxIsDouble( prhs[ 14 ] ) != 1))
mexErrMsgTxt("WW_OLD collocation frequency matrix must be a sparse double precision matrix");
srww_old = mxGetPr(prhs[14]);
irww_old = mxGetIr(prhs[14]);
jcww_old = mxGetJc(prhs[14]);
nzmax_old = mxGetNzmax(prhs[14]);
W_OLD = mxGetM( prhs[14] );
if (W_OLD != W_NEW) mexErrMsgTxt("WW_OLD and WW_NEW matrices have different dimensions");
if (mxGetN( prhs[14] ) != W_NEW) mexErrMsgTxt("WW_OLD matrix should be square");
// dealing with sparse array WP_OLD
if ((mxIsSparse( prhs[ 15 ] ) != 1) || (mxIsDouble( prhs[ 15 ] ) != 1))
mexErrMsgTxt("WP_OLD topic-word matrix must be a sparse double precision matrix");
srwp_old = mxGetPr(prhs[15]);
irwp_old = mxGetIr(prhs[15]);
jcwp_old = mxGetJc(prhs[15]);
if (mxGetM( prhs[15] ) != W_NEW) mexErrMsgTxt("Number of words in WP_OLD matrix does not match WW_NEW number of words");
if (mxGetN( prhs[15] ) != T_NEW) mexErrMsgTxt("Number of topics in WP_OLD matrix does not match given number of topics");
WC_OLD = mxGetPr( prhs[ 16 ]);
if ((mxGetM( prhs[16] ) * mxGetN( prhs[16] )) != W_NEW ) mexErrMsgTxt("Number of words in WC_OLD matrix does not match WW_NEW number of words");
if (mxIsDouble( prhs[ 17 ] ) != 1) mexErrMsgTxt("WS_OLD must be double precision");
WS_OLD = mxGetPr(prhs[ 17 ]);
ntokens_old = mxGetM( prhs[ 17 ] ) * mxGetN( prhs[ 17 ] );
updatecols = (int) mxGetScalar(prhs[18]);
// seeding
seedMT( 1 + SEED * 2 ); // seeding only works on uneven numbers
/* allocate memory */
z = (int *) mxCalloc( ntokens_new , sizeof( int ));
d = (int *) mxCalloc( ntokens_new , sizeof( int ));
w = (int *) mxCalloc( ntokens_new , sizeof( int ));
s = (int *) mxCalloc( ntokens_new , sizeof( int ));
c = (int *) mxCalloc( ntokens_new , sizeof( int ));
sc = (int *) mxCalloc( ntokens_new , sizeof( int ));
for (i=0; i<ntokens_new; i++) w[ i ] = (int) (WS_NEW[ i ] - 1); // Matlab indexing not zero based
for (i=0; i<ntokens_new; i++) d[ i ] = (int) (DS_NEW[ i ] - 1); // Matlab indexing not zero based
for (i=0; i<ntokens_new; i++) s[ i ] = (int) SI_NEW[ i ];
order = (int *) mxCalloc( ntokens_new , sizeof( int ));
wp_old = (int *) mxCalloc( T_NEW*W_NEW , sizeof( int ));
dp_new = (int *) mxCalloc( T_NEW*D_NEW , sizeof( int ));
dp_new2 = (int *) mxCalloc( T_NEW*D_NEW , sizeof( int ));
wc_new = (int *) mxCalloc( W_NEW , sizeof( int ));
ztot_old = (int *) mxCalloc( T_NEW , sizeof( int ));
wtot_new = (int *) mxCalloc( W_NEW , sizeof( int ));
probs = (double *) mxCalloc( T_NEW , sizeof( double ));
// FILL IN WTOT_NEW with OLD COUNTS
for (i=0; i<ntokens_old; i++) {
j = (int) WS_OLD[ i ] - 1;
wtot_new[ j ]++; // calculate number of words of each type
}
// FILL IN WTOT_NEW with NEW COUNTS
if (updatecols==1)
for (i=0; i<ntokens_new; i++) {
j = w[ i ];
wtot_new[ j ]++; // calculate number of words of each type
}
// FILL IN WC WITH COUNTS FROM OLD SET
for (i=0; i<W_NEW; i++) {
wc_new[ i ] = (int) WC_OLD[ i ];
}
// FILL IN wp_old and ztot_old
for (j=0; j<T_NEW; j++) {
startindex = *( jcwp_old + j );
endindex = *( jcwp_old + j + 1 ) - 1;
for (i=startindex; i<=endindex; i++) {
currentrow = *( irwp_old + i );
count = (int) *( srwp_old + i );
wp_old[ j + currentrow*T_NEW ] = count;
ztot_old[ j ] += count;
}
}
// FILL IN SC COUNTS
for (i=1; i<ntokens_new; i++) // start with second item
{
wi = w[ i ]; // word index
count = 0;
// what is the previous word?
wipre = w[ i-1 ];
// calculate how many times the current word follows the previous word IN THE OLD DOCUMENT SET
startindex = *( jcww_old + wipre ); // look up start and end index for column wipre in WW_OLD matrix
endindex = *( jcww_old + wipre + 1 ) - 1;
for (j=startindex; j<=endindex; j++) {
currentrow = *( irww_old + j );
if (currentrow == wi) count = (int) *( srww_old + j );
}
// calculate how many times the current word follows the previous word IN THE NEW DOCUMENT SET
if (updatecols==1) {
startindex = *( jcww_new + wipre ); // look up start and end index for column wipre in WW_NEW matrix
endindex = *( jcww_new + wipre + 1 ) - 1;
for (j=startindex; j<=endindex; j++) {
currentrow = *( irww_new + j );
if (currentrow == wi) count += (int) *( srww_new + j ); // add the count to what was there previously !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
}
}
sc[ i ] = count;
}
/*for (i=0; i<10; i++) {
mexPrintf( "i=%4d w[i]=%3d d[i]=%d z[i]=%d s[i]=%d\n" , i , w[i] , d[i] , z[i] , s[i] );
}*/
if (OUTPUT==2) {
mexPrintf( "Running NEW DOCUMENT LDA COL Gibbs Sampler Version 1.0\n" );
mexPrintf( "Arguments:\n" );
mexPrintf( "\tNumber of words W = %d\n" , W_NEW );
mexPrintf( "\tNumber of docs D = %d\n" , D_NEW );
mexPrintf( "\tNumber of topics T = %d\n" , T_NEW );
mexPrintf( "\tNumber of iterations N = %d\n" , NITER );
mexPrintf( "\tNumber of samples S = %d\n" , NSAMPLE );
mexPrintf( "\tHyperparameter ALPHA = %4.4f\n" , ALPHA );
mexPrintf( "\tHyperparameter BETA = %4.4f\n" , BETA );
mexPrintf( "\tHyperparameter GAMMA0 = %4.4f\n" , GAMMA0 );
mexPrintf( "\tHyperparameter GAMMA1 = %4.4f\n" , GAMMA1 );
mexPrintf( "\tHyperparameter DELTA = %4.4f\n" , DELTA );
mexPrintf( "\tSeed number = %d\n" , SEED );
mexPrintf( "\tNumber of tokens = %d\n" , ntokens_new );
mexPrintf( "\tUpdating collocation counts with new documents? = %d\n" , updatecols );
}
/* ---------------------------------------------
create the PROB vector
-----------------------------------------------*/
plhs[ 3 ] = mxCreateDoubleMatrix( ntokens_new , 1 , mxREAL );
PROBA = mxGetPr( plhs[ 3 ] );
/* run the model */
GibbsSamplerLDACOL( ALPHA, BETA, GAMMA0,GAMMA1,DELTA, W_NEW, T_NEW, D_NEW, NITER, NSAMPLE , OUTPUT, ntokens_new, z, d, w, s, c, sc, wp_old, dp_new, dp_new2, ztot_old, wtot_new, order, probs, wc_new , PROBA , updatecols );
/* ---------------------------------------------
convert the full DP matrix into a sparse matrix
-----------------------------------------------*/
nzmaxdp_new = 0;
for (i=0; i<D_NEW; i++) {
for (j=0; j<T_NEW; j++)
nzmaxdp_new += (int) ( *( dp_new2 + j + i*T_NEW )) > 0; // !!!!!!!!!!!!!!!!!! copy from dp_new2
}
if (OUTPUT==2) {
mexPrintf( "Constructing sparse output matrix dp_new\n" );
mexPrintf( "Number of nonzero entries for DP = %d\n" , nzmaxdp_new );
}
plhs[0] = mxCreateSparse( D_NEW,T_NEW,nzmaxdp_new,mxREAL);
srdp_new = mxGetPr(plhs[0]);
irdp_new = mxGetIr(plhs[0]);
jcdp_new = mxGetJc(plhs[0]);
n = 0;
for (j=0; j<T_NEW; j++) {
*( jcdp_new + j ) = n;
for (i=0; i<D_NEW; i++) {
cc = (int) *( dp_new2 + i*T_NEW + j ); // !!!!!!!!!!!!!!!!!! copy from dp_new2
if (cc >0) {
*( srdp_new + n ) = cc;
*( irdp_new + n ) = i;
n++;
}
}
}
*( jcdp_new + T_NEW ) = n;
/* ---------------------------------------------
create the C route vector
-----------------------------------------------*/
plhs[ 1 ] = mxCreateDoubleMatrix( ntokens_new , 1 , mxREAL );
C = mxGetPr( plhs[ 1 ] );
for (i=0; i<ntokens_new; i++) C[ i ] = (double) c[ i ];
/* ---------------------------------------------
create the topic assignment vector
-----------------------------------------------*/
plhs[ 2 ] = mxCreateDoubleMatrix( ntokens_new , 1 , mxREAL );
Z = mxGetPr( plhs[ 2 ] );
for (i=0; i<ntokens_new; i++) Z[ i ] = (double) z[ i ] + 1;
}
void mexFunction(int nlhs, mxArray *plhs[], int nrhs, const mxArray *prhs[])
{
mwIndex *ir, *jc;
double *pr;
char *filename, *pch, mystring[MAX];
int buflen, d, i, D, W, NNZ;
FILE *doc;
/* Check for proper number of arguments. */
if (nrhs != 1) {
mexErrMsgTxt("Only one input argument required");
} else if (nlhs != 1) {
mexErrMsgTxt("Only one output argument required");
}
// open document file
buflen = (int) (1 + (mxGetM(prhs[0])*mxGetN(prhs[0])));
filename = (char *) mxCalloc(buflen, sizeof(char));
mxGetString( prhs[0], filename, buflen);
doc = fopen(filename, "r");
if (doc == NULL) mexErrMsgTxt("Text file cannot be found.");
// read the file header
fgets(mystring , sizeof(mystring) , doc);
if ((mystring != NULL) && (atoi(mystring) != 0)) {
D = atoi(mystring);
mexPrintf( "#Document: %d \n", D);
} else {
mexErrMsgTxt("File header error.");
}
fgets(mystring , sizeof(mystring) , doc);
if ((mystring != NULL) && (atoi(mystring) != 0)){
W = atoi(mystring);
mexPrintf( "#Vocabulary: %d \n", W);
} else {
mexErrMsgTxt("File header error.");
}
fgets(mystring , sizeof(mystring) , doc);
if ((mystring != NULL) && (atoi(mystring) != 0)){
NNZ = atoi(mystring);
mexPrintf( "#NNZ: %d \n", NNZ);
} else {
mexErrMsgTxt("File header error.");
}
// create W x D sparse matrix for data
plhs[0] = mxCreateSparse(W, D, NNZ, mxREAL);
pr = (double *) mxGetPr(plhs[0]);
ir = mxGetIr(plhs[0]);
jc = mxGetJc(plhs[0]);
// copy data to sparse matrix
i = 0;
jc[0] = 0;
while (fgets(mystring , sizeof(mystring) , doc) != NULL) {
pch = strtok(mystring, " ");
if ((pch != NULL) && (atoi(pch) != 0)) {
d = atoi(pch);
pch = strtok(NULL, " ");
ir[i] = atoi(pch) - 1;
pch = strtok(NULL, " ");
pr[i] = atoi(pch);
i++;
jc[d] = i;
}
}
fclose(doc);
}
/* ************************************************************
PROCEDURE mexFunction - Entry for Matlab
[perm,dz] = incorder(At,Ajc1,ifirst)
************************************************************ */
void mexFunction(int nlhs, mxArray *plhs[],
int nrhs, const mxArray *prhs[])
{
mxArray *myplhs[NPAROUT];
mwIndex i, m, firstPSD, lenud, iwsiz, lenfull, maxnnz;
mwIndex *iwork, *Atjc1, *Ajc, *Air, *perm;
const mwIndex *Atjc2, *Atir;
double *permPr;
const double *Ajc1Pr;
char *cwork;
jcir dz;
/* ------------------------------------------------------------
Check for proper number of arguments
------------------------------------------------------------ */
mxAssert(nrhs >= NPARINMIN, "incorder requires more input arguments.");
mxAssert(nlhs <= NPAROUT, "incorder produces less output arguments.");
/* --------------------------------------------------
GET STATISTICS:
-------------------------------------------------- */
if(nrhs == NPARIN){
firstPSD = (mwIndex) mxGetScalar(IFIRST_IN); /* double to mwIndex */
mxAssert(firstPSD>0,"");
--firstPSD; /* Fortran to C */
}
else
firstPSD = 0; /* default: use all subscripts */
lenfull = mxGetM(AT_IN);
lenud = lenfull - firstPSD;
/* --------------------------------------------------
Check size At, and get At.
-------------------------------------------------- */
mxAssert(mxIsSparse(AT_IN), "At must be a sparse matrix.");
m = mxGetN(AT_IN);
Atjc2 = mxGetJc(AT_IN)+1; /* points to end of constraint */
Atir = mxGetIr(AT_IN); /* subscripts */
/* ------------------------------------------------------------
ALLOCATE WORKING arrays:
mwIndex Atjc1(m), Ajc(1+lenud), Air(maxnnz), perm(m),
iwork(iwsiz). iwsiz = MAX(lenud,1+m)
char cwork(lenud).
