| 27 |
|
int nrows, ncols; /* array size (matches migration) */ |
| 28 |
|
float *price; /* migration prices */ |
| 29 |
|
short *sord; /* sort for each row, low to high */ |
| 30 |
+ |
float *prow; /* current price row */ |
| 31 |
|
} PRICEMAT; /* sorted pricing matrix */ |
| 32 |
|
|
| 33 |
|
#define pricerow(p,i) ((p)->price + (i)*(p)->ncols) |
| 140 |
|
msrt_cmp(void *b, const void *p1, const void *p2) |
| 141 |
|
{ |
| 142 |
|
PRICEMAT *pm = (PRICEMAT *)b; |
| 143 |
< |
int ri = ((const short *)p1 - pm->sord) / pm->ncols; |
| 144 |
< |
float c1 = pricerow(pm,ri)[*(const short *)p1]; |
| 144 |
< |
float c2 = pricerow(pm,ri)[*(const short *)p2]; |
| 143 |
> |
float c1 = pm->prow[*(const short *)p1]; |
| 144 |
> |
float c2 = pm->prow[*(const short *)p2]; |
| 145 |
|
|
| 146 |
|
if (c1 > c2) return(1); |
| 147 |
|
if (c1 < c2) return(-1); |
| 171 |
|
for (i = from_rbf->nrbf; i--; ) { |
| 172 |
|
const double from_ang = R2ANG(from_rbf->rbfa[i].crad); |
| 173 |
|
FVECT vfrom; |
| 174 |
+ |
short *srow; |
| 175 |
|
ovec_from_pos(vfrom, from_rbf->rbfa[i].gx, from_rbf->rbfa[i].gy); |
| 176 |
+ |
pm->prow = pricerow(pm,i); |
| 177 |
+ |
srow = psortrow(pm,i); |
| 178 |
|
for (j = to_rbf->nrbf; j--; ) { |
| 179 |
< |
double dprod = DOT(vfrom, vto[j]); |
| 180 |
< |
pricerow(pm,i)[j] = ((dprod >= 1.) ? .0 : acos(dprod)) + |
| 181 |
< |
fabs(R2ANG(to_rbf->rbfa[j].crad) - from_ang); |
| 182 |
< |
psortrow(pm,i)[j] = j; |
| 179 |
> |
double d; /* quadratic cost function */ |
| 180 |
> |
d = DOT(vfrom, vto[j]); |
| 181 |
> |
d = (d >= 1.) ? .0 : acos(d); |
| 182 |
> |
pm->prow[j] = d*d; |
| 183 |
> |
d = R2ANG(to_rbf->rbfa[j].crad) - from_ang; |
| 184 |
> |
pm->prow[j] += d*d; |
| 185 |
> |
srow[j] = j; |
| 186 |
|
} |
| 187 |
< |
qsort_r(psortrow(pm,i), pm->ncols, sizeof(short), pm, &msrt_cmp); |
| 187 |
> |
qsort_r(srow, pm->ncols, sizeof(short), pm, &msrt_cmp); |
| 188 |
|
} |
| 189 |
|
free(vto); |
| 190 |
|
} |
| 201 |
|
static double |
| 202 |
|
min_cost(double amt2move, const double *avail, const PRICEMAT *pm, int s) |
| 203 |
|
{ |
| 204 |
+ |
const short *srow = psortrow(pm,s); |
| 205 |
+ |
const float *prow = pricerow(pm,s); |
| 206 |
|
double total_cost = 0; |
| 207 |
|
int j; |
| 200 |
– |
|
| 201 |
– |
if (amt2move <= FTINY) /* pre-emptive check */ |
| 202 |
– |
return(.0); |
| 208 |
|
/* move cheapest first */ |
| 209 |
< |
for (j = 0; j < pm->ncols && amt2move > FTINY; j++) { |
| 210 |
< |
int d = psortrow(pm,s)[j]; |
| 209 |
> |
for (j = 0; (j < pm->ncols) & (amt2move > FTINY); j++) { |
| 210 |
> |
int d = srow[j]; |
| 211 |
|
double amt = (amt2move < avail[d]) ? amt2move : avail[d]; |
