You cannot select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
3dpcp/.svn/pristine/35/35452722e3484f4b8a583e474ef...

387 lines
11 KiB
Plaintext

/** @file
* @brief Representation of the optimized k-d tree.
* @author Remus Dumitru. Jacobs University Bremen, Germany
* @author Corneliu-Claudiu Prodescu. Jacobs University Bremen, Germany
* @author Andreas Nuechter. Jacobs University Bremen, Germany
* @author Kai Lingemann. Inst. of CS, University of Osnabrueck, Germany
* @author Thomas Escher. Inst. of CS, University of Osnabrueck, Germany
*/
#ifndef __KD_TREE_IMPL_H__
#define __KD_TREE_IMPL_H__
#include "slam6d/kdparams.h"
#include "globals.icc"
#ifdef _MSC_VER
#if !defined _OPENMP && defined OPENMP
#define _OPENMP
#endif
#endif
#ifdef _OPENMP
#include <omp.h>
#endif
class PointCompare {
public:
bool operator() (const std::pair<Point, double>& left,
const std::pair<Point, double>& right)
{
return left.second > right.second;
}
};
/**
* @brief The optimized k-d tree.
*
* A kD tree for points, with limited
* capabilities (find nearest point to
* a given point, or to a ray).
**/
template<class PointData, class AccessorData, class AccessorFunc>
class KDTreeImpl {
public:
inline KDTreeImpl() { }
virtual inline ~KDTreeImpl() {
if (!npts) {
#ifdef WITH_OPENMP_KD
omp_set_num_threads(OPENMP_NUM_THREADS);
#pragma omp parallel for schedule(dynamic)
#endif
for (int i = 0; i < 2; i++) {
if (i == 0 && node.child1) delete node.child1;
if (i == 1 && node.child2) delete node.child2;
}
} else {
if (leaf.p) delete [] leaf.p;
}
}
virtual void create(PointData pts, AccessorData *indices, size_t n) {
AccessorFunc point;
// Find bbox
double xmin = point(pts, indices[0])[0], xmax = point(pts, indices[0])[0];
double ymin = point(pts, indices[0])[1], ymax = point(pts, indices[0])[1];
double zmin = point(pts, indices[0])[2], zmax = point(pts, indices[0])[2];
for(unsigned int i = 1; i < n; i++) {
xmin = min(xmin, point(pts, indices[i])[0]);
xmax = max(xmax, point(pts, indices[i])[0]);
ymin = min(ymin, point(pts, indices[i])[1]);
ymax = max(ymax, point(pts, indices[i])[1]);
zmin = min(zmin, point(pts, indices[i])[2]);
zmax = max(zmax, point(pts, indices[i])[2]);
}
// Leaf nodes
if ((n > 0) && (n <= 10)) {
npts = n;
leaf.p = new AccessorData[n];
// fill leaf index array with indices
for(unsigned int i = 0; i < n; ++i) {
leaf.p[i] = indices[i];
}
return;
}
// Else, interior nodes
npts = 0;
node.center[0] = 0.5 * (xmin+xmax);
node.center[1] = 0.5 * (ymin+ymax);
node.center[2] = 0.5 * (zmin+zmax);
node.dx = 0.5 * (xmax-xmin);
node.dy = 0.5 * (ymax-ymin);
node.dz = 0.5 * (zmax-zmin);
node.r2 = sqr(node.dx) + sqr(node.dy) + sqr(node.dz);
// Find longest axis
if (node.dx > node.dy) {
if (node.dx > node.dz) {
node.splitaxis = 0;
} else {
node.splitaxis = 2;
}
} else {
if (node.dy > node.dz) {
node.splitaxis = 1;
} else {
node.splitaxis = 2;
}
}
// Partition
double splitval = node.center[node.splitaxis];
if ( fabs(max(max(node.dx,node.dy),node.dz)) < 0.01 ) {
npts = n;
leaf.p = new AccessorData[n];
// fill leaf index array with indices
for(unsigned int i = 0; i < n; ++i) {
leaf.p[i] = indices[i];
}
return;
}
AccessorData* left = indices, * right = indices + n - 1;
while(true) {
while(point(pts, *left)[node.splitaxis] < splitval)
left++;
while(point(pts, *right)[node.splitaxis] >= splitval)
right--;
if(right < left)
break;
std::swap(*left, *right);
}
// Build subtrees
int i;
#ifdef WITH_OPENMP_KD // does anybody know the reason why this is slower ?? --Andreas
omp_set_num_threads(OPENMP_NUM_THREADS);
#pragma omp parallel for schedule(dynamic)
#endif
for (i = 0; i < 2; i++) {
if (i == 0) {
node.child1 = new KDTreeImpl();
node.child1->create(pts, indices, left - indices);
}
if (i == 1) {
node.child2 = new KDTreeImpl();
node.child2->create(pts, left, n - (left - indices));
}
}
}
protected:
/**
* storing the parameters of the k-d tree, i.e., the current closest point,
* the distance to the current closest point and the point itself.
