// ================================
// solve linear equations C = A\B
// ================================
void solve(double* C, double* A, double* B,
const int rA, const int cA, const int rB, const int cB,
const char *mod, double *W, const int LW, ptrdiff_t *S) {
#ifdef USE_BLAS
int i, j, rank;
char uplo = 'U', side = 'L', trans = 'N', unit = 'N';
double one = 1.0, rcond = 0.000000001;
ptrdiff_t rA0 = rA, cA0 = cA, cB0 = cB, Lwork=LW, info;
ptrdiff_t rC0 = (rA>cA) ? rA : cA;
//ptrdiff_t ptr_S = S;
switch (mod[0]){
case 'L':
case 'U':
uplo = mod[0];
dtrsm(&side, &uplo, &trans, &unit, &rC0, &cB0, &one, A, &rA0, C, &rC0);
break;
case 'P':
dposv(&uplo, &rA0, &cB0, W, &rA0, C, &rA0, &info);// A has already been copied into W
break;
default:
if (rA == cA) {
//dgesv(&rA0, &cB0, W, &rA0, S, C, &rA0, &info);// A has already been copied into W
dgesv(&rA0, &cB0, W, &rA0, (ptrdiff_t*)S, C, &rA0, &info);// A has already been copied into W
}
else{
for( i=0; i<cB; i++ )
for( j=0; j<rB; j++ )
C[i*rC0+j] = B[i*rB+j];
//dgelsy(&rA0, &cA0, &cB0, A, &rA0, C, &rC0, S, &rcond, &rank, W, &Lwork, &info);
dgelsy(&rA0, &cA0, &cB0, A, &rA0, C, &rC0,
(ptrdiff_t*)S, &rcond, (ptrdiff_t*)&rank, W, &Lwork, &info);
}
}
#endif
}
void encode(const double* dict_words, const double* im_words,
const double* nn_ind, ptrdiff_t dim, ptrdiff_t dict_size,
ptrdiff_t im_size, ptrdiff_t nn_size, double beta,
double* encoding_sr, mwIndex* encoding_ir, mwIndex* encoding_jc,
double* hist, bool output_full_mat){
char upper_triangle = 'U';
ptrdiff_t INFO;
ptrdiff_t int_one = 1;
double sum;
double* z = new double[dim*nn_size];
double* C = new double[nn_size*nn_size];
double* b = new double[nn_size];
ptrdiff_t* IPIV = new ptrdiff_t[nn_size];
if (output_full_mat) {
// first column with non-zero value in encoding has index 0
encoding_jc[0] = 0;
}
for(mwIndex i=0; i<im_size; i++){
mwIndex tmp_ind;
for(mwIndex n=0;n<nn_size;n++){
for(mwIndex m=0;m<dim;m++){
tmp_ind = (mwIndex)nn_ind[i*nn_size+n];
z[n*dim+m] = dict_words[tmp_ind*dim+m]-im_words[i*dim+m];
}
}
mult_help(z,dim,nn_size,C);
// dgemm(&transpose,&non_trans,&nn_size,&nn_size,&dim,&one,z,&nn_size,z,&dim,&zero,C,&nn_size);
sum = 0;
for(mwIndex m=0;m<nn_size;m++) sum += C[m*nn_size+m];
sum = sum*beta;
for(mwIndex m=0;m<nn_size;m++) C[m*nn_size+m] += sum;
for(mwIndex m=0;m<nn_size;m++) b[m] = 1;
dposv(&upper_triangle,&nn_size,&int_one,C,&nn_size,b,&nn_size,&INFO);
// dgesv(&nn_size,&int_one,C,&nn_size,IPIV,b,&nn_size,&INFO);
sum = 0;
for(mwIndex m=0;m<nn_size;m++) sum += b[m];
for(mwIndex m=0;m<nn_size;m++) b[m] /= sum;
for(mwIndex m=0;m<nn_size;m++){
tmp_ind = (mwIndex)nn_ind[i*nn_size+m];
if (output_full_mat) {
encoding_sr[encoding_jc[i]+m] = b[m];
encoding_ir[encoding_jc[i]+m] = tmp_ind;
} else {
if(hist[tmp_ind]<b[m]) hist[tmp_ind] = b[m];
}
}
if (output_full_mat) {
encoding_jc[i+1] = encoding_jc[i] + nn_size;
}
}
delete [] z;
delete [] C;
delete [] b;
delete [] IPIV;
return;
}
