| 1 |
/* |
| 2 |
* Copyright(C) 1999-2021, 2023 National Technology & Engineering Solutions |
| 3 |
* of Sandia, LLC (NTESS). Under the terms of Contract DE-NA0003525 with |
| 4 |
* NTESS, the U.S. Government retains certain rights in this software. |
| 5 |
* |
| 6 |
* See packages/seacas/LICENSE for details |
| 7 |
*/ |
| 8 |
/**************************************************************************/ |
| 9 |
/* Program to perform orthogonal range searches and nearest neighbor */ |
| 10 |
/* queries in a more sophisticated k-d tree. In this implementation the, */ |
| 11 |
/* nodes on any given level of the tree do not have the same */ |
| 12 |
/* discriminating dimension as the discrimiator is chosen based on the */ |
| 13 |
/* dimension with the "maxspead." */ |
| 14 |
/* */ |
| 15 |
/* References: J.H. Friedman, J.L. Bentley, R.A. Finkel "An Algorithm */ |
| 16 |
/* for Finding Best Matches in Logarithmic Expected Time." */ |
| 17 |
/* ACM Transactions on Mathematical Software, Vol 3 No. 3 Sept. 1977 */ |
| 18 |
/* pp. 209-226. */ |
| 19 |
/**************************************************************************/ |
| 20 |
|
| 21 |
/* Turn off assertions here since SEACAS compiles don't normally do it. */ |
| 22 |
#ifndef NDEBUG |
| 23 |
#define NDEBUG |
| 24 |
#endif |
| 25 |
|
| 26 |
#if defined ADDC_ |
| 27 |
#define KDRECTQUERY kdrectquery_ |
| 28 |
#define KDKILLTREE kdkilltree_ |
| 29 |
#define KDBUILDTREE kdbuildtree_ |
| 30 |
#else |
| 31 |
#define KDRECTQUERY kdrectquery |
| 32 |
#define KDKILLTREE kdkilltree |
| 33 |
#define KDBUILDTREE kdbuildtree |
| 34 |
#endif |
| 35 |
|
| 36 |
#include <float.h> |
| 37 |
#include <math.h> |
| 38 |
#include <stdbool.h> |
| 39 |
#include <stdio.h> |
| 40 |
#include <stdlib.h> |
| 41 |
#include <string.h> |
| 42 |
|
| 43 |
#include <assert.h> |
| 44 |
|
| 45 |
#if defined(Build64) |
| 46 |
#define real double |
| 47 |
#else |
| 48 |
#define real float |
| 49 |
#endif |
| 50 |
|
| 51 |
#include "optkd.h" |
| 52 |
|
| 53 |
/* Used to create a new tree in the k-d tree */ |
| 54 |
#define TESTTREE(PP) ((PP) = (optkdNode *)malloc(sizeof(optkdNode))) |
| 55 |
#define NEWTREE(PP) \ |
| 56 |
if (TESTTREE(PP) == NULL) { \ |
| 57 |
printf("memory error\n"); \ |
| 58 |
return 0; \ |
| 59 |
} |
| 60 |
|
| 61 |
static optkdNode *Root = NULL; |
| 62 |
static int *perm = NULL; /* permutation array */ |
| 63 |
|
| 64 |
/***************************************************************************/ |
| 65 |
/* Makes the perm partition the array Values along the element k. */ |
| 66 |
/* Adapted from Sedgewick's Algorithms in C (p. 128) */ |
| 67 |
/***************************************************************************/ |
| 68 |
#define max(a, b) ((a) > (b) ? (a) : (b)) |
| 69 |
#define min(a, b) ((a) < (b) ? (a) : (b)) |
| 70 |
#define sign(x) ((x) >= 0 ? 1 : -1) |
| 71 |
|
| 72 |
/* |
| 73 |
* Floyd and Rivest SELECT Algorithm (CACM Algorithm 489) |
| 74 |
*/ |
| 75 |
void rf_select(real *a, int count, int l, int r, int k, int discrim) |
| 76 |
{ |
| 77 |
while (r > l) { |
| 78 |
if (r - l > 600) { |
| 79 |
/* |
| 80 |
* Use SELECT recursively on a sample of size S to get an estimate |
| 81 |
* for the (K-L+1)-th smallest element into A[K], biased slightly so |