------------------------------------------------------------ */
Atjc1 = (mwIndex *) mxCalloc(MAX(m,1), sizeof(mwIndex));
Ajc = (mwIndex *) mxCalloc(1+lenud, sizeof(mwIndex));
perm = (mwIndex *) mxCalloc(MAX(1,m), sizeof(mwIndex));
iwsiz = MAX(lenud, 1 + m);
iwork = (mwIndex *) mxCalloc(iwsiz, sizeof(mwIndex)); /* iwork(iwsiz) */
cwork = (char *) mxCalloc(MAX(1,lenud), sizeof(char));
/* ------------------------------------------------------------
Get input AJc1
------------------------------------------------------------ */
if(nrhs == NPARIN){
Ajc1Pr = mxGetPr(AJC1_IN);
mxAssert(mxGetM(AJC1_IN) * mxGetN(AJC1_IN) >= m, "Ajc1 size mismatch");
/* ------------------------------------------------------------
Double to mwIndex: Atjc1
------------------------------------------------------------ */
for(i = 0; i < m; i++)
Atjc1[i] = (mwIndex) Ajc1Pr[i]; /* double to mwIndex */
/* ------------------------------------------------------------
Let maxnnz = number of PSD nonzeros = sum(Atjc2-Atjc1)
------------------------------------------------------------ */
maxnnz = 0;
for(i = 0; i < m; i++)
maxnnz += Atjc2[i] - Atjc1[i];
}
else{
memcpy(Atjc1, mxGetJc(AT_IN), m * sizeof(mwIndex)); /* default column start */
maxnnz = Atjc2[m-1]; /* maxnnz = nnz(At) */
}
/* ------------------------------------------------------------
ALLOCATE WORKING array mwIndex Air(maxnnz)
------------------------------------------------------------ */
Air = (mwIndex *) mxCalloc(MAX(1,maxnnz), sizeof(mwIndex));
/* ------------------------------------------------------------
CREATE OUTPUT ARRAYS PERM(m) and DZ=sparse(lenfull,m,lenud).
------------------------------------------------------------ */
PERM_OUT = mxCreateDoubleMatrix(m,(mwSize)1,mxREAL);
permPr = mxGetPr(PERM_OUT);
DZ_OUT = mxCreateSparse(lenfull,m,lenud,mxREAL);
dz.jc = mxGetJc(DZ_OUT);
dz.ir = mxGetIr(DZ_OUT);
/* ------------------------------------------------------------
Let (Ajc,Air) := At(first:end,:)', the transpose of PSD-part.
Uses iwork(lenud)
------------------------------------------------------------ */
spPartTransp(Air,Ajc, Atir,Atjc1,Atjc2, firstPSD,lenfull, m,iwork);
/* ------------------------------------------------------------
The main job: greedy order of columns of At
------------------------------------------------------------ */
incorder(perm, dz.jc,dz.ir, Atjc1,Atjc2,Atir, Ajc,Air,
m, firstPSD,lenud, iwork, cwork); /* uses iwork(m+1) */
/* ------------------------------------------------------------
REALLOC (shrink) dz to dz.jc[m] nonzeros.
------------------------------------------------------------ */
mxAssert(dz.jc[m] <= lenud,"");
maxnnz = MAX(1,dz.jc[m]); /* avoid realloc to 0 */
if((dz.ir = (mwIndex *) mxRealloc(dz.ir, maxnnz * sizeof(mwIndex))) == NULL)
mexErrMsgTxt("Memory allocation error");
if((dz.pr = (double *) mxRealloc(mxGetPr(DZ_OUT), maxnnz*sizeof(double)))
== NULL)
mexErrMsgTxt("Memory allocation error");
mxSetPr(DZ_OUT,dz.pr);
mxSetIr(DZ_OUT,dz.ir);
mxSetNzmax(DZ_OUT,maxnnz);
for(i = 0; i < maxnnz; i++)
dz.pr[i] = 1.0;
/* ------------------------------------------------------------
Convert C-mwIndex to Fortran-double
------------------------------------------------------------ */
for(i = 0; i < m; i++)
permPr[i] = perm[i] + 1.0;
/* ------------------------------------------------------------
Release working arrays
------------------------------------------------------------ */
mxFree(cwork);
mxFree(iwork);
mxFree(perm);
mxFree(Air);
mxFree(Ajc);
mxFree(Atjc1);
/* ------------------------------------------------------------
Copy requested output parameters (at least 1), release others.
------------------------------------------------------------ */
i = MAX(nlhs, 1);
memcpy(plhs,myplhs, i * sizeof(mxArray *));
for(; i < NPAROUT; i++)
mxDestroyArray(myplhs[i]);
}
/* ************************************************************
PROCEDURE mexFunction - Entry for Matlab
y = bwblksolve(L,b, [y])
y(L.fullperm) = L.L' \ b
************************************************************ */
void mexFunction(const int nlhs, mxArray *plhs[],
const int nrhs, const mxArray *prhs[])
{
const mxArray *L_FIELD;
mwIndex m,n, j, k, nsuper, inz;
double *y, *fwork;
const double *permPr, *b, *xsuperPr;
const mwIndex *yjc, *yir, *bjc, *bir;
mwIndex *perm, *xsuper, *iwork, *snode;
jcir L;
char bissparse;
/* ------------------------------------------------------------
Check for proper number of arguments
------------------------------------------------------------ */
mxAssert(nrhs >= MINNPARIN, "fwblkslv requires more input arguments.");
mxAssert(nlhs == 1, "fwblkslv generates only 1 output argument.");
/* ------------------------------------------------------------
Disassemble block Cholesky structure L
------------------------------------------------------------ */
mxAssert(mxIsStruct(L_IN), "Parameter `L' should be a structure.");
L_FIELD = mxGetField(L_IN,(mwIndex)0,"perm"); /* L.perm */
mxAssert( L_FIELD != NULL, "Missing field L.perm.");
m = mxGetM(L_FIELD) * mxGetN(L_FIELD);
permPr = mxGetPr(L_FIELD);
L_FIELD = mxGetField(L_IN,(mwIndex)0,"L"); /* L.L */
mxAssert( L_FIELD != NULL, "Missing field L.L.");
mxAssert( m == mxGetM(L_FIELD) && m == mxGetN(L_FIELD), "Size L.L mismatch.");
mxAssert(mxIsSparse(L_FIELD), "L.L should be sparse.");
L.jc = mxGetJc(L_FIELD);
L.ir = mxGetIr(L_FIELD);
L.pr = mxGetPr(L_FIELD);
L_FIELD = mxGetField(L_IN,(mwIndex)0,"xsuper"); /* L.xsuper */
mxAssert( L_FIELD != NULL, "Missing field L.xsuper.");
nsuper = mxGetM(L_FIELD) * mxGetN(L_FIELD) - 1;
mxAssert( nsuper <= m, "Size L.xsuper mismatch.");
xsuperPr = mxGetPr(L_FIELD);
/* ------------------------------------------------------------
Get rhs matrix b.
If it is sparse, then we also need the sparsity structure of y.
------------------------------------------------------------ */
b = mxGetPr(B_IN);
mxAssert( mxGetM(B_IN) == m, "Size mismatch b.");
n = mxGetN(B_IN);
if( (bissparse = mxIsSparse(B_IN)) ){
bjc = mxGetJc(B_IN);
bir = mxGetIr(B_IN);
mxAssert(nrhs >= NPARIN, "bwblkslv requires more inputs in case of sparse b.");
mxAssert(mxGetM(Y_IN) == m && mxGetN(Y_IN) == n, "Size mismatch y.");
mxAssert(mxIsSparse(Y_IN), "y should be sparse.");
}
/* ------------------------------------------------------------
Allocate output y. If bissparse, then Y_IN gives the sparsity structure.
------------------------------------------------------------ */
if(!bissparse)
Y_OUT = mxCreateDoubleMatrix(m, n, mxREAL);
else{
yjc = mxGetJc(Y_IN);
yir = mxGetIr(Y_IN);
Y_OUT = mxCreateSparse(m,n, yjc[n],mxREAL);
memcpy(mxGetJc(Y_OUT), yjc, (n+1) * sizeof(mwIndex));
memcpy(mxGetIr(Y_OUT), yir, yjc[n] * sizeof(mwIndex));
}
y = mxGetPr(Y_OUT);
/* ------------------------------------------------------------
Allocate working arrays
------------------------------------------------------------ */
fwork = (double *) mxCalloc(m, sizeof(double));
iwork = (mwIndex *) mxCalloc(2*m+nsuper+1, sizeof(mwIndex));
perm = iwork; /* m */
xsuper = iwork + m; /*nsuper+1*/
snode = xsuper + (nsuper+1); /* m */
/* ------------------------------------------------------------
Convert real to integer array, and from Fortran to C style.
------------------------------------------------------------ */
for(k = 0; k < m; k++)
perm[k] = permPr[k] - 1;
for(k = 0; k <= nsuper; k++)
xsuper[k] = xsuperPr[k] - 1;
/* ------------------------------------------------------------
In case of sparse b, we also create snode, which maps each subnode
to the supernode containing it.
------------------------------------------------------------ */
if(bissparse)
for(j = 0, k = 0; k < nsuper; k++)
while(j < xsuper[k+1])
snode[j++] = k;
/* ------------------------------------------------------------
The actual job is done here: y(perm) = L'\b.
------------------------------------------------------------ */
if(!bissparse)
for(j = 0; j < n; j++){
memcpy(fwork,b, m * sizeof(double));
bwsolve(fwork,L.jc,L.ir,L.pr,xsuper,nsuper,y); /* y(m) as work */
for(k = 0; k < m; k++) /* y(perm) = fwork */
y[perm[k]] = fwork[k];
y += m; b += m;
}
else{ /* sparse y,b: don't use perm */
fzeros(fwork,m);
for(j = 0; j < n; j++){
inz = yjc[j];
for(k = bjc[j]; k < bjc[j+1]; k++) /* fwork = b */
fwork[bir[k]] = b[k];
selbwsolve(fwork,L.jc,L.ir,L.pr,xsuper,nsuper, snode,
yir+inz,yjc[j+1]-inz);
for(k = inz; k < yjc[j+1]; k++)
y[k] = fwork[yir[k]];
for(k = inz; k < yjc[j+1]; k++) /* fwork = all-0 */
fwork[yir[k]] = 0.0;
}
}
/* ------------------------------------------------------------
RELEASE WORKING ARRAYS.
------------------------------------------------------------ */
mxFree(fwork);
mxFree(iwork);
}
void mexFunction( int nlhs, mxArray *plhs[],
int nrhs, const mxArray *prhs[] )
{
mxArray *rhs[2];
mxArray *blk_cell_pr, *A_cell_pr;
double *A, *B, *blksize;
mwIndex *irA, *jcA, *irB, *jcB;
mwIndex *cumblksize, *blknnz;
mwSize iscellA, mblk, mA, nA, m1, n1, rowidx, colidx, isspA, isspB;
mwIndex subs[2];
mwSize nsubs=2;
mwSize n, n2, k, nsub, index, numblk, NZmax;
double ir2=1/sqrt(2);
/* CHECK FOR PROPER NUMBER OF ARGUMENTS */
if (nrhs < 2){
mexErrMsgTxt("mexsmat: requires at least 2 input arguments."); }
else if (nlhs>1){
mexErrMsgTxt("mexsmat: requires 1 output argument."); }
/* CHECK THE DIMENSIONS */
iscellA = mxIsCell(prhs[1]);
if (iscellA) { mA = mxGetM(prhs[1]); nA = mxGetN(prhs[1]); }
else { mA = 1; nA = 1; }
if (mxGetM(prhs[0]) != mA) {
mexErrMsgTxt("mexsmat: blk and Avec not compatible"); }
/***** main body *****/
if (nrhs > 3) {rowidx = (mwSize)*mxGetPr(prhs[3]); } else {rowidx = 1;}
if (rowidx > mA) {
mexErrMsgTxt("mexsmat: rowidx exceeds size(Avec,1)."); }
subs[0] = rowidx-1; /* subtract 1 to adjust for Matlab index */
subs[1] = 1;
index = mxCalcSingleSubscript(prhs[0],nsubs,subs);
blk_cell_pr = mxGetCell(prhs[0],index);
if (blk_cell_pr == NULL) {
mexErrMsgTxt("mexsmat: blk not properly specified"); }
numblk = mxGetN(blk_cell_pr);
blksize = mxGetPr(blk_cell_pr);
cumblksize = mxCalloc(numblk+1,sizeof(mwSize));
blknnz = mxCalloc(numblk+1,sizeof(mwSize));
cumblksize[0] = 0; blknnz[0] = 0;
n = 0; n2 = 0;
for (k=0; k<numblk; ++k) {
nsub = (mwSize) blksize[k];
n += nsub;
n2 += nsub*(nsub+1)/2;
cumblksize[k+1] = n;
blknnz[k+1] = n2;
}
/***** assign pomwSizeers *****/
if (iscellA) {
subs[0] = rowidx-1;
subs[1] = 0;
index = mxCalcSingleSubscript(prhs[1],nsubs,subs);
A_cell_pr = mxGetCell(prhs[1],index);
A = mxGetPr(A_cell_pr);
m1 = mxGetM(A_cell_pr);
n1 = mxGetN(A_cell_pr);
isspA = mxIsSparse(A_cell_pr);
if (isspA) { irA = mxGetIr(A_cell_pr);
jcA = mxGetJc(A_cell_pr); }
} else {
A = mxGetPr(prhs[1]);
m1 = mxGetM(prhs[1]);
n1 = mxGetN(prhs[1]);
isspA = mxIsSparse(prhs[1]);
if (isspA) { irA = mxGetIr(prhs[1]);
jcA = mxGetJc(prhs[1]); }
}
if (numblk > 1)
{ isspB = 1; }
else {
if (nrhs > 2) {isspB = (mwSize)*mxGetPr(prhs[2]);} else {isspB = isspA;}
}
if (nrhs > 4) {colidx = (mwSize)*mxGetPr(prhs[4]) -1;} else {colidx = 0;}
if (colidx > n1) {
mexErrMsgTxt("mexsmat: colidx exceeds size(Avec,2).");
}
/***** create return argument *****/
if (isspB) {
if (isspA) { NZmax = jcA[colidx+1]-jcA[colidx]; }
else { NZmax = blknnz[numblk]; }
rhs[0] = mxCreateSparse(n,n,2*NZmax,mxREAL);
B = mxGetPr(rhs[0]); irB = mxGetIr(rhs[0]); jcB = mxGetJc(rhs[0]);
} else {
plhs[0] = mxCreateDoubleMatrix(n,n,mxREAL);
B = mxGetPr(plhs[0]);
}
/***** Do the computations in a subroutine *****/
if (numblk == 1) {
smat1(n,ir2,A,irA,jcA,isspA,m1,colidx,B,irB,jcB,isspB); }
else {
smat2(n,numblk,cumblksize,blknnz,ir2,A,irA,jcA,isspA,m1,colidx,B,irB,jcB,isspB);
}
if (isspB) {
/*** if isspB, (actual B) = B+B' ****/
mexCallMATLAB(1, &rhs[1], 1, &rhs[0], "transpose");
mexCallMATLAB(1, &plhs[0],2, rhs, "+");
mxDestroyArray(*rhs);
}
return;
}
m = mxGetM(prhs[0]);
n = mxGetN(prhs[0]);
pr = mxGetPr(prhs[0]);
pi = mxGetPi(prhs[0]);
cmplx = (pi == NULL ? 0 : 1);
/* Allocate space for sparse matrix.
