void mexFunction(int nlhs,mxArray *plhs[],int nrhs,const mxArray *prhs[])
{
DOUBLE *lpp, *lpf, *sig;
unsigned int m,n,nn,mm;
int nr,nc,J,kk,lenfil;
/* Check for proper number of arguments */
if (nrhs != 2) {
mexErrMsgTxt("UpDyadLo requires two input arguments.");
} else if (nlhs != 1) {
mexErrMsgTxt("UpDyadLo requires one output argument.");
}
/* Check the dimensions of signal. signal can be n X 1 or 1 X n. */
m = mxGetM(Sig_IN);
n = mxGetN(Sig_IN);
if(m == 1){
nr = (int) n;
nc = 1;
nn = 2*n;
mm = 1;
} else {
nr = (int) m;
nc = (int) n;
nn = 1;
mm = 2*m;
}
J = 0;
for( kk = 1; kk < nr; kk *= 2 )
J ++;
if( kk != nr){
mexErrMsgTxt("UpDyadLo requires dyadic length");
}
/* Create a matrix for the return argument */
LP_OUT = mxCreateDoubleMatrix(mm, nn, mxREAL);
/* Assign pointers to the various parameters */
lpp = mxGetPr(LP_OUT);
sig = mxGetPr(Sig_IN);
lpf = mxGetPr(LPF_IN);
lenfil = (int) (mxGetM(LPF_IN) * mxGetN(LPF_IN)); /* should check this */
/* Do the actual computations in a subroutine */
uplo(sig,nr,lpf,lenfil,lpp);
}
/**
* @brief Perform inverse DWT (periodic boundaries) on 3D data.
*
* @param wc Wavelet presentation of 3D data.
* @param img Reconstructed signal.
*/
void
idpwt3 (const Matrix <T> & wc, Matrix <T> & img)
{
// assign dwt to result image
img = wc;
# pragma omp parallel default (shared) num_threads (_num_threads)
{
T * wcplo, * wcphi, * templo, * temphi, * temptop, * tmp;
size_t stride;
int sl1 = _sl1_scale,
sl2 = _sl2_scale,
sl3 = _sl3_scale;
const int t_num = omp_get_thread_num ();
// loop over levels of backwards DWT
for (int j = _min_level; j < _max_level; j++)
{
// update stride
stride = 6 * sl3 * t_num;
tmp = & _temp [stride];
templo = & _temp [2 * sl3 + stride];
temphi = & _temp [3 * sl3 + stride];
temptop = & _temp [4 * sl3 + stride];
# pragma omp for schedule (OMP_SCHEDULE)
// loop over lines along third dimension ('third') of result image
for (int c1_loc = 0; c1_loc < 2 * sl1 * 2 * sl2; c1_loc++)
{
int c1_glob = (c1_loc % (2 * sl1)) + (c1_loc / (2 * sl1)) * _sl1;
// copy lowpass part of current line to temporary memory
unpackdouble (& img [c1_glob], sl3, _ld12, 0, templo);
// copy highpass part of current line to temporary memory
unpackdouble (& img [c1_glob + sl3 * _ld12], sl3, _ld12, 0, temphi);
// perform lowpass reconstruction
uplo (templo, sl3, tmp);
// perform highpass reconstruction
uphi (temphi, sl3, temptop);
// fusion of reconstruction parts
adddouble (tmp, temptop, sl3 * 2, tmp);
// write back reconstructed line
packdouble (tmp, sl3 * 2, _ld12, 0, & img [c1_glob]);
} // loop over lines along third dimension of result image
// update stride
stride = 6 * sl2 * t_num;
tmp = & _temp [stride];
templo = & _temp [2 * sl2 + stride];
temphi = & _temp [3 * sl2 + stride];
temptop = & _temp [4 * sl2 + stride];
# pragma omp for schedule (OMP_SCHEDULE)
// loop over lines along second dimension ('rows') of result image
for (int c1_loc = 0; c1_loc < 2 * sl1 * 2 * sl3; c1_loc++)
{
int c1_glob = (c1_loc / (2 * sl1)) * _sl1 * _sl2;
// copy lowpass part of current line to temporary memory
unpackdouble (& img [c1_glob], sl2, _sl1, c1_loc % (2 * sl1), templo);
// copy highpass part of current line to temporary memory
unpackdouble (& img [c1_glob + sl2 * _sl1], sl2, _sl1, c1_loc % (2 * sl1), temphi);
// perform lowpass reconstruction
uplo (templo, sl2, tmp);
// perform highpass reconstruction
uphi (temphi, sl2, temptop);
// fusion of reconstruction parts
adddouble (tmp, temptop, sl2 * 2, tmp);
// write back reconstructed line
packdouble (tmp, sl2 * 2, _sl1, c1_loc % (2 * sl1), & img [c1_glob]);
} // loop over lines along second dimension of result image
// update stride
stride = 5 * sl1 * t_num;
tmp = & _temp [stride];
templo = & _temp [ sl1 + stride];
temphi = & _temp [3 * sl1 + stride];
# pragma omp for schedule (OMP_SCHEDULE)
// loop over lines along first dimension ('columns') of result image