| 82 |
* that the (K-L+1)-th element is expected to lie in the smaller set |
| 83 |
* after partitioning |
| 84 |
*/ |
| 85 |
int n = r - l + 1; |
| 86 |
int i = k - l + 1; |
| 87 |
real z = log(n); |
| 88 |
int s = 0.5 * exp(2.0 * z / 3.0); |
| 89 |
int sd = 0.1 * sqrt(z * s * (n - s) / n) * sign(i - n / 2.0); |
| 90 |
|
| 91 |
int tmp = k - (i * s / n) + sd; |
| 92 |
int ll = max(l, tmp); |
| 93 |
tmp = k + (n - i) * s / n + sd; |
| 94 |
int rr = min(r, tmp); |
| 95 |
rf_select(a, count, ll, rr, k, discrim); |
| 96 |
} |
| 97 |
real t = a[discrim * count + perm[k]]; |
| 98 |
|
| 99 |
/* The following code partitions a[L:R] about T. It is similar to |
| 100 |
PARTITION, but will run faster on most machines since subscript |
| 101 |
range checking on I and J has been eliminated. |
| 102 |
*/ |
| 103 |
int i = l; |
| 104 |
int j = r; |
| 105 |
/* exchange(x[l], x[k]); */ |
| 106 |
int tmp = perm[l]; |
| 107 |
perm[l] = perm[k]; |
| 108 |
perm[k] = tmp; |
| 109 |
if (a[discrim * count + perm[r]] > t) { |
| 110 |
/* exchange(x[r], x[l]); */ |
| 111 |
tmp = perm[r]; |
| 112 |
perm[r] = perm[l]; |
| 113 |
perm[l] = tmp; |
| 114 |
} |
| 115 |
while (i < j) { |
| 116 |
/* exchange(x[i], x[j]); */ |
| 117 |
tmp = perm[i]; |
| 118 |
perm[i] = perm[j]; |
| 119 |
perm[j] = tmp; |
| 120 |
i = i + 1; |
| 121 |
j = j - 1; |
| 122 |
while (a[discrim * count + perm[i]] < t) { |
| 123 |
i++; |
| 124 |
} |
| 125 |
while (a[discrim * count + perm[j]] > t) { |
| 126 |
j--; |
| 127 |
} |
| 128 |
} |
| 129 |
if (a[discrim * count + perm[l]] == t) { |
| 130 |
/* exchange(x[l], x[j]); */ |
| 131 |
tmp = perm[l]; |
| 132 |
perm[l] = perm[j]; |
| 133 |
perm[j] = tmp; |
| 134 |
} |
| 135 |
else { |
| 136 |
j++; |
| 137 |
/* exchange(x[j], x[r]); */ |
| 138 |
tmp = perm[j]; |
| 139 |
perm[j] = perm[r]; |
| 140 |
perm[r] = tmp; |
| 141 |
} |
| 142 |
/* Now adjust L, R so they surround the subset containing the (k-l+1)-th element */ |
| 143 |
if (j <= k) { |
| 144 |
l = j + 1; |
| 145 |
} |
| 146 |
if (k <= j) { |
| 147 |
r = j - 1; |
| 148 |
} |
| 149 |
} |
| 150 |
} |
| 151 |
|
| 152 |
/****************************************************************************/ |
| 153 |
int findmaxspread(int l, int u, int dimension, real *points, int N) |
| 154 |
{ |
| 155 |
int maxdim = 0; |
| 156 |
real maxspread = -FLT_MAX; |
| 157 |
|
| 158 |
for (int i = 0; i < dimension; i++) { |
| 159 |
real max = -FLT_MAX; |
| 160 |
real min = FLT_MAX; |
| 161 |
for (int j = l; j <= u; j++) { |
| 162 |
real val = points[N * i + perm[j]]; |
| 163 |
if (max < val) { |
| 164 |
max = val; |
| 165 |
} |
| 166 |
if (min > val) { |
| 167 |
min = val; |
| 168 |
} |
| 169 |
} |
| 170 |
if (maxspread < fabs(max - min)) { |
| 171 |
maxspread = fabs(max - min); |
| 172 |
maxdim = i; |
| 173 |
} |
| 174 |
} |
| 175 |
return (maxdim); |
| 176 |
} |
| 177 |
|
| 178 |
int findmaxvariance(int l, int u, int dimension, real *points, int N) |
| 179 |
{ |
| 180 |
int maxdim = 0; |
| 181 |
real max_var = 0.0; |
| 182 |
|
| 183 |
for (int i = 0; i < dimension; i++) { |
| 184 |
real mean = 0.0; |
| 185 |
real variance = 0.0; |
| 186 |
real count = 0.0; |
| 187 |
|
| 188 |
for (int j = l; j <= u; j++) { |
| 189 |