| 212 |
|
|
| 213 |
< |
total_cost += amt * pricerow(pm,s)[d]; |
| 213 |
> |
total_cost += amt * prow[d]; |
| 214 |
|
amt2move -= amt; |
| 215 |
|
} |
| 216 |
|
return(total_cost); |
| 217 |
|
} |
| 218 |
|
|
| 219 |
< |
/* Take a step in migration by choosing optimal bucket to transfer */ |
| 219 |
> |
/* Compare entries by moving price */ |
| 220 |
> |
static int |
| 221 |
> |
rmovcmp(void *b, const void *p1, const void *p2) |
| 222 |
> |
{ |
| 223 |
> |
PRICEMAT *pm = (PRICEMAT *)b; |
| 224 |
> |
const short *ij1 = (const short *)p1; |
| 225 |
> |
const short *ij2 = (const short *)p2; |
| 226 |
> |
float price_diff; |
| 227 |
> |
|
| 228 |
> |
if (ij1[1] < 0) return(ij2[1] >= 0); |
| 229 |
> |
if (ij2[1] < 0) return(-1); |
| 230 |
> |
price_diff = pricerow(pm,ij1[0])[ij1[1]] - pricerow(pm,ij2[0])[ij2[1]]; |
| 231 |
> |
if (price_diff > 0) return(1); |
| 232 |
> |
if (price_diff < 0) return(-1); |
| 233 |
> |
return(0); |
| 234 |
> |
} |
| 235 |
> |
|
| 236 |
> |
/* Take a step in migration by choosing reasonable bucket to transfer */ |
| 237 |
|
static double |
| 238 |
< |
migration_step(MIGRATION *mig, double *src_rem, double *dst_rem, const PRICEMAT *pm) |
| 238 |
> |
migration_step(MIGRATION *mig, double *src_rem, double *dst_rem, PRICEMAT *pm) |
| 239 |
|
{ |
| 240 |
+ |
const int max2check = 100; |
| 241 |
|
const double maxamt = 1./(double)pm->ncols; |
| 242 |
< |
const double minamt = maxamt*5e-6; |
| 242 |
> |
const double minamt = maxamt*1e-4; |
| 243 |
|
double *src_cost; |
| 244 |
+ |
short (*rord)[2]; |
| 245 |
|
struct { |
| 246 |
|
int s, d; /* source and destination */ |
| 247 |
|
double price; /* price estimate per amount moved */ |
| 248 |
|
double amt; /* amount we can move */ |
| 249 |
|
} cur, best; |
| 250 |
< |
int i; |
| 250 |
> |
int r2check, i, ri; |
| 251 |
> |
/* |
| 252 |
> |
* Check cheapest available routes only -- a higher adjusted |
| 253 |
> |
* destination price implies that another source is closer, so |
| 254 |
> |
* we can hold off considering more expensive options until |
| 255 |
> |
* some other (hopefully better) moves have been made. |
| 256 |
> |
*/ |
| 257 |
> |
/* most promising row order */ |
| 258 |
> |
rord = (short (*)[2])malloc(sizeof(short)*2*pm->nrows); |
| 259 |
> |
if (rord == NULL) |
| 260 |
> |
goto memerr; |
| 261 |
> |
for (ri = pm->nrows; ri--; ) { |
| 262 |
> |
rord[ri][0] = ri; |
| 263 |
> |
rord[ri][1] = -1; |
| 264 |
> |
if (src_rem[ri] <= minamt) /* enough source material? */ |
| 265 |
> |
continue; |
| 266 |
> |
for (i = 0; i < pm->ncols; i++) |
| 267 |
> |
if (dst_rem[ rord[ri][1] = psortrow(pm,ri)[i] ] > minamt) |
| 268 |
