/* Compute (and allocate) migration price matrix for optimization */
static void
price_routes(PRICEMAT *pm, const RBFNODE *from_rbf, const RBFNODE *to_rbf)
{
FVECT *vto = (FVECT *)malloc(sizeof(FVECT) * to_rbf->nrbf);
int i, j;
pm->nrows = from_rbf->nrbf;
pm->ncols = to_rbf->nrbf;
pm->price = (float *)malloc(sizeof(float) * pm->nrows*pm->ncols);
pm->sord = (short *)malloc(sizeof(short) * pm->nrows*pm->ncols);
if ((pm->price == NULL) | (pm->sord == NULL) | (vto == NULL)) {
fprintf(stderr, "%s: Out of memory in migration_costs()\n",
progname);
exit(1);
}
for (j = to_rbf->nrbf; j--; ) /* save repetitive ops. */
ovec_from_pos(vto[j], to_rbf->rbfa[j].gx, to_rbf->rbfa[j].gy);
for (i = from_rbf->nrbf; i--; ) {
const double from_ang = R2ANG(from_rbf->rbfa[i].crad);
FVECT vfrom;
ovec_from_pos(vfrom, from_rbf->rbfa[i].gx, from_rbf->rbfa[i].gy);
for (j = to_rbf->nrbf; j--; ) {
double dprod = DOT(vfrom, vto[j]);
pricerow(pm,i)[j] = ((dprod >= 1.) ? .0 : acos(dprod)) +
fabs(R2ANG(to_rbf->rbfa[j].crad) - from_ang);
psortrow(pm,i)[j] = j;
}
qsort_r(psortrow(pm,i), pm->ncols, sizeof(short), pm, &msrt_cmp);
}
free(vto);
}
/* Cull points for more uniform distribution, leave all nval 0 or 1 */
static void
cull_values(void)
{
FVECT ovec0, ovec1;
double maxang, maxang2;
int i, j, ii, jj, r;
/* simple greedy algorithm */
for (i = 0; i < GRIDRES; i++)
for (j = 0; j < GRIDRES; j++) {
if (!dsf_grid[i][j].nval)
continue;
if (!dsf_grid[i][j].crad)
continue; /* shouldn't happen */
ovec_from_pos(ovec0, i, j);
maxang = 2.*R2ANG(dsf_grid[i][j].crad);
if (maxang > ovec0[2]) /* clamp near horizon */
maxang = ovec0[2];
r = maxang*(2.*GRIDRES/M_PI) + 1;
maxang2 = maxang*maxang;
for (ii = i-r; ii <= i+r; ii++) {
if (ii < 0) continue;
if (ii >= GRIDRES) break;
for (jj = j-r; jj <= j+r; jj++) {
if (jj < 0) continue;
if (jj >= GRIDRES) break;
if (!dsf_grid[ii][jj].nval)
continue;
if ((ii == i) & (jj == j))
continue; /* don't get self-absorbed */
ovec_from_pos(ovec1, ii, jj);
if (2. - 2.*DOT(ovec0,ovec1) >= maxang2)
continue;
/* absorb sum */
dsf_grid[i][j].vsum += dsf_grid[ii][jj].vsum;
dsf_grid[i][j].nval += dsf_grid[ii][jj].nval;
/* keep value, though */
dsf_grid[ii][jj].vsum /= (float)dsf_grid[ii][jj].nval;
dsf_grid[ii][jj].nval = 0;
}
}
}
/* final averaging pass */
for (i = 0; i < GRIDRES; i++)
for (j = 0; j < GRIDRES; j++)
if (dsf_grid[i][j].nval > 1) {
dsf_grid[i][j].vsum /= (float)dsf_grid[i][j].nval;
dsf_grid[i][j].nval = 1;
}
}
/* Evaluate RBF for DSF at the given normalized outgoing direction */
double
eval_rbfrep(const RBFNODE *rp, const FVECT outvec)
{
double minval = bsdf_min*output_orient*outvec[2];
double res = 0;
const RBFVAL *rbfp;
FVECT odir;
double sig2;
int n;
/* use minimum if no information avail. */
if (rp == NULL) {
if (outvec[2] > 0 ^ output_orient > 0)
return(.0);
return(minval);
}
rbfp = rp->rbfa;
for (n = rp->nrbf; n--; rbfp++) {
ovec_from_pos(odir, rbfp->gx, rbfp->gy);
sig2 = R2ANG(rbfp->crad);
sig2 = (DOT(odir,outvec) - 1.) / (sig2*sig2);
if (sig2 > -19.)