for (int c2_loc = 0; c2_loc < 2 * sl2 * 2 * sl3; c2_loc++)
{
int c2_glob = (c2_loc / (2 * sl2)) * _sl2 * _sl1 + (c2_loc % (2 * sl2)) * _sl1;
// assign address of current line's lowpass part
wcplo = & img [c2_glob];
// assign address of current line's highpass part
wcphi = & img [c2_glob + sl1];
// copy lowpass part to temporary memory
copydouble (wcplo, tmp, sl1);
// perform lowpass reconstruction
uplo (wcplo, sl1, templo);
// perform highpass reconstruction
uphi (wcphi, sl1, temphi);
// combine reconstructed parts and write back to current line
adddouble (templo, temphi, sl1 * 2, wcplo);
} // loop over lines along first dimension ('columns') of result image
// update current row / column size
sl2 *= 2;
sl1 *= 2;
sl3 *= 2;
} // loop over levels of backwards DWT
} // omp parallel
}
Matrix operator * (const Matrix& A, const Matrix& B)
{
if (A.Clo() != B.Rlo() || A.Chi() != B.Rhi())
Matpack.Error("Matrix operator * (const Matrix&, const Matrix&): "
"non conformant arguments\n");
// allocate return matrix
Matrix C(A.Rlo(),A.Rhi(),B.Clo(),B.Chi());
//------------------------------------------------------------------------//
// the BLAS version
//------------------------------------------------------------------------//
#if defined ( _MATPACK_USE_BLAS_ )
if ( LT(B) ) { // full matrix * lower triangle
#ifdef DEBUG
cout << "GM*LT\n";
#endif
checksquare(B);
// copy A to C to protect from overwriting
copyvec(C.Store(),A.Store(),A.Elements());
charT side('L'), uplo('U'), transc('N'), diag('N');
intT m(C.Cols()), n(C.Rows()),
ldb(B.Cols()), ldc(C.Cols());
doubleT alpha(1.0);
F77NAME(dtrmm)(&side,&uplo,&transc,&diag,&m,&n,
&alpha,B.Store(),&ldb, C.Store(),&ldc);
} else if ( UT(B) ) { // full matrix * upper triangle
#ifdef DEBUG
cout << "GM*UT\n";
#endif
checksquare(B);
// copy A to C to protect from overwriting
copyvec(C.Store(),A.Store(),A.Elements());
charT side('L'), uplo('L'), transc('N'), diag('N');
intT m(C.Cols()), n(C.Rows()),
ldb(B.Cols()), ldc(C.Cols());
doubleT alpha(1.0);
F77NAME(dtrmm)(&side,&uplo,&transc,&diag,&m,&n,
&alpha,B.Store(),&ldb, C.Store(),&ldc);
} else if ( LT(A) ) { // lower triangle * full matrix
#ifdef DEBUG
cout << "LT*GM\n";
#endif
checksquare(A);
// copy B to C to protect from overwriting
copyvec(C.Store(),B.Store(),B.Elements());
charT side('R'), uplo('U'), transc('N'), diag('N');
intT m(C.Cols()), n(C.Rows()),
ldb(A.Cols()), ldc(C.Cols());
doubleT alpha(1.0);
F77NAME(dtrmm)(&side,&uplo,&transc,&diag,&m,&n,
&alpha,A.Store(),&ldb, C.Store(),&ldc);
} else if ( UT(A) ) { // upper triangle * full matrix
#ifdef DEBUG
cout << "UT*GM\n";
#endif
checksquare(A);
// copy A to C to protect from overwriting
copyvec(C.Store(),B.Store(),B.Elements());
charT side('R'), uplo('L'), transc('N'), diag('N');
intT m(C.Cols()), n(C.Rows()),
ldb(A.Cols()), ldc(C.Cols());
doubleT alpha(1.0);
F77NAME(dtrmm)(&side,&uplo,&transc,&diag,&m,&n,
&alpha,A.Store(),&ldb, C.Store(),&ldc);
} else /* GM(A) and GM(B) */ { // GM*GM: full matrix * full matrix
#ifdef DEBUG
cout << "GM*GM\n";
#endif
charT t('N');
intT m(B.Cols()), n(A.Rows()), k(B.Rows()),
lda(A.Cols()), ldb(B.Cols()), ldc(C.Cols());
doubleT alpha(1.0), beta(0.0);
F77NAME(dgemm)(&t,&t, &m,&n,&k,
&alpha,B.Store(),&ldb, A.Store(),&lda,
&beta,C.Store(),&ldc);
}
//------------------------------------------------------------------------//
// the non-BLAS version
//------------------------------------------------------------------------//
#else
int cl = A.cl, ch = A.ch,
arl = A.rl, arh = A.rh,
bcl = B.cl, bch = B.ch;
// avoid call to index operator that optimizes very badely
double **a = A.M, **b = B.M, **c = C.M;
for (int i = arl; i <= arh; i++) {
for (int j = bcl; j <= bch; j++) c[i][j] = 0.0;
for (int l = cl; l <= ch; l++) {
if ( a[i][l] != 0.0 ) {
double temp = a[i][l];
for (int j = bcl; j <= bch; j++)
c[i][j] += temp * b[l][j];
}
}
}
#endif
return C.Value();
}