> |
break; |
| 269 |
> |
if (i >= pm->ncols) { /* moved all we can? */ |
| 270 |
> |
free(rord); |
| 271 |
> |
return(.0); |
| 272 |
> |
} |
| 273 |
> |
} |
| 274 |
> |
if (pm->nrows > max2check) /* sort if too many sources */ |
| 275 |
> |
qsort_r(rord, pm->nrows, sizeof(short)*2, pm, &rmovcmp); |
| 276 |
|
/* allocate cost array */ |
| 277 |
|
src_cost = (double *)malloc(sizeof(double)*pm->nrows); |
| 278 |
< |
if (src_cost == NULL) { |
| 279 |
< |
fprintf(stderr, "%s: Out of memory in migration_step()\n", |
| 231 |
< |
progname); |
| 232 |
< |
exit(1); |
| 233 |
< |
} |
| 278 |
> |
if (src_cost == NULL) |
| 279 |
> |
goto memerr; |
| 280 |
|
for (i = pm->nrows; i--; ) /* starting costs for diff. */ |
| 281 |
|
src_cost[i] = min_cost(src_rem[i], dst_rem, pm, i); |
| 236 |
– |
|
| 282 |
|
/* find best source & dest. */ |
| 283 |
|
best.s = best.d = -1; best.price = FHUGE; best.amt = 0; |
| 284 |
< |
for (cur.s = pm->nrows; cur.s--; ) { |
| 284 |
> |
if ((r2check = pm->nrows) > max2check) |
| 285 |
> |
r2check = max2check; /* put a limit on search */ |
| 286 |
> |
for (ri = 0; ri < r2check; ri++) { /* check each source row */ |
| 287 |
|
double cost_others = 0; |
| 288 |
< |
|
| 289 |
< |
if (src_rem[cur.s] <= minamt) |
| 290 |
< |
continue; |
| 291 |
< |
/* examine cheapest dest. */ |
| 292 |
< |
for (i = 0; i < pm->ncols; i++) |
| 293 |
< |
if (dst_rem[ cur.d = psortrow(pm,cur.s)[i] ] > minamt) |
| 247 |
< |
break; |
| 248 |
< |
if (i >= pm->ncols) |
| 249 |
< |
break; |
| 250 |
< |
if ((cur.price = pricerow(pm,cur.s)[cur.d]) >= best.price) |
| 251 |
< |
continue; /* no point checking further */ |
| 288 |
> |
cur.s = rord[ri][0]; |
| 289 |
> |
if ((cur.d = rord[ri][1]) < 0 || |
| 290 |
> |
(cur.price = pricerow(pm,cur.s)[cur.d]) >= best.price) { |
| 291 |
> |
if (pm->nrows > max2check) break; /* sorted end */ |
| 292 |
> |
continue; /* else skip this one */ |
| 293 |
> |
} |
| 294 |
|
cur.amt = (src_rem[cur.s] < dst_rem[cur.d]) ? |
| 295 |
|
src_rem[cur.s] : dst_rem[cur.d]; |
| 296 |
< |
if (cur.amt > maxamt) cur.amt = maxamt; |
| 297 |
< |
dst_rem[cur.d] -= cur.amt; /* add up differential costs */ |
| 296 |
> |
/* don't just leave smidgen */ |
| 297 |
> |
if (cur.amt > maxamt*1.02) cur.amt = maxamt; |
| 298 |
> |
dst_rem[cur.d] -= cur.amt; /* add up opportunity costs */ |
| 299 |
|
for (i = pm->nrows; i--; ) |
| 300 |
|
if (i != cur.s) |
| 301 |
< |
cost_others += min_cost(src_rem[i], dst_rem, pm, i) |
| 301 |
> |
cost_others += min_cost(src_rem[i], dst_rem, pm, i) |
| 302 |
|
- src_cost[i]; |
| 303 |
|
dst_rem[cur.d] += cur.amt; /* undo trial move */ |
| 304 |
|