real val = points[N * i + perm[j]]; |
| 190 |
real prev_mean = mean; |
| 191 |
count += 1.0; |
| 192 |
mean += (val - prev_mean) / count; |
| 193 |
variance += (val - prev_mean) * (val - mean); |
| 194 |
} |
| 195 |
/* True variance would be divided by (count-1.0) */ |
| 196 |
|
| 197 |
if (max_var < variance) { |
| 198 |
max_var = variance; |
| 199 |
maxdim = i; |
| 200 |
} |
| 201 |
} |
| 202 |
return (maxdim); |
| 203 |
} |
| 204 |
|
| 205 |
/*******************************************************************************/ |
| 206 |
int check(real *points, int N, int l, int u, int m, int discrim) |
| 207 |
{ |
| 208 |
for (int i = l; i < m; i++) { |
| 209 |
assert(points[discrim * N + perm[i]] <= points[discrim * N + perm[m]]); |
| 210 |
} |
| 211 |
for (int i = m; i < u; i++) { |
| 212 |
assert(points[discrim * N + perm[i]] >= points[discrim * N + perm[m]]); |
| 213 |
} |
| 214 |
return 1; |
| 215 |
} |
| 216 |
|
| 217 |
optkdNode *BuildkdTree(real *points, int N, int l, int u, int dimension) |
| 218 |
{ |
| 219 |
optkdNode *p; |
| 220 |
|
| 221 |
NEWTREE(p); |
| 222 |
if (u - l + 1 <= BUCKETSIZE) { |
| 223 |
p->bucket = 1; |
| 224 |
p->lopt = l; |
| 225 |
p->hipt = u; |
| 226 |
p->loson = NULL; |
| 227 |
p->hison = NULL; |
| 228 |
} |
| 229 |
else { |
| 230 |
p->bucket = 0; |
| 231 |
|
| 232 |
#if 1 |
| 233 |
p->discrim = findmaxspread(l, u, dimension, points, N); |
| 234 |
#else |
| 235 |
p->discrim = findmaxvariance(l, u, dimension, points, N); |
| 236 |
#endif |
| 237 |
|
| 238 |
int m = (l + u) / 2; |
| 239 |
|
| 240 |
rf_select(points, N, l, u, m, p->discrim); |
| 241 |
assert(check(points, N, l, u, m, p->discrim)); |
| 242 |
|
| 243 |
p->cutval = points[p->discrim * N + perm[m]]; |
| 244 |
p->loson = BuildkdTree(points, N, l, m, dimension); |
| 245 |
p->hison = BuildkdTree(points, N, m + 1, u, dimension); |
| 246 |
} |
| 247 |
return (p); |
| 248 |
} |
| 249 |
|
| 250 |
/*******************************************************************************/ |
| 251 |
void KDBUILDTREE(real *points, int *numPoints, int *dimension) |
| 252 |
{ |
| 253 |
/* initialize perm array */ |
| 254 |
assert(perm == NULL); |
| 255 |
perm = (int *)malloc(*numPoints * sizeof(int)); |
| 256 |
assert(perm != NULL); |
| 257 |
for (int j = 0; j < *numPoints; j++) { |
| 258 |
perm[j] = j; |
| 259 |
} |
| 260 |
assert(Root == NULL); |
| 261 |
Root = BuildkdTree(points, *numPoints, 0, *numPoints - 1, *dimension); |
| 262 |
assert(Root != NULL); |
| 263 |
} |
| 264 |
|
| 265 |
/***************************************************************************/ |
| 266 |
void KillOptTree(optkdNode *P) |
| 267 |
{ |
| 268 |
/* Kills a kd-tree to avoid memory holes. */ |
| 269 |
if (P == NULL) { |
| 270 |
return; |
| 271 |
} /* just to be sure */ |
| 272 |
|
| 273 |
if (P->loson != NULL) { |
| 274 |
KillOptTree(P->loson); |
| 275 |
} |
| 276 |
|
| 277 |
if (P->hison != NULL) { |
| 278 |
KillOptTree(P->hison); |
| 279 |
} |
| 280 |
|
| 281 |
free(P); |
| 282 |
} |
| 283 |
|
| 284 |
void KDKILLTREE(void) |
| 285 |
{ |
| 286 |
if (perm != NULL) { |
| 287 |
free(perm); |
| 288 |
perm = NULL; |
| 289 |
} /* free permutation array */ |
| 290 |
|
| 291 |
KillOptTree(Root); |
| 292 |
Root = NULL; |
| 293 |