res += rbfp->peak * exp(sig2);
}
if (res < minval) /* never return less than minval */
return(minval);
return(res);
}
/* Compute distance between two RBF lobes */
double
lobe_distance(RBFVAL *rbf1, RBFVAL *rbf2)
{
FVECT vfrom, vto;
double d, res;
/* quadratic cost function */
ovec_from_pos(vfrom, rbf1->gx, rbf1->gy);
ovec_from_pos(vto, rbf2->gx, rbf2->gy);
d = Acos(DOT(vfrom, vto));
res = d*d;
d = R2ANG(rbf2->crad) - R2ANG(rbf1->crad);
res += d*d;
/* neighborhood difference */
res += NEIGH_FACT2 * neighborhood_dist2( rbf1->gx, rbf1->gy,
rbf2->gx, rbf2->gy );
return(res);
}
/* Compute normalized distribution scattering functions for comparison */
static void
compute_nDSFs(const RBFNODE *rbf0, const RBFNODE *rbf1)
{
const double nf0 = (GRIDRES*GRIDRES) / rbf0->vtotal;
const double nf1 = (GRIDRES*GRIDRES) / rbf1->vtotal;
int x, y;
FVECT dv;
for (x = GRIDRES; x--; )
for (y = GRIDRES; y--; ) {
ovec_from_pos(dv, x, y); /* cube root (brightness) */
dsf_grid[x][y].val[0] = pow(nf0*eval_rbfrep(rbf0, dv), .3333);
dsf_grid[x][y].val[1] = pow(nf1*eval_rbfrep(rbf1, dv), .3333);
}
}
/* Rotate RBF to correspond to given incident vector */
void
rotate_rbf(RBFNODE *rbf, const FVECT invec)
{
static const FVECT vnorm = {.0, .0, 1.};
const double phi = atan2(invec[1],invec[0]) -
atan2(rbf->invec[1],rbf->invec[0]);
FVECT outvec;
int pos[2];
int n;
for (n = ((-.01 > phi) | (phi > .01))*rbf->nrbf; n-- > 0; ) {
ovec_from_pos(outvec, rbf->rbfa[n].gx, rbf->rbfa[n].gy);
spinvector(outvec, outvec, vnorm, phi);
pos_from_vec(pos, outvec);
rbf->rbfa[n].gx = pos[0];
rbf->rbfa[n].gy = pos[1];
}
VCOPY(rbf->invec, invec);
}
/* (distance to furthest empty bin for which non-empty bin is the closest) */
static void
compute_radii(void)
{
unsigned int fill_grid[GRIDRES][GRIDRES];
unsigned short fill_cnt[GRIDRES][GRIDRES];
FVECT ovec0, ovec1;
double ang2, lastang2;
int r, i, j, jn, ii, jj, inear, jnear;
r = GRIDRES/2; /* proceed in zig-zag */
for (i = 0; i < GRIDRES; i++)
for (jn = 0; jn < GRIDRES; jn++) {
j = (i&1) ? jn : GRIDRES-1-jn;
if (dsf_grid[i][j].nval) /* find empty grid pos. */
continue;
ovec_from_pos(ovec0, i, j);
inear = jnear = -1; /* find nearest non-empty */
lastang2 = M_PI*M_PI;
for (ii = i-r; ii <= i+r; ii++) {
if (ii < 0) continue;
if (ii >= GRIDRES) break;
for (jj = j-r; jj <= j+r; jj++) {
if (jj < 0) continue;
if (jj >= GRIDRES) break;
if (!dsf_grid[ii][jj].nval)
continue;
ovec_from_pos(ovec1, ii, jj);
ang2 = 2. - 2.*DOT(ovec0,ovec1);
if (ang2 >= lastang2)
continue;
lastang2 = ang2;
inear = ii; jnear = jj;
}
}
if (inear < 0) {
fprintf(stderr,
"%s: Could not find non-empty neighbor!\n",
progname);
exit(1);
}
ang2 = sqrt(lastang2);
r = ANG2R(ang2); /* record if > previous */
if (r > dsf_grid[inear][jnear].crad)
dsf_grid[inear][jnear].crad = r;
/* next search radius */
r = ang2*(2.*GRIDRES/M_PI) + 3;
}
/* blur radii over hemisphere */
memset(fill_grid, 0, sizeof(fill_grid));
memset(fill_cnt, 0, sizeof(fill_cnt));
for (i = 0; i < GRIDRES; i++)
for (j = 0; j < GRIDRES; j++) {