* NOTE: Assume at most 20% of the data is sparse. Use ceil
* to cause it to round up.
*/
percent_sparse = 0.2;
nzmax = (int)ceil((double)m*(double)n*percent_sparse);
plhs[0] = mxCreateSparse(m,n,nzmax,cmplx);
sr = mxGetPr(plhs[0]);
si = mxGetPi(plhs[0]);
irs = mxGetIr(plhs[0]);
jcs = mxGetJc(plhs[0]);
/* Copy nonzeros. */
k = 0;
isfull = 0;
for (j = 0; (j < n); j++) {
int i;
jcs[j] = k;
for (i = 0; (i < m); i++) {
if (IsNonZero(pr[i]) || (cmplx && IsNonZero(pi[i]))) {
/* Check to see if non-zero element will fit in
void mexFunction(int nlhs, mxArray *plhs[], int nrhs,
const mxArray *prhs[])
{
double *srwp, *srdp, *srmp, *probs, *d, *w, *ZIN, *XIN, *Z, *X;
double ALPHA,BETA, GAMMA;
mwIndex *irwp, *jcwp, *irdp, *jcdp, *irmp, *jcmp;
int *z,*x, *wp, *mp, *dp, *ztot, *mtot, *stot, *sp, *first, *second, *third;
int W,T,S,SINPUT,D,NN,SEED,OUTPUT, nzmax, nzmaxwp,nzmaxmp, nzmaxdp, ntokens;
int NWS,NDS,i,j,c,n,wi,di,i1,i2,i3,i4,S2,S3, startcond;
// Syntax
// [ WP , DP , MP , Z , X ] = GibbsSamplerHMMLDA( WS , DS , T , S , N , ALPHA , BETA , GAMMA , SEED , OUTPUT , ZIN , XIN )
/* Check for proper number of arguments. */
if (nrhs < 10) {
mexErrMsgTxt("At least 10 input arguments required");
} else if (nlhs != 5) {
mexErrMsgTxt("5 output arguments required");
}
startcond = 0;
if (nrhs > 10) startcond = 1;
if (sizeof( int ) != sizeof( mxINT32_CLASS ))
mexErrMsgTxt("Problem with internal integer representation -- contact programmer" );
w = mxGetPr( prhs[0] );
NWS = mxGetN( prhs[0] );
d = mxGetPr( prhs[1] );
NDS = mxGetN( prhs[1] );
if ((mxIsSparse( prhs[ 0 ] ) == 1) || (mxIsDouble( prhs[ 0 ] ) != 1) || (mxGetM( prhs[0]) != 1))
mexErrMsgTxt("WS input stream must be a one-dimensional, non-sparse, double precision vector of word indices");
if ((mxIsSparse( prhs[ 1 ] ) == 1) || (mxIsDouble( prhs[ 1 ] ) != 1) || (mxGetM( prhs[1]) != 1))
mexErrMsgTxt("DS input stream must be a one-dimensional, non-sparse, double precision vector of document indices");
if (NWS != NDS)
mexErrMsgTxt("WS and DS input streams must have equal dimensions" );
ntokens = NWS;
T = (int) mxGetScalar(prhs[2]);
if (T<=0) mexErrMsgTxt("Number of topics must be greater than zero");
SINPUT = (int) mxGetScalar(prhs[3]);
if (SINPUT<=0) mexErrMsgTxt("Number of syntactic states must be greater than zero");
// need one syntactic state for sentence start
// need one syntactic state for topic state
S = SINPUT + 2;
NN = (int) mxGetScalar(prhs[4]);
if (NN<0) mexErrMsgTxt("Number of iterations must be greater than zero");
ALPHA = (double) mxGetScalar(prhs[5]);
if (ALPHA<=0) mexErrMsgTxt("ALPHA must be greater than zero");
BETA = (double) mxGetScalar(prhs[6]);
if (BETA<=0) mexErrMsgTxt("BETA must be greater than zero");
GAMMA = (double) mxGetScalar(prhs[7]);
if (GAMMA<=0) mexErrMsgTxt("GAMMA must be greater than zero");
SEED = (int) mxGetScalar(prhs[8]);
// set the seed of the random number generator
OUTPUT = (int) mxGetScalar(prhs[9]);
if (startcond==1) {
if ((mxGetN( prhs[ 10 ] )*mxGetM( prhs[ 10 ] )) != ntokens) mexErrMsgTxt( "The ZIN vector should have have the same number of tokens as WS" );
if ((mxGetN( prhs[ 11 ] )*mxGetM( prhs[ 11 ] )) != ntokens) mexErrMsgTxt( "The ZIN vector should have have the same number of tokens as WS" );
ZIN = mxGetPr( prhs[ 10 ]);
XIN = mxGetPr( prhs[ 11 ]);
}
// seeding
seedMT( 1 + SEED * 2 ); // seeding only works on uneven numbers
W = 0;
D = 0;
for (i=0; i<ntokens; i++) {
// in Matlab WI=1...W word occurence, WI=0 sentence marker
wi = (int) w[ i ] - 1; // word indices are not zero based in Matlab
di = (int) d[ i ] - 1; // document indices are not zero based in Matlab
//mexPrintf( "i=%d wi=%d di=%d\n" , i , wi , di );
if (wi > W) W = wi;
if (di > D) D = di;
if (wi < -1) mexErrMsgTxt("Unrecognized code in word stream (<-2)");
if (di < 0) mexErrMsgTxt("Unrecognized code in document stream (<0)");
}
W = W + 1;
D = D + 1;
if ( wi != -1 )
mexErrMsgTxt("Word stream should end with sentence marker");
n = ntokens;
//mexPrintf( "W=%d D=%d n=%d w[n-1]=%d\n" , W , D , n , (int) w[ n-1 ] );
// allocate memory
z = (int *) mxCalloc( ntokens , sizeof( int ));
x = (int *) mxCalloc( ntokens , sizeof( int ));
if (startcond==1) {
for (i=0; i<ntokens; i++) {
z[ i ] = (int) ZIN[ i ] - 1;
x[ i ] = (int) XIN[ i ] - 1;
}
}
first = (int *) mxCalloc( ntokens , sizeof( int ));
second = (int *) mxCalloc( ntokens , sizeof( int ));
third = (int *) mxCalloc( ntokens , sizeof( int ));
mp = (int *) mxCalloc( W*S , sizeof( int ));
wp = (int *) mxCalloc( T*W , sizeof( int ));
dp = (int *) mxCalloc( T*D , sizeof( int ));
ztot = (int *) mxCalloc( T , sizeof( int ));
mtot = (int *) mxCalloc( S , sizeof( int ));
stot = (int *) mxCalloc( S * S * S , sizeof( int ));
sp = (int *) mxCalloc( S * S * S * S , sizeof( int ));
probs = (double *) mxCalloc( T+S , sizeof( double ));
//for (i=0; i<n; i++) mexPrintf( "i=%4d w[i]=%3d d[i]=%d\n" , i , w[i] , d[i] , z[i] );
if (OUTPUT==2) {
mexPrintf( "Running HMM-LDA Gibbs Sampler Version 1.0\n" );
if (startcond==1) mexPrintf( "Starting sampler from previous state\n" );
mexPrintf( "Arguments:\n" );
mexPrintf( "\tNumber of words W = %d\n" , W );
mexPrintf( "\tNumber of docs D = %d\n" , D );
mexPrintf( "\tNumber of topics T = %d\n" , T );
mexPrintf( "\tNumber of syntactic states S = %d\n" , S );
mexPrintf( "\tNumber of iterations N = %d\n" , NN );
mexPrintf( "\tHyperparameter ALPHA = %4.4f\n" , ALPHA );
mexPrintf( "\tHyperparameter BETA = %4.4f\n" , BETA );
mexPrintf( "\tHyperparameter GAMMA = %4.4f\n" , GAMMA );
mexPrintf( "\tSeed number SEED = %d\n" , SEED );
mexPrintf( "Properties of WS stream\n" );
mexPrintf( "\tNumber of tokens NZ = %d\n" , ntokens );
mexPrintf( "Internal Memory Allocation\n" );
mexPrintf( "\tz,x,first,second,third indices combined = %d bytes\n" , 5 * sizeof( int) * ntokens );
mexPrintf( "\twp matrix = %d bytes\n" , sizeof( int ) * W * T );
mexPrintf( "\tmp matrix = %d bytes\n" , sizeof( int ) * W * S );
mexPrintf( "\tdp matrix = %d bytes\n" , sizeof( int ) * D * T );
mexPrintf( "\tstot matrix = %d bytes\n" , sizeof( int ) * S * S * S );
mexPrintf( "\tsp matrix = %d bytes\n" , sizeof( int ) * S * S * S * S );
//mexPrintf( "Checking: sizeof(int)=%d sizeof(long)=%d sizeof(double)=%d\n" , sizeof(int) , sizeof(long) , sizeof(double));
}
// run the model
GibbsSamplerHMMLDA( ALPHA, BETA, GAMMA, W, T, S , D, NN, OUTPUT, n, z, x, d, w, wp, mp, dp, ztot, mtot, stot , sp , first,second,third, probs, startcond );
// convert the full wp matrix into a sparse matrix
nzmaxwp = 0;
for (i=0; i<W; i++) {
for (j=0; j<T; j++)
nzmaxwp += (int) ( *( wp + j + i*T )) > 0;
}
if (OUTPUT==2) mexPrintf( "Constructing sparse output matrix WP nnz=%d\n" , nzmaxwp );
plhs[0] = mxCreateSparse( W,T,nzmaxwp,mxREAL);
srwp = mxGetPr(plhs[0]);
irwp = mxGetIr(plhs[0]);
jcwp = mxGetJc(plhs[0]);
n = 0;
for (j=0; j<T; j++) {
*( jcwp + j ) = n;
for (i=0; i<W; i++) {
c = (int) *( wp + i*T + j );
if (c >0) {
*( srwp + n ) = c;
*( irwp + n ) = i;
n++;
}
}
}
*( jcwp + T ) = n;
// processing DP as sparse output matrix
nzmaxdp = 0;
for (i=0; i<D; i++) {
for (j=0; j<T; j++)
nzmaxdp += (int) ( *( dp + j + i*T )) > 0;
}
if (OUTPUT==2) mexPrintf( "Constructing sparse output matrix DP nnz=%d\n" , nzmaxdp );
plhs[1] = mxCreateSparse( D,T,nzmaxdp,mxREAL);
srdp = mxGetPr(plhs[1]);
irdp = mxGetIr(plhs[1]);
jcdp = mxGetJc(plhs[1]);
n = 0;
for (j=0; j<T; j++) {
*( jcdp + j ) = n;
for (i=0; i<D; i++) {
c = (int) *( dp + i*T + j );
if (c >0) {
*( srdp + n ) = c;
*( irdp + n ) = i;
n++;
}
}
}
*( jcdp + T ) = n;
// processing MP as sparse output matrix
nzmaxmp = 0;
for (i=0; i<W; i++) {
for (j=0; j<S; j++) {
nzmaxmp += (int) ( mp[ j + i*S ] > 0 );
}
}
if (OUTPUT==2) mexPrintf( "Constructing sparse output matrix MP nnz=%d\n" , nzmaxmp );
plhs[2] = mxCreateSparse( W,SINPUT,nzmaxmp,mxREAL);
srmp = mxGetPr(plhs[2]);
irmp = mxGetIr(plhs[2]);
jcmp = mxGetJc(plhs[2]);
n = 0;
for (j=0; j<SINPUT; j++) {
*( jcmp + j ) = n;
for (i=0; i<W; i++) {
c = mp[ i*S + (j+2) ];
if (c >0) {
*( srmp + n ) = c;
*( irmp + n ) = i;
n++;
}
}
}
*( jcmp + SINPUT ) = n;
/* ---------------------------------------------
create the topic assignment vector
-----------------------------------------------*/
plhs[ 3 ] = mxCreateDoubleMatrix( ntokens , 1 , mxREAL );
Z = mxGetPr( plhs[ 3 ] );
for (i=0; i<ntokens; i++) Z[ i ] = (double) z[ i ] + 1;
/* ---------------------------------------------
create the HMM state assignment vector
-----------------------------------------------*/
plhs[ 4 ] = mxCreateDoubleMatrix( ntokens , 1 , mxREAL );
X = mxGetPr( plhs[ 4 ] );
for (i=0; i<ntokens; i++) X[ i ] = (double) x[ i ] + 1;
}
void mexFunction(
int nlhs, mxArray *plhs[],
int nrhs, const mxArray *prhs[] )
{ double *A, *U, *V, *VT, *S, *flag, *work, *AA;
mwIndex subs[2];
mwSize nsubs=2;
mwIndex *irS, *jcS, *iwork;
mwSize M, N, lwork, info, options, minMN, maxMN, k, j;
mwSize LDA, LDU, LDVT, nU, mVT, mS, nS;
char *jobz;
/* CHECK FOR PROPER NUMBER OF ARGUMENTS */
if (nrhs > 2){
mexErrMsgTxt("mexsvd: requires at most 2 input arguments."); }
if (nlhs > 4){
mexErrMsgTxt("mexsvd: requires at most 4 output argument."); }
/* CHECK THE DIMENSIONS */
M = mxGetM(prhs[0]);
N = mxGetN(prhs[0]);
if (mxIsSparse(prhs[0])) {
mexErrMsgTxt("mexeig: sparse matrix not allowed."); }
A = mxGetPr(prhs[0]);
LDA = M;
minMN = MIN(M,N);
maxMN = MAX(M,N);
options = 0;
if (nrhs==2) { options = (int)*mxGetPr(prhs[1]); }
if (options==0) {
/***** economical SVD ******/
jobz="S"; nU = minMN; mVT = minMN; mS = minMN; nS = minMN;
} else {
/***** full SVD ******/
jobz="A"; nU = M; mVT = N; mS = M; nS = N;
}
LDU = M;
LDVT = mVT;
/***** create return argument *****/
if (nlhs >=3) {
/***** compute singular values and vectors ******/
plhs[0] = mxCreateDoubleMatrix(M,nU,mxREAL);
U = mxGetPr(plhs[0]);
plhs[1] = mxCreateSparse(mS,nS,minMN,mxREAL);
S = mxGetPr(plhs[1]);
irS = mxGetIr(plhs[1]);
jcS = mxGetJc(plhs[1]);
plhs[2] = mxCreateDoubleMatrix(N,mVT,mxREAL);
V = mxGetPr(plhs[2]);
plhs[3] = mxCreateDoubleMatrix(1,1,mxREAL);
flag = mxGetPr(plhs[3]);
} else {
/***** compute only singular values ******/
plhs[0]= mxCreateDoubleMatrix(minMN,1,mxREAL);
S = mxGetPr(plhs[0]);
plhs[1] = mxCreateDoubleMatrix(1,1,mxREAL);
flag = mxGetPr(plhs[1]);
U = mxCalloc(M*nU,sizeof(double));
V = mxCalloc(N*mVT,sizeof(double));
jobz="N";
}
/***** Do the computations in a subroutine *****/
lwork = 4*minMN*minMN + MAX(maxMN,5*minMN*minMN+4*minMN);
work = mxCalloc(lwork,sizeof(double));
iwork = mxCalloc(8*minMN,sizeof(int));
VT = mxCalloc(mVT*N,sizeof(double));
AA = mxCalloc(M*N,sizeof(double));
memcpy(AA,mxGetPr(prhs[0]),(M*N)*sizeof(double));
dgesdd(jobz,&M,&N, AA,&LDA,S,U,&LDU,VT,&LDVT,work,&lwork,iwork, &info);
flag[0] = (double)info;
if (nlhs >= 3) {
for (k=0; k<minMN; k++) { irS[k] = k; }
jcS[0] = 0;
for (k=1; k<=nS; k++) {
if (k<minMN) { jcS[k] = k; } else { jcS[k] = minMN; }
}
for (k=0; k<mVT; k++) {
for (j=0; j<N; j++) { V[j+k*N] = VT[k+j*mVT]; }
}
}
return;
}
mxArray* omp_ls(const double m_dict[],
const double m_x[],
mwSize M,
mwSize N,
mwSize S,
mwSize K,
double res_norm_bnd,
int sparse_output,
int verbose){
// List of indices of selected atoms
mwIndex *selected_atoms = 0;
// Simple binary mask of selected atoms
int* selected_atoms_mask = 0;
// The submatrix of selected atoms
double* m_subdict = 0;
// Copy of subdictionary for least squares
double* m_subdict_copy = 0;
// The proxy D' x
double* v_proxy = 0;
// The inner product of residual with atoms
double* v_h = 0;
// The residual
double* v_r = 0;
// Result of orthogonal projection LL' c = p_I
double* v_c = 0;
// Some temporary vectors
double *v_t1 = 0, *v_t2 = 0;
// Pointer to new atom
const double* wv_new_atom;
// residual norm squared
double res_norm_sqr;
// square of upper bound on residual norm
double res_norm_bnd_sqr = SQR(res_norm_bnd);
// Pointer to current signal
const double *wv_x = 0;
/// Output array
mxArray* p_alpha;
double* m_alpha;
// row indices for non-zero entries in Alpha
mwIndex *ir_alpha;
// indices for first non-zero entry in column
mwIndex *jc_alpha;
/// Index for non-zero entries in alpha
mwIndex nz_index;
// counters
int i, j , k, s;
// index of new atom
mwIndex new_atom_index;
// misc variables
double d1, d2;
// Maximum number of columns to be used in representations
mwSize max_cols;
// structure for tracking time spent.