} |
| 294 |
|
| 295 |
/***************************************************************************/ |
| 296 |
/* Determines if the treenode P falls inside the rectangular query */ |
| 297 |
/* xmin,xmax. If so, adds the array index of the point to the found */ |
| 298 |
/* array. */ |
| 299 |
/***************************************************************************/ |
| 300 |
void optInRegion(optkdNode *P, int Dimension, real *Points, int N, real *xmin, real *xmax, |
| 301 |
int *found, int *count) |
| 302 |
{ |
| 303 |
for (int index = P->lopt; index <= P->hipt; index++) { |
| 304 |
bool InsideRange = true; |
| 305 |
|
| 306 |
if (Dimension == 3) { |
| 307 |
int inval = perm[index]; |
| 308 |
if (Points[0 * N + inval] < xmin[0] || Points[0 * N + inval] > xmax[0] || |
| 309 |
Points[1 * N + inval] < xmin[1] || Points[1 * N + inval] > xmax[1] || |
| 310 |
Points[2 * N + inval] < xmin[2] || |
| 311 |
Points[2 * N + inval] > xmax[2]) { /* P is not in the region */ |
| 312 |
InsideRange = false; |
| 313 |
} |
| 314 |
} |
| 315 |
else { |
| 316 |
for (int dc = 0; dc < Dimension; dc++) { |
| 317 |
if (Points[dc * N + perm[index]] < xmin[dc] || |
| 318 |
Points[dc * N + perm[index]] > xmax[dc]) { /* P is not in the region */ |
| 319 |
InsideRange = false; |
| 320 |
break; |
| 321 |
} |
| 322 |
} |
| 323 |
} |
| 324 |
if (InsideRange) { |
| 325 |
found[(*count)++] = perm[index] + 1; |
| 326 |
} |
| 327 |
} |
| 328 |
} |
| 329 |
|
| 330 |
/***************************************************************************/ |
| 331 |
/* Adds the array index of each point in the bucket the point to the found */ |
| 332 |
/* array. There is no need to check if the points are in it because we */ |
| 333 |
/* have proven so already. */ |
| 334 |
/***************************************************************************/ |
| 335 |
void optAddRegion(optkdNode *P, int Dimension, int *found, int *count) |
| 336 |
{ |
| 337 |
for (int index = P->lopt; index <= P->hipt; index++) { |
| 338 |
found[(*count)++] = perm[index] + 1; |
| 339 |
} |
| 340 |
} |
| 341 |
|
| 342 |
/***************************************************************************/ |
| 343 |
/* Returns true iff the hyper-rectangle defined by bounds array B */ |
| 344 |
/* intersects the rectangular query xmin,xmax. */ |
| 345 |
/***************************************************************************/ |
| 346 |
int optBoundsIntersectRegion(real *B, real *xmin, real *xmax, int Dimension) |
| 347 |
{ |
| 348 |
if (Dimension == 3) { |
| 349 |
if (B[0] > xmax[0] || B[1] < xmin[0] || B[2] > xmax[1] || B[3] < xmin[1] || B[4] > xmax[2] || |
| 350 |
B[5] < xmin[2]) { |
| 351 |
return 0; |
| 352 |
} |
| 353 |
} |
| 354 |
else { |
| 355 |
int dc; |
| 356 |
|
| 357 |
for (dc = 0; dc < Dimension; dc++) { |
| 358 |
if (B[2 * dc] > xmax[dc] || B[2 * dc + 1] < xmin[dc]) { |
| 359 |
return (0); |
| 360 |
} |
| 361 |
} |
| 362 |
} |
| 363 |
return (1); |
| 364 |
} |
| 365 |
|
| 366 |
/***************************************************************************/ |
| 367 |
/* Returns true iff the hyper-rectangle defined by bounds array B */ |
| 368 |
/* is completely contained inside the rectangular query xmin,xmax. */ |
| 369 |