if (!dsf_grid[i][j].crad)
continue; /* missing distance */
r = R2ANG(dsf_grid[i][j].crad)*(2.*RSCA*GRIDRES/M_PI);
for (ii = i-r; ii <= i+r; ii++) {
if (ii < 0) continue;
if (ii >= GRIDRES) break;
for (jj = j-r; jj <= j+r; jj++) {
if (jj < 0) continue;
if (jj >= GRIDRES) break;
if ((ii-i)*(ii-i) + (jj-j)*(jj-j) > r*r)
continue;
fill_grid[ii][jj] += dsf_grid[i][j].crad;
fill_cnt[ii][jj]++;
}
}
}
/* copy back blurred radii */
for (i = 0; i < GRIDRES; i++)
for (j = 0; j < GRIDRES; j++)
if (fill_cnt[i][j])
dsf_grid[i][j].crad = fill_grid[i][j]/fill_cnt[i][j];
}
/* Count up filled nodes and build RBF representation from current grid */
RBFNODE *
make_rbfrep(void)
{
int niter = 16;
double lastVar, thisVar = 100.;
int nn;
RBFNODE *newnode;
RBFVAL *itera;
int i, j;
/* compute RBF radii */
compute_radii();
/* coagulate lobes */
cull_values();
nn = 0; /* count selected bins */
for (i = 0; i < GRIDRES; i++)
for (j = 0; j < GRIDRES; j++)
nn += dsf_grid[i][j].nval;
/* compute minimum BSDF */
comp_bsdf_min();
/* allocate RBF array */
newnode = (RBFNODE *)malloc(sizeof(RBFNODE) + sizeof(RBFVAL)*(nn-1));
if (newnode == NULL)
goto memerr;
newnode->ord = -1;
newnode->next = NULL;
newnode->ejl = NULL;
newnode->invec[2] = sin((M_PI/180.)*theta_in_deg);
newnode->invec[0] = cos((M_PI/180.)*phi_in_deg)*newnode->invec[2];
newnode->invec[1] = sin((M_PI/180.)*phi_in_deg)*newnode->invec[2];
newnode->invec[2] = input_orient*sqrt(1. - newnode->invec[2]*newnode->invec[2]);
newnode->vtotal = 0;
newnode->nrbf = nn;
nn = 0; /* fill RBF array */
for (i = 0; i < GRIDRES; i++)
for (j = 0; j < GRIDRES; j++)
if (dsf_grid[i][j].nval) {
newnode->rbfa[nn].peak = dsf_grid[i][j].vsum;
newnode->rbfa[nn].crad = RSCA*dsf_grid[i][j].crad + .5;
newnode->rbfa[nn].gx = i;
newnode->rbfa[nn].gy = j;
++nn;
}
/* iterate to improve interpolation accuracy */
itera = (RBFVAL *)malloc(sizeof(RBFVAL)*newnode->nrbf);
if (itera == NULL)
goto memerr;
memcpy(itera, newnode->rbfa, sizeof(RBFVAL)*newnode->nrbf);
do {
double dsum = 0, dsum2 = 0;
nn = 0;
for (i = 0; i < GRIDRES; i++)
for (j = 0; j < GRIDRES; j++)
if (dsf_grid[i][j].nval) {
FVECT odir;
double corr;
ovec_from_pos(odir, i, j);
itera[nn++].peak *= corr =
dsf_grid[i][j].vsum /
eval_rbfrep(newnode, odir);
dsum += 1. - corr;
dsum2 += (1.-corr)*(1.-corr);
}
memcpy(newnode->rbfa, itera, sizeof(RBFVAL)*newnode->nrbf);
lastVar = thisVar;
thisVar = dsum2/(double)nn;
#ifdef DEBUG
fprintf(stderr, "Avg., RMS error: %.1f%% %.1f%%\n",
100.*dsum/(double)nn,
100.*sqrt(thisVar));
#endif
} while (--niter > 0 && lastVar-thisVar > 0.02*lastVar);
free(itera);
nn = 0; /* compute sum for normalization */
while (nn < newnode->nrbf)
newnode->vtotal += rbf_volume(&newnode->rbfa[nn++]);
#ifdef DEBUG
fprintf(stderr, "Integrated DSF at (%.1f,%.1f) deg. is %.2f\n",
get_theta180(newnode->invec), get_phi360(newnode->invec),
newnode->vtotal);
#endif
insert_dsf(newnode);
return(newnode);
memerr:
fprintf(stderr, "%s: Out of memory in make_rbfrep()\n", progname);
exit(1);
}