omp_profile profile;
if (K < 0 || K > M) {
// K cannot be greater than M.
K = M;
}
max_cols = (mwSize)(ceil(sqrt((double)M)/2.0) + 1.01);
if(max_cols < K){
max_cols = K;
}
// Memory allocations
// Number of selected atoms cannot exceed M
selected_atoms = (mwIndex*) mxMalloc(M*sizeof(mwIndex));
// Total number of atoms is N
selected_atoms_mask = (int*) mxMalloc(N*sizeof(int));
// Coefficients of solution of least square problem
v_c = (double*)mxMalloc(M*sizeof(double));
// Giving enough space for temporary vectors
v_t1 = (double*)mxMalloc(N*sizeof(double));
v_t2 = (double*)mxMalloc(N*sizeof(double));
// Keeping max_cols space for subdictionary.
m_subdict = (double*)mxMalloc(max_cols*M*sizeof(double));
m_subdict_copy = (double*)mxMalloc(max_cols*M*sizeof(double));
// Proxy vector is in R^N
v_proxy = (double*)mxMalloc(N*sizeof(double));
// h is in R^N.
v_h = (double*)mxMalloc(N*sizeof(double));
// Residual is in signal space R^M.
v_r = (double*)mxMalloc(M*sizeof(double));
if (sparse_output == 0){
p_alpha = mxCreateDoubleMatrix(N, S, mxREAL);
m_alpha = mxGetPr(p_alpha);
ir_alpha = 0;
jc_alpha = 0;
}else{
p_alpha = mxCreateSparse(N, S, max_cols*S, mxREAL);
m_alpha = mxGetPr(p_alpha);
ir_alpha = mxGetIr(p_alpha);
jc_alpha = mxGetJc(p_alpha);
nz_index = 0;
jc_alpha[0] = 0;
}
omp_profile_init(&profile);
for(s=0; s<S; ++s){
wv_x = m_x + M*s;
// Initialization
res_norm_sqr = inner_product(wv_x, wv_x, M);
//Compute proxy p = D' * x
mult_mat_t_vec(1, m_dict, wv_x, v_proxy, M, N);
omp_profile_toctic(&profile, TIME_DtR);
// h = p = D' * r
copy_vec_vec(v_proxy, v_h, N);
for (i=0; i<N; ++i){
selected_atoms_mask[i] = 0;
}
// Number of atoms selected so far.
k = 0;
// Iterate for each atom
while (k < K && res_norm_sqr > res_norm_bnd_sqr){
omp_profile_tic(&profile);
// Pick the index of (k+1)-th atom
new_atom_index = abs_max_index(v_h, N);
omp_profile_toctic(&profile, TIME_MaxAbs);
// If this atom is already selected, we will break
if (selected_atoms_mask[new_atom_index]){
// This is unlikely due to orthogonal structure of OMP
if (verbose){
mexPrintf("This atom is already selected.");
}
break;
}
// Check for small values
d2 = v_h[new_atom_index];
if (SQR(d2) < 1e-14){
// The inner product of residual with new atom is way too small.
break;
}
// Store the index of new atom
selected_atoms[k] = new_atom_index;
selected_atoms_mask[new_atom_index] = 1;
// Copy the new atom to the sub-dictionary
wv_new_atom = m_dict + new_atom_index*M;
copy_vec_vec(wv_new_atom, m_subdict+k*M, M);
omp_profile_toctic(&profile, TIME_DictSubMatrixUpdate);
// It is time to increase the count of selected atoms
++k;
// Least squares
copy_vec_vec(m_subdict, m_subdict_copy, M*k);
copy_vec_vec(wv_x, v_t1, M);
least_square(m_subdict_copy, v_t1, v_c, M, k, 1);
omp_profile_toctic(&profile, TIME_LeastSquares);
// Compute residual
// r = x - D_I c
mult_mat_vec(-1, m_subdict, v_c, v_r, M, k);
sum_vec_vec(1, wv_x, v_r, M);
omp_profile_toctic(&profile, TIME_RUpdate);
// Update h = D' r
mult_mat_t_vec(1, m_dict, v_r, v_h, M, N);
// Update residual norm squared
res_norm_sqr = inner_product(v_r, v_r, M);
omp_profile_toctic(&profile, TIME_DtR);
//mexPrintf(".\n");
}
// Write the output vector
if(sparse_output == 0){
// Write the output vector
double* wv_alpha = m_alpha + N*s;
fill_vec_sparse_vals(v_c, selected_atoms, wv_alpha, N, k);
}
else{
// Sort the row indices
quicksort_indices(selected_atoms, v_c, k);
// add the non-zero entries for this column
for(j=0; j <k; ++j){
m_alpha[nz_index] = v_c[j];
ir_alpha[nz_index] = selected_atoms[j];
++nz_index;
}
// fill in the total number of nonzero entries in the end.
jc_alpha[s+1] = jc_alpha[s] + k;
}
}
if(verbose){
omp_profile_print(&profile);
}
// Memory cleanup
mxFree(selected_atoms);
mxFree(selected_atoms_mask);
mxFree(v_c);
mxFree(v_t1);
mxFree(v_t2);
mxFree(m_subdict);
mxFree(m_subdict_copy);
mxFree(v_proxy);
mxFree(v_h);
mxFree(v_r);
// Return the result
return p_alpha;
}
void mexFunction(int nlhs, mxArray *plhs[],
int nrhs, const mxArray *prhs[])
{
/*input for C function */
double *in; /*input image*/
double *weight; /*output weight*/
double h;
int nx,ny,nxny,newnx,newny;
int nwin, bloc;
double a, percent_sparse,v,c,sum,e,dist;
int i,j,k,nzmax;
int *dadr,*dd,x,y,xp,yp,adr,adrp,psize,wsize,ds,realadr,realadrp,li;
double *w,*ww,*lw;
/*symetry image*/
double* in_sym;
mwIndex *irs,*jcs;
char s;
/*Output for C function */
int l;
if (nrhs < 1 )
mexErrMsgTxt("At least one inputs required.");
if (nrhs > 6 )
mexErrMsgTxt("Too many inputs required.");
if (nlhs != 1)
mexErrMsgTxt("One output required.");
/*Input Image */
nx = mxGetM(prhs[0]);
ny = mxGetN(prhs[0]);
in = mxGetPr(prhs[0]);
h = mxGetScalar(prhs[1]);
nxny=nx*ny;
if (nrhs<3)
nwin=3;
else
nwin=mxGetScalar(prhs[2]);
if (nrhs<4)
bloc=6;
else
bloc=mxGetScalar(prhs[3]);
/*Other parameters*/
percent_sparse=(2*bloc+1)*(2*bloc+1)/(double)nxny;
nzmax=(mwSize)ceil((double)nxny*(double)nxny*percent_sparse);
plhs[0] = mxCreateSparse(nxny,nxny,nzmax,0);
weight= mxGetPr(plhs[0]);
/* Create a C pointer to a copy of the output matrix,but be careful that the matrix is tranposed */
a=(2*nwin)/4.;
c=1;
/*Sparse weight paramters*/
irs= mxGetIr(plhs[0]);
jcs= mxGetJc(plhs[0]);
/* precompute weights */
ds =nwin;
psize = (2*nwin+1)*(2*nwin+1); /* patch size; patch = [-ds,ds]x[-ds,ds] */
wsize = (2*bloc+1)*(2*bloc+1); /* window size; patch = [-d,d]x[-d,d] */
w = (double *)malloc(psize*sizeof(double)); /*patch weight */
dadr = (int *)malloc(psize*sizeof(int));/*patch index */
lw = (double *)malloc(wsize*sizeof(double)); /*local window weight */
/*symstry bord of image*/
newnx=nx+2*ds;
newny=ny+2*ds;
in_sym = (double *)malloc(newny*newnx*sizeof(double));
/* extend the original image */
for (j=0,k=0;j<newny;j++)
{
if (j<ds)
y=ds-j;
else if (j>ny+ds-1)
y=2*ny+ds-j-2;
else y=j-ds;
for (i=0;i<newnx;i++,k++)
{
if (i<=ds)
x=ds-i;
else if (i>nx+ds-1) x=2*nx+ds-i-2;
else x=i-ds;
in_sym[k]=in[y*nx+x];
}
}
/*Precompute patch weight and index */
for(sum=0.,i=0,x=-ds;x<=ds;x++)
for(y=-ds;y<=ds;y++,i++) {
dadr[i] = y*newnx+x;
w[i] = exp(-(double)(x*x+y*y)/(2.*a*a));
sum += w[i];
}
for (i=psize;i--;) w[i] /= 2*sum*h*h;
/* Principal loop */
k=0;
for (y=ds;y<newny-ds;y++)
for (x=ds;x<newnx-ds;x++)
{
adr = y*newnx+x;
realadr=(y-ds)*nx+x-ds;
sum = 0.;
jcs[realadr] = k;
/* loop on patches */
/* Count non zero weight for realadr columns */
for (xp=MAX(x-bloc,ds);xp<=MIN(x+bloc,nx-1+ds);xp++)
for (yp=MAX(y-bloc,ds);yp<=MIN(y+bloc,ny-1+ds);yp++)
{
adrp = yp*newnx+xp;
realadrp=(yp-ds)*nx+xp-ds;
for (i=psize,dist=0.,ww=w,dd=dadr;i--;ww++,dd++) {
v = in_sym[adr+*dd]-in_sym[adrp+*dd];
dist += *ww*(double)(v*v);
}
e = (adrp==adr?c:exp(-dist));
weight[k] = e;
irs[k]=realadrp;
k++;
}
}
jcs[nxny]=k;
free(dadr);
free(w);
free(in_sym);
}
void mexFunction
(
int nargout,
mxArray *pargout [ ],
int nargin,
const mxArray *pargin [ ]
)
{
double dummy = 0, beta [2], *px, *C, *Ct, *C2, *fil, *Zt, *zt, done=1.0, *zz, dzero=0.0;
cholmod_sparse Amatrix, *A, *Lsparse ;
cholmod_factor *L ;
cholmod_common Common, *cm ;
Int minor, *It2, *Jt2 ;
mwIndex l, k2, h, k, i, j, ik, *I, *J, *Jt, *It, *I2, *J2, lfi, *w, *w2, *r;
mwSize nnz, nnzlow, m, n;
int nz = 0;
mwSignedIndex one=1, lfi_si;
mxArray *Am, *Bm;
char *uplo="L", *trans="N";
/* ---------------------------------------------------------------------- */
/* Only one input. We have to find first the Cholesky factorization. */
/* start CHOLMOD and set parameters */
/* ---------------------------------------------------------------------- */
if (nargin == 1) {
cm = &Common ;
cholmod_l_start (cm) ;
sputil_config (SPUMONI, cm) ;
/* convert to packed LDL' when done */
cm->final_asis = FALSE ;
cm->final_super = FALSE ;
cm->final_ll = FALSE ;
cm->final_pack = TRUE ;
cm->final_monotonic = TRUE ;
/* since numerically zero entries are NOT dropped from the symbolic
* pattern, we DO need to drop entries that result from supernodal
* amalgamation. */
cm->final_resymbol = TRUE ;
cm->quick_return_if_not_posdef = (nargout < 2) ;
}
/* This will disable the supernodal LL', which will be slow. */
/* cm->supernodal = CHOLMOD_SIMPLICIAL ; */
/* ---------------------------------------------------------------------- */
/* get inputs */
/* ---------------------------------------------------------------------- */
if (nargin > 3)
{
mexErrMsgTxt ("usage: Z = sinv(A), or Z = sinv(LD, 1)") ;
}
n = mxGetM (pargin [0]) ;
m = mxGetM (pargin [0]) ;
if (!mxIsSparse (pargin [0]))
{
mexErrMsgTxt ("A must be sparse") ;
}
if (n != mxGetN (pargin [0]))
{
mexErrMsgTxt ("A must be square") ;
}
/* Only one input. We have to find first the Cholesky factorization. */
if (nargin == 1) {
/* get sparse matrix A, use tril(A) */
A = sputil_get_sparse (pargin [0], &Amatrix, &dummy, -1) ;
A->stype = -1 ; /* use lower part of A */
beta [0] = 0 ;
beta [1] = 0 ;
/* ---------------------------------------------------------------------- */
/* analyze and factorize */
/* ---------------------------------------------------------------------- */
L = cholmod_l_analyze (A, cm) ;
cholmod_l_factorize_p (A, beta, NULL, 0, L, cm) ;
if (cm->status != CHOLMOD_OK)
{
mexErrMsgTxt ("matrix is not positive definite") ;
}
/* ---------------------------------------------------------------------- */
/* convert L to a sparse matrix */
/* ---------------------------------------------------------------------- */
Lsparse = cholmod_l_factor_to_sparse (L, cm) ;
if (Lsparse->xtype == CHOLMOD_COMPLEX)
{
mexErrMsgTxt ("matrix is complex") ;
}
/* ---------------------------------------------------------------------- */
/* Set the sparse Cholesky factorization in Matlab format */