cur.price += cost_others/cur.amt; /* adjust effective price */ |
| 305 |
|
if (cur.price < best.price) /* are we better than best? */ |
| 306 |
< |
best = cur; |
| 306 |
> |
best = cur; |
| 307 |
|
} |
| 308 |
< |
free(src_cost); /* finish up */ |
| 309 |
< |
|
| 308 |
> |
free(src_cost); /* clean up */ |
| 309 |
> |
free(rord); |
| 310 |
|
if ((best.s < 0) | (best.d < 0)) /* nothing left to move? */ |
| 311 |
|
return(.0); |
| 312 |
|
/* else make the actual move */ |
| 314 |
|
src_rem[best.s] -= best.amt; |
| 315 |
|
dst_rem[best.d] -= best.amt; |
| 316 |
|
return(best.amt); |
| 317 |
+ |
memerr: |
| 318 |
+ |
fprintf(stderr, "%s: Out of memory in migration_step()\n", progname); |
| 319 |
+ |
exit(1); |
| 320 |
|
} |
| 321 |
|
|
| 322 |
|
/* Compute and insert migration along directed edge (may fork child) */ |
| 336 |
|
return(NULL); |
| 337 |
|
/* else allocate */ |
| 338 |
|
#ifdef DEBUG |
| 339 |
< |
fprintf(stderr, "Building path from (theta,phi) (%.0f,%.0f) ", |
| 339 |
> |
fprintf(stderr, "Building path from (theta,phi) (%.1f,%.1f) ", |
| 340 |
|
get_theta180(from_rbf->invec), |
| 341 |
|
get_phi360(from_rbf->invec)); |
| 342 |
< |
fprintf(stderr, "to (%.0f,%.0f) with %d x %d matrix\n", |
| 342 |
> |
fprintf(stderr, "to (%.1f,%.1f) with %d x %d matrix\n", |
| 343 |
|
get_theta180(to_rbf->invec), |
| 344 |
|
get_phi360(to_rbf->invec), |
| 345 |
|
from_rbf->nrbf, to_rbf->nrbf); |
| 407 |
|
return(vother[im_rev] != NULL); |
| 408 |
|
} |
| 409 |
|
|
| 410 |
< |
/* Find context hull vertex to complete triangle (oriented call) */ |
| 410 |
> |
/* Find convex hull vertex to complete triangle (oriented call) */ |
| 411 |
|
static RBFNODE * |
| 412 |
|
find_chull_vert(const RBFNODE *rbf0, const RBFNODE *rbf1) |
| 413 |
|
{ |
| 428 |
|
if (DOT(vp, vmid) <= FTINY) |
| 429 |
|
continue; /* wrong orientation */ |
| 430 |
|
area2 = .25*DOT(vp,vp); |
| 431 |
< |
VSUB(vp, rbf->invec, rbf0->invec); |
| 431 |
> |
VSUB(vp, rbf->invec, vmid); |
| 432 |
|
dprod = -DOT(vp, vejn); |
| 433 |
|
VSUM(vp, vp, vejn, dprod); /* above guarantees non-zero */ |
| 434 |
|
dprod = DOT(vp, vmid) / VLEN(vp); |
| 485 |
|
} |
| 486 |
|
} |
| 487 |
|
|
| 488 |
< |
/* Compute minimum BSDF from histogram and clear it */ |
| 488 |
> |
/* Add normal direction if missing */ |
| 489 |
|
static void |
| 490 |
< |
comp_bsdf_min() |
| 490 |
> |
check_normal_incidence(void) |
| 491 |
|
{ |
| 492 |
< |
int cnt; |
| 493 |
< |
int i, target; |
| 492 |
> |
const int saved_nprocs = nprocs; |
| 493 |
> |
RBFNODE *near_rbf, *mir_rbf, *rbf; |
| 494 |
> |
double bestd; |
| 495 |
> |
int n, i, j; |
| 496 |
|
|
| 497 |
< |
cnt = 0; |
| 498 |
< |
for (i = HISTLEN; i--; ) |