/***************************************************************************/ |
| 370 |
int optBoundsContainsRegion(real *B, real *xmin, real *xmax, int Dimension) |
| 371 |
{ |
| 372 |
if (Dimension == 3) { |
| 373 |
if (xmin[0] <= B[0] && xmax[0] >= B[1] && xmin[1] <= B[2] && xmax[1] >= B[3] && |
| 374 |
xmin[2] <= B[4] && xmax[2] >= B[5]) { |
| 375 |
return 1; |
| 376 |
} |
| 377 |
else { |
| 378 |
return 0; |
| 379 |
} |
| 380 |
} |
| 381 |
else { |
| 382 |
int dc; |
| 383 |
|
| 384 |
for (dc = 0; dc < Dimension; dc++) { |
| 385 |
if (!(B[2 * dc] >= xmin[dc] && B[2 * dc + 1] <= xmax[dc])) { |
| 386 |
return (0); |
| 387 |
} |
| 388 |
} |
| 389 |
} |
| 390 |
return (1); |
| 391 |
} |
| 392 |
|
| 393 |
#define MAX_DEPTH 60 |
| 394 |
/***************************************************************************/ |
| 395 |
void optRangeSearch(optkdNode *P, real *Points, int N, int Dimension, real *xmin, real *xmax, |
| 396 |
real *B, int *found, int *count, int depth) |
| 397 |
{ |
| 398 |
static real BLow[MAX_DEPTH][6]; |
| 399 |
static real BHigh[MAX_DEPTH][6]; |
| 400 |
|
| 401 |
if (depth >= MAX_DEPTH) { |
| 402 |
fprintf(stderr, "Internal Error in optRangeSearch -- recursion depth too large.\n"); |
| 403 |
abort(); |
| 404 |
} |
| 405 |
|
| 406 |
if (P == NULL) { |
| 407 |
fprintf(stderr, "Internal Error in optRangeSearch -- P is NULL.\n"); |
| 408 |
abort(); |
| 409 |
} |
| 410 |
|
| 411 |
if (P->bucket) { |
| 412 |
if (optBoundsContainsRegion(B, xmin, xmax, Dimension)) { |
| 413 |
optAddRegion(P, Dimension, found, count); |
| 414 |
} |
| 415 |
else { |
| 416 |
optInRegion(P, Dimension, Points, N, xmin, xmax, found, count); |
| 417 |
} |
| 418 |
return; |
| 419 |
} |
| 420 |
|
| 421 |
/* Claim: P is not a bucket node */ |
| 422 |
int disc = P->discrim; |
| 423 |
/* copy the region B into BLow, BHigh */ |
| 424 |
|
| 425 |
if (Dimension == 3) { |
| 426 |
memcpy(&BLow[depth][0], &B[0], 6 * sizeof(real)); |
| 427 |
memcpy(&BHigh[depth][0], &B[0], 6 * sizeof(real)); |
| 428 |
} |
| 429 |
else { |
| 430 |
memcpy(&BLow[depth][0], &B[0], 2 * Dimension * sizeof(real)); |
| 431 |
memcpy(&BHigh[depth][0], &B[0], 2 * Dimension * sizeof(real)); |
| 432 |
} |
| 433 |
|
| 434 |
/* Improve the Bounds for the subtrees */ |
| 435 |
BLow[depth][2 * disc + 1] = P->cutval; |
| 436 |
BHigh[depth][2 * disc] = P->cutval; |
| 437 |
|
| 438 |
if (optBoundsIntersectRegion(&BLow[depth][0], xmin, xmax, Dimension)) { |
| 439 |
optRangeSearch(P->loson, Points, N, Dimension, xmin, xmax, &BLow[depth][0], found, count, |
| 440 |
depth + 1); |
| 441 |
} |
| 442 |
if (optBoundsIntersectRegion(&BHigh[depth][0], xmin, xmax, Dimension)) { |
| 443 |
optRangeSearch(P->hison, Points, N, Dimension, xmin, xmax, &BHigh[depth][0], found, count, |
| 444 |
depth + 1); |
| 445 |
} |
| 446 |
} |
| 447 |
|
| 448 |
/***************************************************************************/ |
| 449 |
void KDRECTQUERY(real *Points, int *N, int *Dimension, real *xmin, real *xmax, int *found, |
| 450 |
int *count) |
| 451 |
{ |
| 452 |
real B[6] = {-FLT_MAX, FLT_MAX, -FLT_MAX, FLT_MAX, -FLT_MAX, FLT_MAX}; |
| 453 |
*count = 0; |
| 454 |
optRangeSearch(Root, Points, *N, *Dimension, xmin, xmax, B, found, count, 0); |
| 455 |
} |