/* ---------------------------------------------------------------------- */
/*Am = sputil_put_sparse (&Lsparse, cm) ;
I = mxGetIr(Am);
J = mxGetJc(Am);
C = mxGetPr(Am);
nnz = mxGetNzmax(Am); */
It2 = Lsparse->i;
Jt2 = Lsparse->p;
Ct = Lsparse->x;
nnz = (mwSize) Lsparse->nzmax;
Am = mxCreateSparse(m, m, nnz, mxREAL) ;
I = mxGetIr(Am);
J = mxGetJc(Am);
C = mxGetPr(Am);
for (j = 0 ; j < n+1 ; j++) J[j] = (mwIndex) Jt2[j];
for ( i = 0 ; i < nnz ; i++) {
I[i] = (mwIndex) It2[i];
C[i] = Ct[i];
}
cholmod_l_free_sparse (&Lsparse, cm) ;
/*FILE *out1 = fopen( "output1.txt", "w" );
if( out1 != NULL )
fprintf( out1, "Hello %d\n", nnz );
fclose (out1);*/
} else {
/* The cholesky factorization is given as an input. */
/* We have to copy it into workspace */
It = mxGetIr(pargin [0]);
Jt = mxGetJc(pargin [0]);
Ct = mxGetPr(pargin [0]);
nnz = mxGetNzmax(pargin [0]);
Am = mxCreateSparse(m, m, nnz, mxREAL) ;
I = mxGetIr(Am);
J = mxGetJc(Am);
C = mxGetPr(Am);
for (j = 0 ; j < n+1 ; j++) J[j] = Jt[j];
for ( i = 0 ; i < nnz ; i++) {
I[i] = It[i];
C[i] = Ct[i];
}
}
/* Evaluate the sparse inverse */
C[nnz-1] = 1.0/C[J[m-1]]; /* set the last element of sparse inverse */
fil = mxCalloc((mwSize)1,sizeof(double));
zt = mxCalloc((mwSize)1,sizeof(double));
Zt = mxCalloc((mwSize)1,sizeof(double));
zz = mxCalloc((mwSize)1,sizeof(double));
for (j=m-2;j!=-1;j--){
lfi = J[j+1]-(J[j]+1);
/* if (lfi > 0) */
if ( J[j+1] > (J[j]+1) )
{
/* printf("lfi = %u \n ", lfi);
printf("lfi*double = %u \n", (mwSize)lfi*sizeof(double));
printf("lfi*lfi*double = %u \n", (mwSize)lfi*(mwSize)lfi*sizeof(double));
printf("\n \n");
*/
fil = mxRealloc(fil,(mwSize)lfi*sizeof(double));
for (i=0;i<lfi;i++) fil[i] = C[J[j]+i+1]; /* take the j'th lower triangular column of the Cholesky */
zt = mxRealloc(zt,(mwSize)lfi*sizeof(double)); /* memory for the sparse inverse elements to be evaluated */
Zt = mxRealloc(Zt,(mwSize)lfi*(mwSize)lfi*sizeof(double)); /* memory for the needed sparse inverse elements */
/* Set the lower triangular for Zt */
k2 = 0;
for (k=J[j]+1;k<J[j+1];k++){
ik = I[k];
h = k2;
for (l=J[ik];l<=J[ik+1];l++){
if (I[l] == I[ J[j]+h+1 ]){
Zt[h+lfi*k2] = C[l];
h++;
}
}
k2++;
}
/* evaluate zt = fil*Zt */
lfi_si = (mwSignedIndex) lfi;
dsymv(uplo, &lfi_si, &done, Zt, &lfi_si, fil, &one, &dzero, zt, &one);
/* Set the evaluated sparse inverse elements, zt, into C */
k=lfi-1;
for (i = J[j+1]-1; i!=J[j] ; i--){
C[i] = -zt[k];
k--;
}
/* evaluate the j'th diagonal of sparse inverse */
dgemv(trans, &one, &lfi_si, &done, fil, &one, zt, &one, &dzero, zz, &one);
C[J[j]] = 1.0/C[J[j]] + zz[0];
}
else
{
/* evaluate the j'th diagonal of sparse inverse */
C[J[j]] = 1.0/C[J[j]];
}
}
/* Free the temporary variables */
mxFree(fil);
mxFree(zt);
mxFree(Zt);
mxFree(zz);
/* ---------------------------------------------------------------------- */
/* Permute the elements according to r(q) = 1:n */
/* Done only if the Cholesky was evaluated here */
/* ---------------------------------------------------------------------- */
if (nargin == 1) {
Bm = mxCreateSparse(m, m, nnz, mxREAL) ;
It = mxGetIr(Bm);
Jt = mxGetJc(Bm);
Ct = mxGetPr(Bm); /* Ct = C(r,r) */
r = (mwIndex *) L->Perm; /* fill reducing ordering */
w = mxCalloc(m,sizeof(mwIndex)); /* column counts of Am */
/* count entries in each column of Bm */
for (j=0; j<m; j++){
k = r ? r[j] : j ; /* column j of Bm is column k of Am */
for (l=J[j] ; l<J[j+1] ; l++){
i = I[l];
ik = r ? r[i] : i ; /* row i of Bm is row ik of Am */
w[ max(ik,k) ]++;
}
}
cumsum2(Jt, w, m);
for (j=0; j<m; j++){
k = r ? r[j] : j ; /* column j of Bm is column k of Am */
for (l=J[j] ; l<J[j+1] ; l++){
i= I[l];
ik = r ? r[i] : i ; /* row i of Bm is row ik of Am */
It [k2 = w[max(ik,k)]++ ] = min(ik,k);
Ct[k2] = C[l];
}
}
mxFree(w);
/* ---------------------------------------------------------------------- */
/* Transpose the permuted (upper triangular) matrix Bm into Am */
/* (this way we get sorted columns) */
/* ---------------------------------------------------------------------- */
w = mxCalloc(m,sizeof(mwIndex));
for (i=0 ; i<Jt[m] ; i++) w[It[i]]++; /* row counts of Bm */
cumsum2(J, w, m); /* row pointers */
for (j=0 ; j<m ; j++){
for (i=Jt[j] ; i<Jt[j+1] ; i++){
I[ l=w[ It[i] ]++ ] = j;
C[l] = Ct[i];
}
}
mxFree(w);
mxDestroyArray(Bm);
}
/* ---------------------------------------------------------------------- */
/* Fill the upper triangle of the sparse inverse */
/* ---------------------------------------------------------------------- */
w = mxCalloc(m,sizeof(mwIndex)); /* workspace */
w2 = mxCalloc(m,sizeof(mwIndex)); /* workspace */
for (k=0;k<J[m];k++) w[I[k]]++; /* row counts of the lower triangular */
for (k=0;k<m;k++) w2[k] = w[k] + J[k+1] - J[k] - 1; /* column counts of the sparse inverse */
nnz = (mwSize)2*nnz - m; /* The number of nonzeros in Z */
pargout[0] = mxCreateSparse(m,m,nnz,mxREAL); /* The sparse matrix */
It = mxGetIr(pargout[0]);
Jt = mxGetJc(pargout[0]);
Ct = mxGetPr(pargout[0]);
cumsum2(Jt, w2, m); /* column starting points */
for (j = 0 ; j < m ; j++){ /* fill the upper triangular */
for (k = J[j] ; k < J[j+1] ; k++){
It[l = w2[ I[k]]++] = j ; /* place C(i,j) as entry Ct(j,i) */
if (Ct) Ct[l] = C[k] ;
}
}
for (j = 0 ; j < m ; j++){ /* fill the lower triangular */
for (k = J[j]+1 ; k < J[j+1] ; k++){
It[l = w2[j]++] = I[k] ; /* place C(j,i) as entry Ct(j,i) */
if (Ct) Ct[l] = C[k] ;
}
}
mxFree(w2);
mxFree(w);
/* ---------------------------------------------------------------------- */
/* return to MATLAB */
/* ---------------------------------------------------------------------- */
/* ---------------------------------------------------------------------- */
/* free workspace and the CHOLMOD L, except for what is copied to MATLAB */
/* ---------------------------------------------------------------------- */
if (nargin == 1) {
cholmod_l_free_factor (&L, cm) ;
cholmod_l_finish (cm) ;
cholmod_l_print_common (" ", cm) ;
}
mxDestroyArray(Am);
}
void mexFunction(int nlhs, mxArray *plhs[],int nrhs, const mxArray *prhs[])
{
double *v, *x, sigma, *lambda, *pr; mwIndex *ir, *jc;
char *mode; int imode;
//Check Inputs
if(nrhs < 1) {
printInfo();
return;
}
if(nrhs < 2) {
mexErrMsgTxt("You must supply the callback mode and input vector.");
return;
}
if(mxIsEmpty(prhs[0]) || !mxIsChar(prhs[0])) {
mexErrMsgTxt("The mode must be a string!");
return;
}
if(!mxIsEmpty(prhs[1])) {
if(mxIsClass(prhs[1],"scipvar") || mxIsClass(prhs[1],"barvec")) {
mexErrMsgTxt("SCIP and BARON cannot be used with this callback function - please specify 'mcode' via symbset as the cbmode.");
return;
}
if(!mxIsDouble(prhs[1]) || mxIsComplex(prhs[1]) || mxIsSparse(prhs[1])) {
mexErrMsgTxt("The input vector must be a dense real double vector!");
return;
}
}
else {
mexErrMsgTxt("The input vector must be a dense real double vector!");
return;
}
//Check x input size
if(mxGetNumberOfElements(prhs[1]) != getNoVar()) {
mexErrMsgTxt("The input vector is not the right size!");
}
//Get x
x = mxGetPr(prhs[1]);
//Determine input mode and setup return variable
mode = mxArrayToString(prhs[0]);
lower(mode);
if(!strcmp(mode,"obj")) {
imode = 0;
plhs[0] = mxCreateDoubleMatrix(1,1, mxREAL);
v = mxGetPr(plhs[0]);
}
else if(!strcmp(mode,"grad")) {
imode = 1;
plhs[0] = mxCreateDoubleMatrix(1,getNoVar(), mxREAL);
v = mxGetPr(plhs[0]);
}
else if(!strcmp(mode,"con")) {
imode = 2;
plhs[0] = mxCreateDoubleMatrix(getNoCon(),1, mxREAL);
v = mxGetPr(plhs[0]);
}
else if(!strcmp(mode,"jac")) {
imode = 3;
plhs[0] = mxCreateSparse(getNoCon(),getNoVar(),getNNZJac(),mxREAL);
pr = mxGetPr(plhs[0]);
ir = mxGetIr(plhs[0]);
jc = mxGetJc(plhs[0]);
}
else if(!strcmp(mode,"hess")) {
if(nrhs < 4) {
mexErrMsgTxt("You must supply the callback mode, input vector, sigma and lambda for Hessian Evaluations.");
return;
}
//Check length of Sigma
if(mxIsEmpty(prhs[2]) || !mxIsDouble(prhs[2]) || mxIsComplex(prhs[2]) || mxGetNumberOfElements(prhs[2]) != 1)
mexErrMsgTxt("Sigma must be a real, double scalar.");
//Check length of Lambda
if(!mxIsDouble(prhs[3]) || mxIsComplex(prhs[3]) || mxIsSparse(prhs[3]) || mxGetNumberOfElements(prhs[3]) != getNoCon())
mexErrMsgTxt("Lambda must be a real, double, dense vector with ncon elements.");
//Get Sigma, Lambda
sigma = *mxGetPr(prhs[2]);
lambda = mxGetPr(prhs[3]);
imode = 4;
plhs[0] = mxCreateSparse(getNoVar(),getNoVar(),getNNZHess(),mxREAL);
pr = mxGetPr(plhs[0]);
ir = mxGetIr(plhs[0]);
jc = mxGetJc(plhs[0]);
}
else
mexErrMsgTxt("Unknown mode - options are 'obj', 'grad', 'con', 'jac', or 'hess'");
mxFree(mode);
//Call Req Callback
switch(imode)
{
case 0: //objective
*v = objective(x);
break;
case 1: //gradient
gradient(x,v);
break;
case 2: //constraints
constraints(x,v);
break;
case 3: //jacobian
jacobian(x,pr,ir,jc);
break;
case 4: //hessian
hessian(x,sigma,lambda,pr,ir,jc);
break;
}
}
void
mexFunction(int nlhs, mxArray *plhs[], int nrhs, const mxArray *prhs[])
{
ASL_pfgh *asl = (ASL_pfgh*)cur_ASL;
FILE *nl;
Jmp_buf err_jmp0;
cgrad *cg, **cgp;
char *buf1, buf[512], *what, **whatp;
fint *hcs, *hr, i, nerror;
int *cs;
mwIndex *Ir, *Jc;
real *H, *He, *J1, *W, *c, *f, *g, *v, *t, *x;
static fint n, nc, nhnz, nz;
static real *Hsp;
static char ignore_complementarity[] =
"Warning: ignoring %d complementarity conditions.\n";
if (nrhs == 1 && mxIsChar(prhs[0])) {
if (nlhs < 6 || nlhs > 7)
usage();
if (mxGetString(prhs[0], buf1 = buf, sizeof(buf)))
mexErrMsgTxt("Expected 'stub' as argument\n");
at_end();
mexAtExit(at_end);
asl = (ASL_pfgh*)ASL_alloc(ASL_read_pfgh);
return_nofile = 1;
if (!(nl = jac0dim(buf1,strlen(buf)))) {
sprintf(msgbuf, "Can't open %.*s\n",
sizeof(msgbuf)-20, buf);
mexErrMsgTxt(msgbuf);
}
if (n_obj <= 0)