| 499 |
< |
cnt += bsdf_hist[i]; |
| 500 |
< |
|
| 501 |
< |
target = cnt/100; /* ignore bottom 1% */ |
| 502 |
< |
cnt = 0; |
| 503 |
< |
for (i = 0; cnt <= target; i++) |
| 504 |
< |
cnt += bsdf_hist[i]; |
| 505 |
< |
bsdf_min = histval(i-1); |
| 506 |
< |
memset(bsdf_hist, 0, sizeof(bsdf_hist)); |
| 497 |
> |
if (dsf_list == NULL) |
| 498 |
> |
return; /* XXX should be error? */ |
| 499 |
> |
near_rbf = dsf_list; |
| 500 |
> |
bestd = input_orient*near_rbf->invec[2]; |
| 501 |
> |
if (single_plane_incident) { /* ordered plane incidence? */ |
| 502 |
> |
if (bestd >= 1.-2.*FTINY) |
| 503 |
> |
return; /* already have normal */ |
| 504 |
> |
} else { |
| 505 |
> |
switch (inp_coverage) { |
| 506 |
> |
case INP_QUAD1: |
| 507 |
> |
case INP_QUAD2: |
| 508 |
> |
case INP_QUAD3: |
| 509 |
> |
case INP_QUAD4: |
| 510 |
> |
break; /* quadrilateral symmetry? */ |
| 511 |
> |
default: |
| 512 |
> |
return; /* else we can interpolate */ |
| 513 |
> |
} |
| 514 |
> |
for (rbf = near_rbf->next; rbf != NULL; rbf = rbf->next) { |
| 515 |
> |
const double d = input_orient*rbf->invec[2]; |
| 516 |
> |
if (d >= 1.-2.*FTINY) |
| 517 |
> |
return; /* seems we have normal */ |
| 518 |
> |
if (d > bestd) { |
| 519 |
> |
near_rbf = rbf; |
| 520 |
> |
bestd = d; |
| 521 |
> |
} |
| 522 |
> |
} |
| 523 |
> |
} |
| 524 |
> |
if (mig_list != NULL) { /* need to be called first */ |
| 525 |
> |
fprintf(stderr, "%s: Late call to check_normal_incidence()\n", |
| 526 |
> |
progname); |
| 527 |
> |
exit(1); |
| 528 |
> |
} |
| 529 |
> |
#ifdef DEBUG |
| 530 |
> |
fprintf(stderr, "Interpolating normal incidence by mirroring (%.1f,%.1f)\n", |
| 531 |
> |
get_theta180(near_rbf->invec), get_phi360(near_rbf->invec)); |
| 532 |
> |
#endif |
| 533 |
> |
/* mirror nearest incidence */ |
| 534 |
> |
n = sizeof(RBFNODE) + sizeof(RBFVAL)*(near_rbf->nrbf-1); |
| 535 |
> |
mir_rbf = (RBFNODE *)malloc(n); |
| 536 |
> |
if (mir_rbf == NULL) |
| 537 |
> |
goto memerr; |
| 538 |
> |
memcpy(mir_rbf, near_rbf, n); |
| 539 |
> |
mir_rbf->ord = near_rbf->ord - 1; /* not used, I think */ |
| 540 |
> |
mir_rbf->next = NULL; |
| 541 |
> |
rev_rbf_symmetry(mir_rbf, MIRROR_X|MIRROR_Y); |
| 542 |
> |
nprocs = 1; /* compute migration matrix */ |
| 543 |
> |
if (mig_list != create_migration(mir_rbf, near_rbf)) |
| 544 |
> |
exit(1); /* XXX should never happen! */ |
| 545 |
> |
n = 0; /* count migrating particles */ |
| 546 |
> |
for (i = 0; i < mtx_nrows(mig_list); i++) |
| 547 |
> |
for (j = 0; j < mtx_ncols(mig_list); j++) |
| 548 |
> |
n += (mtx_coef(mig_list,i,j) > FTINY); |
| 549 |
> |
rbf = (RBFNODE *)malloc(sizeof(RBFNODE) + sizeof(RBFVAL)*(n-1)); |