* These global variable are needed in this search.
*
* Padded in the parallel case.
*/
#ifdef _OPENMP
#ifdef __INTEL_COMPILER
__declspec (align(16)) static KDParams params[MAX_OPENMP_NUM_THREADS];
#else
static KDParams params[MAX_OPENMP_NUM_THREADS];
#endif //__INTEL_COMPILER
#else
static KDParams params[MAX_OPENMP_NUM_THREADS];
#endif
/**
* number of points. If this is 0: intermediate node. If nonzero: leaf.
*/
int npts;
/**
* Cue the standard rant about anon unions but not structs in C++
*/
union {
/**
* in case of internal node...
*/
struct {
double center[3]; ///< storing the center of the voxel (R^3)
double dx, ///< defining the voxel itself
dy, ///< defining the voxel itself
dz, ///< defining the voxel itself
r2; ///< defining the voxel itself
int splitaxis; ///< defining the kind of splitaxis
KDTreeImpl *child1; ///< pointers to the childs
KDTreeImpl *child2; ///< pointers to the childs
} node;
/**
* in case of leaf node ...
*/
struct {
/**
* store the value itself
* Here we store just a pointer to the data
*/
AccessorData* p;
} leaf;
};
void _FindClosest(const PointData& pts, int threadNum) const {
AccessorFunc point;
// Leaf nodes
if (npts) {
for (int i = 0; i < npts; i++) {
double myd2 = Dist2(params[threadNum].p, point(pts, leaf.p[i]));
if (myd2 < params[threadNum].closest_d2) {
params[threadNum].closest_d2 = myd2;
params[threadNum].closest = point(pts, leaf.p[i]);
}
}
return;
}
// Quick check of whether to abort
double approx_dist_bbox =
max(max(fabs(params[threadNum].p[0]-node.center[0])-node.dx,
fabs(params[threadNum].p[1]-node.center[1])-node.dy),
fabs(params[threadNum].p[2]-node.center[2])-node.dz);
if (approx_dist_bbox >= 0 &&
sqr(approx_dist_bbox) >= params[threadNum].closest_d2)
return;
// Recursive case
double myd = node.center[node.splitaxis] - params[threadNum].p[node.splitaxis];
if (myd >= 0.0) {
node.child1->_FindClosest(pts, threadNum);
if (sqr(myd) < params[threadNum].closest_d2) {
node.child2->_FindClosest(pts, threadNum);
}
} else {
node.child2->_FindClosest(pts, threadNum);
if (sqr(myd) < params[threadNum].closest_d2) {
node.child1->_FindClosest(pts, threadNum);
}
}
}
void _FindClosestAlongDir(const PointData& pts, int threadNum) const {
AccessorFunc point;
// Leaf nodes
if (npts) {
for (int i=0; i < npts; i++) {
double p2p[] = { params[threadNum].p[0] - point(pts, leaf.p[i])[0],
params[threadNum].p[1] - point(pts, leaf.p[i])[1],
params[threadNum].p[2] - point(pts, leaf.p[i])[2] };
double myd2 = Len2(p2p) - sqr(Dot(p2p, params[threadNum].dir));
if ((myd2 < params[threadNum].closest_d2)) {
params[threadNum].closest_d2 = myd2;
params[threadNum].closest = point(pts, leaf.p[i]);
}
}
return;
}
// Quick check of whether to abort
double p2c[] = { params[threadNum].p[0] - node.center[0],
params[threadNum].p[1] - node.center[1],
params[threadNum].p[2] - node.center[2] };
double myd2center = Len2(p2c) - sqr(Dot(p2c, params[threadNum].dir));
if (myd2center > node.r2 + params[threadNum].closest_d2 + 2.0f * max(node.r2, params[threadNum].closest_d2))