printf("Warning: objectve == 0\n");
n = n_var;
nc = n_con;
nz = nzc;
X0 = mxGetPr(plhs[0] = mxCreateDoubleMatrix(n, 1, mxREAL));
LUv = mxGetPr(plhs[1] = mxCreateDoubleMatrix(n, 1, mxREAL));
Uvx = mxGetPr(plhs[2] = mxCreateDoubleMatrix(n, 1, mxREAL));
pi0 = mxGetPr(plhs[3] = mxCreateDoubleMatrix(nc, 1, mxREAL));
LUrhs = mxGetPr(plhs[4] = mxCreateDoubleMatrix(nc, 1, mxREAL));
Urhsx = mxGetPr(plhs[5] = mxCreateDoubleMatrix(nc, 1, mxREAL));
if (nlhs == 7) {
cvar = (int*)M1alloc(nc*sizeof(int));
plhs[6] = mxCreateDoubleMatrix(nc, 1, mxREAL);
x = mxGetPr(plhs[6]);
}
else if (n_cc)
printf(ignore_complementarity, n_cc);
pfgh_read(nl, ASL_findgroups);
if (nlhs == 7)
for(i = 0; i < nc; i++)
x[i] = cvar[i];
/* Arrange to compute the whole sparese Hessian */
/* of the Lagrangian function (both triangles). */
nhnz = sphsetup(0, 0, nc > 0, 0);
Hsp = (real *)M1alloc(nhnz*sizeof(real));
return;
}
if (!asl)
mexErrMsgTxt("spamfunc(\"stub\") has not been called\n");
nerror = -1;
err_jmp1 = &err_jmp0;
what = "(?)";
whatp = &what;
if (nlhs == 2) {
if (nrhs != 2)
usage();
x = sizechk(prhs[0],"x",n);
t = sizechk(prhs[1],"0 or 1", 1);
if (setjmp(err_jmp0.jb)) {
sprintf(msgbuf, "Trouble evaluating %s\n", *whatp);
mexErrMsgTxt(msgbuf);
}
if (t[0] == 0.) {
f = mxGetPr(plhs[0] = mxCreateDoubleMatrix(1, 1, mxREAL));
c = mxGetPr(plhs[1] = mxCreateDoubleMatrix(nc, 1, mxREAL));
what = "f";
*f = objval(0, x, &nerror);
what = "c";
conval(x, c, &nerror);
return;
}
g = mxGetPr(plhs[0] = mxCreateDoubleMatrix(n, 1, mxREAL));
J1 = mxGetPr(plhs[1] = mxCreateSparse(nc, n, nz, mxREAL));
what = "g";
objgrd(0, x, g, &nerror);
if (nc) {
what = "J";
jacval(x, J1, &nerror);
Ir = mxGetIr(plhs[1]);
/*memcpy(mxGetJc(plhs[1]), A_colstarts, (n+1)*sizeof(int));*/
for(Jc = mxGetJc(plhs[1]), cs = A_colstarts, i = 0; i <= n; ++i)
Jc[i] = cs[i];
cgp = Cgrad;
for(i = 0; i < nc; i++)
for(cg = *cgp++; cg; cg = cg->next)
Ir[cg->goff] = i;
}
return;
}
if (nlhs == 0 && (nrhs == 3 || nrhs == 4)) {
/* eval2('solution message', x, v): x = primal, v = dual */
/* optional 4th arg = solve_result_num */
if (!mxIsChar(prhs[0]))
usage();
x = sizechk(prhs[1],"x",n);
v = sizechk(prhs[2],"v",nc);
if (mxGetString(prhs[0], buf, sizeof(buf)))
mexErrMsgTxt(
"Expected 'solution message' as first argument\n");
solve_result_num = nrhs == 3 ? -1 /* unknown */
: (int)*sizechk(prhs[3],"solve_result_num",1);
write_sol(buf, x, v, 0);
return;
}
if (nlhs != 1 || nrhs != 1)
usage();
v = sizechk(prhs[0],"v",nc);
W = mxGetPr(plhs[0] = mxCreateSparse(n, n, nhnz, mxREAL));
what = "W";
sphes(H = Hsp, 0, 0, v);
/* Expand the Hessian lower triangle into the full Hessian... */
Ir = mxGetIr(plhs[0]);
Jc = mxGetJc(plhs[0]);
hcs = sputinfo->hcolstarts;
hr = sputinfo->hrownos;
for(i = 0; i <= n; i++)
Jc[i] = hcs[i];
He = H + hcs[n];
while(H < He) {
*W++ = *H++;
*Ir++ = *hr++;
}
}
static mxArray *readdata(file_t *fp, mxArray **header, int start, int howmany){
/*
if start!=0 || howmany!=0 and data is cell, will only read cell from start to start+howmany
if data is not cell and start==-1, will skip the data.
Only the first call to readdata will possibly have howmany!=0
*/
if(fp->eof) return NULL;
header_t header2;
if(read_header2(&header2, fp)){
return NULL;
}
uint32_t magic=header2.magic & 0xFFFF;
if(magic==0){//end of file or empty file
fp->eof=1;
return NULL;
}
if(header){
if(header2.str)
*header=mxCreateString(header2.str);
else
*header=mxCreateString("");
}
free(header2.str); header2.str=NULL;
long nx=header2.nx;
long ny=header2.ny;
int start_save=start;
if(iscell(magic)){
if(start!=-1 && howmany==0){
if(start!=0){
error("Invalid use\n");
}
howmany=nx*ny-start;
}
}else if(fp->isfits){
if(howmany!=0){
/*first read of fits file. determine if we need it*/
if(start>0){
start=-1;//skip this block.
}
}
}else{
if((start!=0 && start!=-1) || howmany!=0){
error("invalid use");
}
}
int iscell=0;
if(fp->eof) return NULL;
mxArray *out=NULL;
switch(magic){
case MCC_ANY:
case MCC_DBL:
case MCC_CMP:
case MC_CSP:
case MC_SP:
case MC_DBL:
case MC_CMP:
case MC_INT32:
case MC_INT64:
{
iscell=1;
mwIndex ix;
if(fp->eof) return NULL;
if(nx*ny>1 && skip_unicell){
out=mxCreateCellMatrix(nx,ny);
}
mxArray *header0=mxCreateCellMatrix(nx*ny+1,1);
for(ix=0; ix<nx*ny; ix++){
int start2=0;
if(start==-1 || ix<start || ix>=start+howmany){
start2=-1;
}
mxArray *header3=NULL;
mxArray *tmp=readdata(fp, &header3, start2, 0);
if(fp->eof){
break;
}
if(tmp && tmp!=SKIPPED){
if(nx*ny==1 && skip_unicell){//only one entry
out=tmp;
}else{
mxSetCell(out, ix, tmp);
}
if(header3){
mxSetCell(header0, ix, header3);
}
}
}
if(header){
mxSetCell(header0, nx*ny, *header);
*header=header0;
}
}
break;
case M_SP64:
case M_SP32:
{
if(start==-1) error("Invalid use\n");
size_t size;
if(magic==M_SP32){
size=4;
}else if(magic==M_SP64){
size=8;
}else{
size=0;
error("Invalid magic\n");
}
uint64_t nzmax;
if(nx!=0 && ny!=0){
zfread(&nzmax,sizeof(uint64_t),1,fp);
}else{
nzmax=0;
}
if(fp->eof) return NULL;
out=mxCreateSparse(nx,ny,nzmax,mxREAL);
if(nx!=0 && ny!=0 && nzmax!=0){
if(sizeof(mwIndex)==size){/*Match*/
zfread(mxGetJc(out), size,ny+1,fp);
zfread(mxGetIr(out), size,nzmax, fp);
}else{
long i;
mwIndex *Jc0=mxGetJc(out);
mwIndex *Ir0=mxGetIr(out);
void *Jc=malloc(size*(ny+1));
void *Ir=malloc(size*nzmax);
zfread(Jc, size, ny+1, fp);
zfread(Ir, size, nzmax, fp);
if(size==4){
uint32_t* Jc2=Jc;
uint32_t* Ir2=Ir;
for(i=0; i<ny+1; i++){
Jc0[i]=Jc2[i];
}
for(i=0; i<nzmax; i++){
Ir0[i]=Ir2[i];
}
free(Jc);
free(Ir);
}else if(size==8){
uint64_t* Jc2=Jc;
uint64_t* Ir2=Ir;
for(i=0; i<ny+1; i++){
Jc0[i]=Jc2[i];
}
for(i=0; i<nzmax; i++){
Ir0[i]=Ir2[i];
}
free(Jc);
free(Ir);
}else{
mexErrMsgTxt("Invalid sparse format\n");
}
}
zfread(mxGetPr(out), sizeof(double), nzmax, fp);
}
}
break;
case M_CSP64:
case M_CSP32:/*complex sparse*/
{
if(start==-1) error("Invalid use\n");
size_t size;
switch(magic){
case M_CSP32:
size=4;break;
case M_CSP64:
size=8;break;
default:
size=0;
}
uint64_t nzmax;
if(nx!=0 && ny!=0){
zfread(&nzmax,sizeof(uint64_t),1,fp);
}else{
nzmax=0;
}
if(fp->eof) return NULL;
out=mxCreateSparse(nx,ny,nzmax,mxCOMPLEX);
if(nx!=0 && ny!=0){
long i;
if(sizeof(mwIndex)==size){
zfread(mxGetJc(out), size,ny+1,fp);
zfread(mxGetIr(out), size,nzmax, fp);
}else{
mwIndex *Jc0=mxGetJc(out);
mwIndex *Ir0=mxGetIr(out);
void *Jc=malloc(size*(ny+1));
void *Ir=malloc(size*nzmax);
zfread(Jc, size, ny+1, fp);
zfread(Ir, size, nzmax, fp);
if(size==4){
uint32_t* Jc2=Jc;
uint32_t* Ir2=Ir;
for(i=0; i<ny+1; i++){
Jc0[i]=Jc2[i];
}
for(i=0; i<nzmax; i++){
Ir0[i]=Ir2[i];
}
free(Jc);
free(Ir);
}else if(size==8){
uint64_t* Jc2=Jc;
uint64_t* Ir2=Ir;
for(i=0; i<ny+1; i++){
Jc0[i]=Jc2[i];
}
for(i=0; i<nzmax; i++){
Ir0[i]=Ir2[i];
}
free(Jc);
free(Ir);
}else{
info("size=%lu\n", size);
error("Invalid sparse format\n");
}
}
dcomplex *tmp=malloc(sizeof(dcomplex)*nzmax);
zfread(tmp, sizeof(dcomplex), nzmax, fp);
double *Pr=mxGetPr(out);
double *Pi=mxGetPi(out);
for(i=0; i<nzmax; i++){
Pr[i]=tmp[i].x;
Pi[i]=tmp[i].y;
}
free(tmp);
}
}
break;
case M_DBL:/*double array*/
if(start==-1){
if(zfseek(fp, sizeof(double)*nx*ny, SEEK_CUR)){
error("Seek failed\n");
}
out=SKIPPED;
}else{
out=mxCreateDoubleMatrix(nx,ny,mxREAL);
if(nx!=0 && ny!=0){
zfread(mxGetPr(out), sizeof(double),nx*ny,fp);
}
}
break;
case M_FLT:/*float array*/
if(start==-1){
if(zfseek(fp, sizeof(float)*nx*ny, SEEK_CUR)){
error("Seek failed\n");
}
out=SKIPPED;
}else{
mwSize nxy[2]={nx,ny};
out=mxCreateNumericArray(2, nxy,mxSINGLE_CLASS, mxREAL);
if(nx!=0 && ny!=0){
zfread((float*)mxGetPr(out), sizeof(float),nx*ny, fp);
}
}break;
case M_INT64:/*long array*/
case M_INT32:
case M_INT16:
case M_INT8:
{
int byte=0;
mxClassID id;
switch(magic){
case M_INT64: byte=8; id=mxINT64_CLASS; break;
case M_INT32: byte=4; id=mxINT32_CLASS; break;
case M_INT16: byte=2; id=mxINT16_CLASS; break;
case M_INT8: byte=1; id=mxINT8_CLASS; break;
default: id=0;
}
if(start==-1){
if(zfseek(fp, byte*nx*ny, SEEK_CUR)){
error("Seek failed\n");
}
out=SKIPPED;
}else{
out=mxCreateNumericMatrix(nx,ny,id,mxREAL);
if(nx!=0 && ny!=0){
/*Don't use sizeof(mxINT64_CLASS), it is just an integer,
not a valid C type.*/
zfread(mxGetPr(out), byte,nx*ny,fp);
}
}
}
break;
case M_CMP:/*double complex array*/
if(start==-1){
if(zfseek(fp, 16*nx*ny, SEEK_CUR)){
error("Seek failed\n");
}
out=SKIPPED;
}else{
out=mxCreateDoubleMatrix(nx,ny,mxCOMPLEX);
if(nx!=0 && ny!=0){
dcomplex*tmp=malloc(sizeof(dcomplex)*nx*ny);
zfread(tmp,sizeof(dcomplex),nx*ny,fp);
double *Pr=mxGetPr(out);
double *Pi=mxGetPi(out);
long i;
for(i=0; i<nx*ny; i++){
Pr[i]=tmp[i].x;
Pi[i]=tmp[i].y;
}
free(tmp);
}
}
break;
case M_ZMP:/*float complex array. convert to double*/
if(start==-1){
if(zfseek(fp, 8*nx*ny, SEEK_CUR)){
error("Seek failed\n");
}
out=SKIPPED;
}else{
mwSize nxy[2]={nx, ny};
out=mxCreateNumericArray(2, nxy,mxSINGLE_CLASS, mxCOMPLEX);
if(nx!=0 && ny!=0){
fcomplex*tmp=malloc(sizeof(fcomplex)*nx*ny);
zfread(tmp,sizeof(fcomplex),nx*ny,fp);
float *Pr=(float*)mxGetPr(out);
float *Pi=(float*)mxGetPi(out);
long i;
for(i=0; i<nx*ny; i++){
Pr[i]=tmp[i].x;
Pi[i]=tmp[i].y;
}
free(tmp);
}
}
break;
case M_HEADER:
break;
default:
fprintf(stderr,"magic=%x\n",magic);
warning("Unrecognized file. Please recompile the mex routines in the newest code\n");
out=NULL;
}
start=start_save;
if(!iscell && fp->isfits==1){/*fits file may contain extra extensions.*/
fp->isfits++;
int icell=0;
mxArray **outarr=NULL;
mxArray **headerarr=NULL;
while(out){
icell++;
outarr=realloc(outarr, sizeof(mxArray*)*icell);
headerarr=realloc(headerarr, sizeof(mxArray*)*icell);
if(out==SKIPPED) out=NULL;
outarr[icell-1]=out;
if(header) headerarr[icell-1]=*header;
int start2=0;
if(howmany!=0){//selective reading.