| 550 |
> |
if (rbf == NULL) |
| 551 |
> |
goto memerr; |
| 552 |
> |
rbf->next = NULL; rbf->ejl = NULL; |
| 553 |
> |
rbf->invec[0] = rbf->invec[1] = 0; rbf->invec[2] = 1.; |
| 554 |
> |
rbf->nrbf = n; |
| 555 |
> |
rbf->vtotal = .5 + .5*mig_list->rbfv[1]->vtotal/mig_list->rbfv[0]->vtotal; |
| 556 |
> |
n = 0; /* advect RBF lobes halfway */ |
| 557 |
> |
for (i = 0; i < mtx_nrows(mig_list); i++) { |
| 558 |
> |
const RBFVAL *rbf0i = &mig_list->rbfv[0]->rbfa[i]; |
| 559 |
> |
const float peak0 = rbf0i->peak; |
| 560 |
> |
const double rad0 = R2ANG(rbf0i->crad); |
| 561 |
> |
FVECT v0; |
| 562 |
> |
float mv; |
| 563 |
> |
ovec_from_pos(v0, rbf0i->gx, rbf0i->gy); |
| 564 |
> |
for (j = 0; j < mtx_ncols(mig_list); j++) |
| 565 |
> |
if ((mv = mtx_coef(mig_list,i,j)) > FTINY) { |
| 566 |
> |
const RBFVAL *rbf1j = &mig_list->rbfv[1]->rbfa[j]; |
| 567 |
> |
double rad2; |
| 568 |
> |
FVECT v; |
| 569 |
> |
int pos[2]; |
| 570 |
> |
rad2 = R2ANG(rbf1j->crad); |
| 571 |
> |
rad2 = .5*(rad0*rad0 + rad2*rad2); |
| 572 |
> |
rbf->rbfa[n].peak = peak0 * mv * rbf->vtotal * |
| 573 |
> |
rad0*rad0/rad2; |
| 574 |
> |
rbf->rbfa[n].crad = ANG2R(sqrt(rad2)); |
| 575 |
> |
ovec_from_pos(v, rbf1j->gx, rbf1j->gy); |
| 576 |
> |
geodesic(v, v0, v, .5, GEOD_REL); |
| 577 |
> |
pos_from_vec(pos, v); |
| 578 |
> |
rbf->rbfa[n].gx = pos[0]; |
| 579 |
> |
rbf->rbfa[n].gy = pos[1]; |
| 580 |
> |
++n; |
| 581 |
> |
} |
| 582 |
> |
} |
| 583 |
> |
rbf->vtotal *= mig_list->rbfv[0]->vtotal; |
| 584 |
> |
nprocs = saved_nprocs; /* final clean-up */ |
| 585 |
> |
free(mir_rbf); |
| 586 |
> |
free(mig_list); |
| 587 |
> |
mig_list = near_rbf->ejl = NULL; |
| 588 |
> |
insert_dsf(rbf); /* insert interpolated normal */ |
| 589 |
> |
return; |
| 590 |
> |
memerr: |
| 591 |
> |
fprintf(stderr, "%s: Out of memory in check_normal_incidence()\n", |
| 592 |
> |
progname); |
| 593 |
> |
exit(1); |
| 594 |
|
} |
| 595 |
|
|
| 596 |
|
/* Build our triangle mesh from recorded RBFs */ |
| 600 |
|
double best2 = M_PI*M_PI; |
| 601 |
|
RBFNODE *shrt_edj[2]; |
| 602 |
|
RBFNODE *rbf0, *rbf1; |
| 603 |
+ |
/* add normal if needed */ |
| 604 |
+ |
check_normal_incidence(); |
| 605 |
|
/* check if isotropic */ |
| 606 |
|
if (single_plane_incident) { |
| 607 |
|
for (rbf0 = dsf_list; rbf0 != NULL; rbf0 = rbf0->next) |
| 629 |
|
mesh_from_edge(create_migration(shrt_edj[0], shrt_edj[1])); |
| 630 |
|
else |
| 631 |
|
mesh_from_edge(create_migration(shrt_edj[1], shrt_edj[0])); |
| 495 |
– |
/* compute minimum BSDF */ |
| 496 |
– |
comp_bsdf_min(); |
| 632 |
|
/* complete migrations */ |
| 633 |
|
await_children(nchild); |
| 634 |
|
} |