return;
// Recursive case
if (params[threadNum].p[node.splitaxis] < node.center[node.splitaxis] ) {
node.child1->_FindClosestAlongDir(pts, threadNum);
node.child2->_FindClosestAlongDir(pts, threadNum);
} else {
node.child2->_FindClosestAlongDir(pts, threadNum);
node.child1->_FindClosestAlongDir(pts, threadNum);
}
}
void _FixedRangeSearch(const PointData& pts, int threadNum) const {
AccessorFunc point;
// Leaf nodes
if (npts) {
for (int i = 0; i < npts; i++) {
double myd2 = Dist2(params[threadNum].p, point(pts, leaf.p[i]));
if (myd2 < params[threadNum].closest_d2) {
params[threadNum].closest = point(pts, leaf.p[i]);
Point newPt;
double* currPt = point(pts, leaf.p[i]);
newPt.x = currPt[0];
newPt.y = currPt[1];
newPt.z = currPt[2];
params[threadNum].heap.push_back(std::make_pair(newPt, myd2));
std::push_heap(params[threadNum].heap.begin(),
params[threadNum].heap.end(),
PointCompare());
}
}
return;
}
// Quick check of whether to abort
double approx_dist_bbox =
max(max(fabs(params[threadNum].p[0]-node.center[0])-node.dx,
fabs(params[threadNum].p[1]-node.center[1])-node.dy),
fabs(params[threadNum].p[2]-node.center[2])-node.dz);
if (approx_dist_bbox >= 0 &&
sqr(approx_dist_bbox) >= params[threadNum].closest_d2)
return;
// Recursive case
double myd = node.center[node.splitaxis] - params[threadNum].p[node.splitaxis];
if (myd >= 0.0) {
node.child1->_FixedRangeSearch(pts, threadNum);
if (sqr(myd) < params[threadNum].closest_d2) {
node.child2->_FixedRangeSearch(pts, threadNum);
}
} else {
node.child2->_FixedRangeSearch(pts, threadNum);
if (sqr(myd) < params[threadNum].closest_d2) {
node.child1->_FixedRangeSearch(pts, threadNum);
}
}
}
void _KNNSearch(const PointData& pts, int threadNum) const {
AccessorFunc point;
// Leaf nodes
if (npts) {
for (int i = 0; i < npts; i++) {
double myd2 = Dist2(params[threadNum].p, point(pts, leaf.p[i]));
if (myd2 < params[threadNum].closest_d2) {
Point newPt;
double* currPt = point(pts, leaf.p[i]);
newPt.x = currPt[0];
newPt.y = currPt[1];
newPt.z = currPt[2];
params[threadNum].heap.push_back(std::make_pair(newPt, myd2));
std::push_heap(params[threadNum].heap.begin(),
params[threadNum].heap.end(),
PointCompare());
params[threadNum].closest = point(pts, leaf.p[i]);
}
}
return;
}
// Quick check of whether to abort
double approx_dist_bbox =
max(max(fabs(params[threadNum].p[0]-node.center[0])-node.dx,
fabs(params[threadNum].p[1]-node.center[1])-node.dy),
fabs(params[threadNum].p[2]-node.center[2])-node.dz);
if (approx_dist_bbox >= 0 &&
sqr(approx_dist_bbox) >= params[threadNum].closest_d2)
return;
// Recursive case
double myd = node.center[node.splitaxis] - params[threadNum].p[node.splitaxis];
if (myd >= 0.0) {
node.child1->_KNNSearch(pts, threadNum);
if (sqr(myd) < params[threadNum].closest_d2) {
node.child2->_KNNSearch(pts, threadNum);
}
} else {
node.child2->_KNNSearch(pts, threadNum);
if (sqr(myd) < params[threadNum].closest_d2) {
node.child1->_KNNSearch(pts, threadNum);
}
}
}
};
#endif