if(icell<start || icell+1>start+howmany){
start2=-1;//don't read next data.
}
}
out=readdata(fp, header, start2, 0);
}
if(icell>1){/*set output.*/
out=mxCreateCellMatrix(icell, 1);
if(header) *header=mxCreateCellMatrix(icell, 1);
int i;
for(i=0; i<icell; i++){
mxSetCell(out, i, outarr[i]);
if(header) mxSetCell(*header, i, headerarr[i]);
}
free(outarr);
free(headerarr);
}else{
out=outarr[0];
if(header) *header=headerarr[0];
}
}
if(!out){
out=mxCreateDoubleMatrix(0,0,mxREAL);
}
return out;
}
/* ************************************************************
PROCEDURE mexFunction - Entry for Matlab
y = fwblksolve(L,b, [y])
y = L.L \ b(L.perm)
************************************************************ */
void mexFunction(const int nlhs, mxArray *plhs[],
const int nrhs, const mxArray *prhs[])
{
const mxArray *L_FIELD;
mwIndex m,n, j, k, nsuper, inz;
double *y,*fwork;
const double *permPr, *b, *xsuperPr;
const mwIndex *yjc, *yir, *bjc, *bir;
mwIndex *perm, *invperm, *snode, *xsuper, *iwork;
jcir L;
char bissparse;
/* ------------------------------------------------------------
Check for proper number of arguments
------------------------------------------------------------ */
mxAssert(nrhs >= MINNPARIN, "fwblkslv requires more input arguments.");
mxAssert(nlhs <= NPAROUT, "fwblkslv generates only 1 output argument.");
/* ------------------------------------------------------------
Disassemble block Cholesky structure L
------------------------------------------------------------ */
mxAssert(mxIsStruct(L_IN), "Parameter `L' should be a structure.");
if( (L_FIELD = mxGetField(L_IN,(mwIndex)0,"perm")) == NULL) /* L.perm */
mexErrMsgTxt("Missing field L.perm.");
m = mxGetM(L_FIELD) * mxGetN(L_FIELD);
permPr = mxGetPr(L_FIELD);
if( (L_FIELD = mxGetField(L_IN,(mwIndex)0,"L")) == NULL) /* L.L */
mexErrMsgTxt("Missing field L.L.");
if( m != mxGetM(L_FIELD) || m != mxGetN(L_FIELD) )
mexErrMsgTxt("Size L.L mismatch.");
if(!mxIsSparse(L_FIELD))
mexErrMsgTxt("L.L should be sparse.");
L.jc = mxGetJc(L_FIELD);
L.ir = mxGetIr(L_FIELD);
L.pr = mxGetPr(L_FIELD);
if( (L_FIELD = mxGetField(L_IN,(mwIndex)0,"xsuper")) == NULL) /* L.xsuper */
mexErrMsgTxt("Missing field L.xsuper.");
nsuper = mxGetM(L_FIELD) * mxGetN(L_FIELD) - 1;
if( nsuper > m )
mexErrMsgTxt("Size L.xsuper mismatch.");
xsuperPr = mxGetPr(L_FIELD);
/* ------------------------------------------------------------
Get rhs matrix b.
If it is sparse, then we also need the sparsity structure of y.
------------------------------------------------------------ */
b = mxGetPr(B_IN);
if( mxGetM(B_IN) != m )
mexErrMsgTxt("Size mismatch b.");
n = mxGetN(B_IN);
if( (bissparse = mxIsSparse(B_IN)) ){
bjc = mxGetJc(B_IN);
bir = mxGetIr(B_IN);
if(nrhs < NPARIN)
mexErrMsgTxt("fwblkslv requires more inputs in case of sparse b.");
if(mxGetM(Y_IN) != m || mxGetN(Y_IN) != n)
mexErrMsgTxt("Size mismatch y.");
if(!mxIsSparse(Y_IN))
mexErrMsgTxt("y should be sparse.");
}
/* ------------------------------------------------------------
Allocate output y. If bissparse, then Y_IN gives the sparsity structure.
------------------------------------------------------------ */
if(!bissparse)
Y_OUT = mxCreateDoubleMatrix(m, n, mxREAL);
else{
yjc = mxGetJc(Y_IN);
yir = mxGetIr(Y_IN);
Y_OUT = mxCreateSparse(m,n, yjc[n],mxREAL);
memcpy(mxGetJc(Y_OUT), yjc, (n+1) * sizeof(mwIndex));
memcpy(mxGetIr(Y_OUT), yir, yjc[n] * sizeof(mwIndex));
}
y = mxGetPr(Y_OUT);
/* ------------------------------------------------------------
Allocate working arrays fwork(m) and iwork(2*m + nsuper+1)
------------------------------------------------------------ */
fwork = (double *) mxCalloc(m, sizeof(double));
iwork = (mwIndex *) mxCalloc(2*m+nsuper+1, sizeof(mwIndex));
perm = iwork;
invperm = perm;
xsuper = iwork + m;
snode = xsuper + (nsuper+1);
/* ------------------------------------------------------------
Convert real to integer array, and from Fortran to C style.
In case of sparse b, we store the inverse perm, instead of perm itself.
------------------------------------------------------------ */
for(k = 0; k <= nsuper; k++)
xsuper[k] = xsuperPr[k] - 1;
if(!bissparse)
for(k = 0; k < m; k++) /* Get perm if !bissparse */
perm[k] = permPr[k] - 1;
else{
for(k = 0; k < m; k++){ /* Get invperm if bissparse */
j = permPr[k];
invperm[--j] = k;
}
/* ------------------------------------------------------------
In case of sparse b, we also create snode, which maps each subnode
to the supernode containing it.
------------------------------------------------------------ */
for(j = 0, k = 0; k < nsuper; k++)
while(j < xsuper[k+1])
snode[j++] = k;
}
/* ------------------------------------------------------------
The actual job is done here: y = L\b(perm).
------------------------------------------------------------ */
if(!bissparse)
for(j = 0; j < n; j++){
for(k = 0; k < m; k++) /* y = b(perm) */
y[k] = b[perm[k]];
fwsolve(y,L.jc,L.ir,L.pr,xsuper,nsuper,fwork);
y += m; b += m;
}
else
for(j = 0, inz = 0; j < n; j++){
for(k = inz; k < yjc[j+1]; k++) /* fwork = all-0 */
fwork[yir[k]] = 0.0;
for(k = bjc[j]; k < bjc[j+1]; k++) /* fwork = b(perm) */
fwork[invperm[bir[k]]] = b[k];
selfwsolve(fwork,L.jc,L.ir,L.pr,xsuper,nsuper, snode,
yir+inz,yjc[j+1]-inz);
for(; inz < yjc[j+1]; inz++)
y[inz] = fwork[yir[inz]];
}
/* ------------------------------------------------------------
RELEASE WORKING ARRAYS.
------------------------------------------------------------ */
mxFree(iwork);
mxFree(fwork);
}
const char *model_to_matlab_structure(mxArray *plhs[], int num_of_feature, struct svm_model *model)
{
int i, j, n;
SVM_REAL *ptr;
mxArray *return_model, **rhs;
int out_id = 0;
rhs = (mxArray **)mxMalloc(sizeof(mxArray *)*NUM_OF_RETURN_FIELD);
/* Parameters */
#ifdef _ALL_FLOAT
rhs[out_id] = mxCreateNumericMatrix(5, 1, mxSINGLE_CLASS, mxREAL);
ptr = (float *) mxGetData(rhs[out_id]);
#else
rhs[out_id] = mxCreateDoubleMatrix(5, 1, mxREAL);
ptr = mxGetPr(rhs[out_id]);
#endif
ptr[0] = (SVM_REAL)model->param.svm_type;
ptr[1] = (SVM_REAL)model->param.kernel_type;
ptr[2] = (SVM_REAL)model->param.degree;
ptr[3] = model->param.gamma;
ptr[4] = model->param.coef0;
out_id++;
/* nr_class */
#ifdef _ALL_FLOAT
rhs[out_id] = mxCreateNumericMatrix(1, 1, mxSINGLE_CLASS,mxREAL);
ptr = (float *) mxGetData(rhs[out_id]);
#else
rhs[out_id] = mxCreateDoubleMatrix(1, 1, mxREAL);
ptr = mxGetPr(rhs[out_id]);
#endif
ptr[0] = (SVM_REAL)model->nr_class;
out_id++;
/* total SV */
#ifdef _ALL_FLOAT
rhs[out_id] = mxCreateNumericMatrix(1, 1, mxSINGLE_CLASS,mxREAL);
ptr = (float *) mxGetData(rhs[out_id]);
#else
rhs[out_id] = mxCreateDoubleMatrix(1, 1, mxREAL);
ptr = mxGetPr(rhs[out_id]);
#endif
ptr[0] = (SVM_REAL)model->l;
out_id++;
/* rho */
n = model->nr_class*(model->nr_class-1)/2;
#ifdef _ALL_FLOAT
rhs[out_id] = mxCreateNumericMatrix(n, 1, mxSINGLE_CLASS,mxREAL);
ptr = (float *) mxGetData(rhs[out_id]);
#else
rhs[out_id] = mxCreateDoubleMatrix(n, 1, mxREAL);
ptr = mxGetPr(rhs[out_id]);
#endif
for(i = 0; i < n; i++)
ptr[i] = model->rho[i];
out_id++;
/* Label */
if(model->label)
{
#ifdef _ALL_FLOAT
rhs[out_id] = mxCreateNumericMatrix(model->nr_class, 1, mxSINGLE_CLASS,mxREAL);
ptr = (float *) mxGetData(rhs[out_id]);
#else
rhs[out_id] = mxCreateDoubleMatrix(model->nr_class, 1, mxREAL);
ptr = mxGetPr(rhs[out_id]);
#endif
for(i = 0; i < model->nr_class; i++)
ptr[i] = (SVM_REAL)model->label[i];
}
else
#ifdef _ALL_FLOAT
rhs[out_id] = mxCreateNumericMatrix(0, 0, mxSINGLE_CLASS,mxREAL);
#else
rhs[out_id] = mxCreateDoubleMatrix(0, 0, mxREAL);
#endif
out_id++;
/* probA */
if(model->probA != NULL)
{
#ifdef _ALL_FLOAT
rhs[out_id] = mxCreateNumericMatrix(n, 1, mxSINGLE_CLASS,mxREAL);
ptr = (float *) mxGetData(rhs[out_id]);
#else
rhs[out_id] = mxCreateDoubleMatrix(n, 1, mxREAL);
ptr = mxGetPr(rhs[out_id]);
#endif
for(i = 0; i < n; i++)
ptr[i] = model->probA[i];
}
else
#ifdef _ALL_FLOAT
rhs[out_id] = mxCreateNumericMatrix(0, 0, mxSINGLE_CLASS,mxREAL);
#else
rhs[out_id] = mxCreateDoubleMatrix(0, 0, mxREAL);
#endif
out_id ++;
/* probB */
if(model->probB != NULL)
{
#ifdef _ALL_FLOAT
rhs[out_id] = mxCreateNumericMatrix(n, 1, mxSINGLE_CLASS,mxREAL);
ptr = (float *) mxGetData(rhs[out_id]);
#else
rhs[out_id] = mxCreateDoubleMatrix(n, 1, mxREAL);
ptr = mxGetPr(rhs[out_id]);
#endif
for(i = 0; i < n; i++)
ptr[i] = model->probB[i];
}
else
#ifdef _ALL_FLOAT
rhs[out_id] = mxCreateNumericMatrix(0, 0, mxSINGLE_CLASS,mxREAL);
#else
rhs[out_id] = mxCreateDoubleMatrix(0, 0, mxREAL);
#endif
out_id++;
/* nSV */
if(model->nSV)
{
#ifdef _ALL_FLOAT
rhs[out_id] = mxCreateNumericMatrix(model->nr_class, 1, mxSINGLE_CLASS,mxREAL);
ptr = (float *) mxGetData(rhs[out_id]);
#else
rhs[out_id] = mxCreateDoubleMatrix(model->nr_class, 1, mxREAL);
ptr = mxGetPr(rhs[out_id]);
#endif
for(i = 0; i < model->nr_class; i++)
ptr[i] = (SVM_REAL)model->nSV[i];
}
else
#ifdef _ALL_FLOAT
rhs[out_id] = mxCreateNumericMatrix(0, 0, mxSINGLE_CLASS,mxREAL);
#else
rhs[out_id] = mxCreateDoubleMatrix(0, 0, mxREAL);
#endif
out_id++;
/* sv_coef */
#ifdef _ALL_FLOAT
rhs[out_id] = mxCreateNumericMatrix(model->l, model->nr_class-1, mxSINGLE_CLASS,mxREAL);
ptr = (float *) mxGetData(rhs[out_id]);
#else
rhs[out_id] = mxCreateDoubleMatrix(model->l, model->nr_class-1, mxREAL);
ptr = mxGetPr(rhs[out_id]);
#endif
for(i = 0; i < model->nr_class-1; i++)
for(j = 0; j < model->l; j++)
ptr[(i*(model->l))+j] = model->sv_coef[i][j];
out_id++;
/* SVs */
{
#ifdef _DENSE_REP
#ifdef _ALL_FLOAT
rhs[out_id] = mxCreateNumericMatrix(model->l, num_of_feature, mxSINGLE_CLASS,mxREAL);
ptr = (float *) mxGetData(rhs[out_id]);
#else
rhs[out_id] = mxCreateDoubleMatrix(model->l, num_of_feature, mxREAL);
ptr = mxGetPr(rhs[out_id]);
#endif
for(i=0;i < num_of_feature;i++)
for(j=0;j < model->l;j++)
ptr[(i*(model->l))+j] = model->SV[j].values[i];
#else
int ir_index, nonzero_element;
mwIndex *ir, *jc;
mxArray *pprhs[1], *pplhs[1];
if(model->param.kernel_type == PRECOMPUTED)
{
nonzero_element = model->l;
num_of_feature = 1;
}
else
{
nonzero_element = 0;
for(i = 0; i < model->l; i++) {
j = 0;
while(model->SV[i][j].index != -1)
{
nonzero_element++;
j++;
}
}
}
/* SV in column, easier accessing */
double *ptr2;
rhs[out_id] = mxCreateSparse(num_of_feature, model->l, nonzero_element, mxREAL);
ptr2 = mxGetPr(rhs[out_id]);
ir = mxGetIr(rhs[out_id]);
jc = mxGetJc(rhs[out_id]);
jc[0] = ir_index = 0;
for(i = 0;i < model->l; i++)
{
if(model->param.kernel_type == PRECOMPUTED)
{
/* make a (1 x model->l) matrix */
ir[ir_index] = 0;
ptr2[ir_index] = model->SV[i][0].value;
ir_index++;
jc[i+1] = jc[i] + 1;
}
else
{
int x_index = 0;
while (model->SV[i][x_index].index != -1)
{
ir[ir_index] = model->SV[i][x_index].index - 1;
ptr2[ir_index] = model->SV[i][x_index].value;
ir_index++, x_index++;
}
jc[i+1] = jc[i] + x_index;
}
}
/* transpose back to SV in row */
pprhs[0] = rhs[out_id];
if(mexCallMATLAB(1, pplhs, 1, pprhs, "transpose"))
return "cannot transpose SV matrix";
rhs[out_id] = pplhs[0];
#endif
out_id++;
}
#ifdef _USE_CHI_SQUARE
rhs[out_id] = mxCreateString(model->param.feat_parm_text);
#endif
/* Create a struct matrix contains NUM_OF_RETURN_FIELD fields */
return_model = mxCreateStructMatrix(1, 1, NUM_OF_RETURN_FIELD, field_names);
/* Fill struct matrix with input arguments */
for(i = 0; i < NUM_OF_RETURN_FIELD; i++)
mxSetField(return_model,0,field_names[i],mxDuplicateArray(rhs[i]));
/* return */
plhs[0] = return_model;
mxFree(rhs);
return NULL;
}
/* ************************************************************
PROCEDURE mexFunction - Entry for Matlab
************************************************************ */
void mexFunction(int nlhs, mxArray *plhs[],
int nrhs, const mxArray *prhs[])
{
const mxArray *L_FIELD;
mwIndex maxnnz, i,j, nsuper,m,n;
const mwIndex *ljc,*lir,*bjc,*bir;
mwIndex *xjc,*xir, *snode,*xlindx,*lindx, *iwork,*xsuper, *invperm;
bool *cwork;
double *xpr;
const double *permPr, *xsuperPr;
/* ------------------------------------------------------------
Check for proper number of arguments
------------------------------------------------------------ */
mxAssert(nrhs >= NPARIN, "symbfwblk requires more input arguments");
mxAssert(nlhs <= NPAROUT, "symbfwblk produces 1 output argument");
/* ------------------------------------------------------------
Get rhs-input B
------------------------------------------------------------ */
mxAssert(mxIsSparse(B_IN), "B must be sparse");
m = mxGetM(B_IN);
n = mxGetN(B_IN);
bjc = mxGetJc(B_IN);
bir = mxGetIr(B_IN);
/* ------------------------------------------------------------
Disassemble block Cholesky structure L
------------------------------------------------------------ */
mxAssert(mxIsStruct(L_IN), "Parameter `L' should be a structure.");
L_FIELD = mxGetField(L_IN,(mwIndex)0,"perm");
mxAssert( L_FIELD != NULL, "Missing field L.perm."); /* L.perm */
mxAssert(m == mxGetM(L_FIELD) * mxGetN(L_FIELD), "L.perm size mismatches B");
permPr = mxGetPr(L_FIELD);
L_FIELD = mxGetField(L_IN,(mwIndex)0,"L");
mxAssert( L_FIELD!= NULL, "Missing field L.L."); /* L.L */
mxAssert( m == mxGetM(L_FIELD) && m == mxGetN(L_FIELD), "Size L.L mismatch.");
mxAssert(mxIsSparse(L_FIELD), "L.L should be sparse.");
ljc = mxGetJc(L_FIELD);
lir = mxGetIr(L_FIELD);
L_FIELD = mxGetField(L_IN,(mwIndex)0,"xsuper");
mxAssert( L_FIELD!= NULL, "Missing field L.xsuper."); /* L.xsuper */
nsuper = mxGetM(L_FIELD) * mxGetN(L_FIELD) - 1;
mxAssert( nsuper <= m , "Size L.xsuper mismatch.");
xsuperPr = mxGetPr(L_FIELD);
/* ------------------------------------------------------------
Allocate mwIndex-part of sparse output matrix X(m x n)
Heuristically set nnz to nnz(B) + 4*m.
------------------------------------------------------------ */
maxnnz = bjc[n] + 4 * m;
xjc = (mwIndex *) mxCalloc(n + 1, sizeof(mwIndex));
xir = (mwIndex *) mxCalloc(maxnnz, sizeof(mwIndex));
/* ------------------------------------------------------------
Allocate working arrays:
mwIndex invperm(m), snode(m), xsuper(nsuper+1), xlindx(nsuper+1), lindx(nnz(L)),
iwork(nsuper).
char cwork(nsuper).
------------------------------------------------------------ */
invperm = (mwIndex *) mxCalloc(m,sizeof(mwIndex));
snode = (mwIndex *) mxCalloc(m,sizeof(mwIndex));
xsuper = (mwIndex *) mxCalloc(nsuper+1,sizeof(mwIndex));
xlindx = (mwIndex *) mxCalloc(nsuper+1,sizeof(mwIndex));
lindx = (mwIndex *) mxCalloc(ljc[m], sizeof(mwIndex));
iwork = (mwIndex *) mxCalloc(nsuper, sizeof(mwIndex));
cwork = (bool *) mxCalloc(nsuper, sizeof(bool));
/* ------------------------------------------------------------
Convert PERM, XSUPER to integer and C-Style
------------------------------------------------------------ */
for(i = 0; i < m; i++){
j = (mwIndex) permPr[i];
mxAssert(j>0,"");
invperm[--j] = i; /* so that invperm[perm[i]] = i */
}
for(i = 0; i <= nsuper; i++){
j = (mwIndex) xsuperPr[i];
mxAssert(j>0,"");
xsuper[i] = --j;
}
/* ------------------------------------------------------------
Create "snode" from xsuper, and get compact subscript (xlindx,lindx)
from (ljc,lir,xsuper), i.e. nz-pattern per supernode.
------------------------------------------------------------ */
snodeCompress(xlindx,lindx,snode, ljc,lir,xsuper,nsuper);
/* ------------------------------------------------------------
Compute nz structure after forward solve
------------------------------------------------------------ */
symbfwmat(xjc, &xir, &maxnnz, bjc, bir, invperm, snode, xsuper,
xlindx, lindx,
nsuper, m, n, iwork, cwork);
/* ------------------------------------------------------------
Create output matrix x
------------------------------------------------------------ */
X_OUT = mxCreateSparse(m,n, (mwSize)1,mxREAL);
mxFree(mxGetJc(X_OUT)); /* jc */
mxFree(mxGetIr(X_OUT)); /* ir */
mxFree(mxGetPr(X_OUT)); /* pr */
xpr = (double *) mxCalloc(maxnnz,sizeof(double));
mxSetJc(X_OUT, xjc);
mxSetIr(X_OUT, xir);
mxSetPr(X_OUT, xpr);
mxSetNzmax(X_OUT, maxnnz);
for(i = 0; i < maxnnz; i++)
xpr[i] = 1.0;
/* ------------------------------------------------------------
Release working arrays.
------------------------------------------------------------ */
mxFree(cwork);
mxFree(iwork);
mxFree(lindx);
mxFree(xlindx);
mxFree(xsuper);
mxFree(snode);
mxFree(invperm);
}
void mexFunction(int nlhs, mxArray *plhs[],
int nrhs, const mxArray *prhs[] )
{ mxArray *blk_cell_pr;
double *A, *B, *AI, *BI, *blksize;
mwIndex *irA, *jcA, *irB, *jcB;
int *cumblksize, *blknnz;
int mblk, isspA, isspB, iscmpA;
mwIndex subs[2];
mwSize nsubs=2;
int m, n, n2, nsub, k, index, numblk, NZmax, type;
double r2;
/* CHECK FOR PROPER NUMBER OF ARGUMENTS */
if (nrhs < 2){
mexErrMsgTxt("mexsvec: requires at least 2 input arguments."); }
if (nlhs > 1){
mexErrMsgTxt("mexsvec: requires 1 output argument."); }
/* CHECK THE DIMENSIONS */
mblk = mxGetM(prhs[0]);
if (mblk > 1) {
mexErrMsgTxt("mexsvec: blk can have only 1 row."); }
m = mxGetM(prhs[1]);
n = mxGetN(prhs[1]);
if (m != n) {
mexErrMsgTxt("mexsvec: matrix must be square."); }
subs[0] = 0; subs[1] = 1;
index = mxCalcSingleSubscript(prhs[0],nsubs,subs);
blk_cell_pr = mxGetCell(prhs[0],index);
numblk = mxGetN(blk_cell_pr);
blksize = mxGetPr(blk_cell_pr);
if (numblk == 1) {
n2 = n*(n+1)/2;
} else {
cumblksize = mxCalloc(numblk+1,sizeof(int));
blknnz = mxCalloc(numblk+1,sizeof(int));
cumblksize[0] = 0; blknnz[0] = 0;
n = 0; n2 = 0;
for (k=0; k<numblk; ++k) {
nsub = (int) blksize[k];
n += nsub;
n2 += nsub*(nsub+1)/2;
cumblksize[k+1] = n;
blknnz[k+1] = n2; }
}
/***** assign pointers *****/
A = mxGetPr(prhs[1]);
isspA = mxIsSparse(prhs[1]);
iscmpA = mxIsComplex(prhs[1]);
if (isspA) { irA = mxGetIr(prhs[1]);
jcA = mxGetJc(prhs[1]);
NZmax = mxGetNzmax(prhs[1]);
} else {
NZmax = n2;
}
if (iscmpA) { AI = mxGetPi(prhs[1]); }
if ((numblk > 1) & (!isspA)) {
mexErrMsgTxt("mexsvec: matrix must be sparse for numblk > 1"); }
if (nrhs > 2) {
if (mxGetM(prhs[2])>1) { isspB = (int)*mxGetPr(prhs[2]); }
else if (NZmax < n2/2) { isspB = 1; }
else { isspB = 0; }
} else {
if (NZmax < n2/2) { isspB = 1; }
else { isspB = 0; }
}
if (nrhs > 3) { type = (int)*mxGetPr(prhs[3]); }
else { type = 0; }
/***** create return argument *****/
if (isspB) {
if (iscmpA) {
plhs[0] = mxCreateSparse(n2,1,NZmax,mxCOMPLEX);
} else {
plhs[0] = mxCreateSparse(n2,1,NZmax,mxREAL);
}
B = mxGetPr(plhs[0]);
irB = mxGetIr(plhs[0]);
jcB = mxGetJc(plhs[0]);
jcB[0] = 0;
} else {
if (iscmpA) {
plhs[0] = mxCreateDoubleMatrix(n2,1,mxCOMPLEX);
} else {
plhs[0] = mxCreateDoubleMatrix(n2,1,mxREAL);
}
B = mxGetPr(plhs[0]);
}
if (iscmpA) { BI = mxGetPi(plhs[0]); }
/***** Do the computations in a subroutine *****/
r2 = sqrt(2);
if (type == 0) {
if (iscmpA) {
if (numblk == 1) {
svec1cmp(n,r2,A,irA,jcA,isspA,B,irB,jcB,isspB,AI,BI); }
else {
svec2cmp(n,numblk,cumblksize,blknnz,r2,A,irA,jcA,isspA,B,irB,jcB,isspB,AI,BI);
}
} else {
if (numblk == 1) {
svec1(n,r2,A,irA,jcA,isspA,B,irB,jcB,isspB); }
else {
svec2(n,numblk,cumblksize,blknnz,r2,A,irA,jcA,isspA,B,irB,jcB,isspB);
}
}
} else {
if (iscmpA) {
if (numblk == 1) {
svec3cmp(n,r2,A,irA,jcA,isspA,B,irB,jcB,isspB,AI,BI); }
else {
svec4cmp(n,numblk,cumblksize,blknnz,r2,A,irA,jcA,isspA,B,irB,jcB,isspB,AI,BI);
}
} else {
if (numblk == 1) {
svec3(n,r2,A,irA,jcA,isspA,B,irB,jcB,isspB); }
else {
svec4(n,numblk,cumblksize,blknnz,r2,A,irA,jcA,isspA,B,irB,jcB,isspB);
}
}
}
return;
}
