size_t BVH4MB::rotate(Base* nodeID, size_t depth)
{
/*! nothing to rotate if we reached a leaf node. */
if (nodeID->isLeaf()) return 0;
Node* parent = nodeID->node();
/*! rotate all children first */
ssei cdepth;
for (size_t c=0; c<4; c++)
cdepth[c] = (int)rotate(parent->child[c],depth+1);
/* compute current area of all children */
ssef sizeX = parent->upper_x-parent->lower_x;
ssef sizeY = parent->upper_y-parent->lower_y;
ssef sizeZ = parent->upper_z-parent->lower_z;
ssef childArea = sizeX*(sizeY + sizeZ) + sizeY*sizeZ;
/*! transpose node bounds */
ssef plower0,plower1,plower2,plower3; transpose(parent->lower_x,parent->lower_y,parent->lower_z,ssef(zero),plower0,plower1,plower2,plower3);
ssef pupper0,pupper1,pupper2,pupper3; transpose(parent->upper_x,parent->upper_y,parent->upper_z,ssef(zero),pupper0,pupper1,pupper2,pupper3);
BBox<ssef> other0(plower0,pupper0), other1(plower1,pupper1), other2(plower2,pupper2), other3(plower3,pupper3);
/*! Find best rotation. We pick a target child of a first child,
and swap this with an other child. We perform the best such
swap. */
float bestCost = pos_inf;
int bestChild = -1, bestTarget = -1, bestOther = -1;
for (size_t c=0; c<4; c++)
{
/*! ignore leaf nodes as we cannot descent into */
if (parent->child[c]->isLeaf()) continue;
Node* child = parent->child[c]->node();
/*! transpose child bounds */
ssef clower0,clower1,clower2,clower3; transpose(child->lower_x,child->lower_y,child->lower_z,ssef(zero),clower0,clower1,clower2,clower3);
ssef cupper0,cupper1,cupper2,cupper3; transpose(child->upper_x,child->upper_y,child->upper_z,ssef(zero),cupper0,cupper1,cupper2,cupper3);
BBox<ssef> target0(clower0,cupper0), target1(clower1,cupper1), target2(clower2,cupper2), target3(clower3,cupper3);
/*! put other0 at each target position */
float cost00 = halfArea3f(merge(other0 ,target1,target2,target3));
float cost01 = halfArea3f(merge(target0,other0 ,target2,target3));
float cost02 = halfArea3f(merge(target0,target1,other0 ,target3));
float cost03 = halfArea3f(merge(target0,target1,target2,other0 ));
ssef cost0 = ssef(cost00,cost01,cost02,cost03);
ssef min0 = vreduce_min(cost0);
int pos0 = (int)__bsf(movemask(min0 == cost0));
/*! put other1 at each target position */
float cost10 = halfArea3f(merge(other1 ,target1,target2,target3));
float cost11 = halfArea3f(merge(target0,other1 ,target2,target3));
float cost12 = halfArea3f(merge(target0,target1,other1 ,target3));
float cost13 = halfArea3f(merge(target0,target1,target2,other1 ));
ssef cost1 = ssef(cost10,cost11,cost12,cost13);
ssef min1 = vreduce_min(cost1);
int pos1 = (int)__bsf(movemask(min1 == cost1));
/*! put other2 at each target position */
float cost20 = halfArea3f(merge(other2 ,target1,target2,target3));
float cost21 = halfArea3f(merge(target0,other2 ,target2,target3));
float cost22 = halfArea3f(merge(target0,target1,other2 ,target3));
float cost23 = halfArea3f(merge(target0,target1,target2,other2 ));
ssef cost2 = ssef(cost20,cost21,cost22,cost23);
ssef min2 = vreduce_min(cost2);
int pos2 = (int)__bsf(movemask(min2 == cost2));
/*! put other3 at each target position */
float cost30 = halfArea3f(merge(other3 ,target1,target2,target3));
float cost31 = halfArea3f(merge(target0,other3 ,target2,target3));
float cost32 = halfArea3f(merge(target0,target1,other3 ,target3));
float cost33 = halfArea3f(merge(target0,target1,target2,other3 ));
ssef cost3 = ssef(cost30,cost31,cost32,cost33);
ssef min3 = vreduce_min(cost3);
int pos3 = (int)__bsf(movemask(min3 == cost3));
/*! find best other child */
ssef otherCost = ssef(extract<0>(min0),extract<0>(min1),extract<0>(min2),extract<0>(min3));
int pos[4] = { pos0,pos1,pos2,pos3 };
sseb valid = ssei(int(depth+1))+cdepth <= ssei(maxDepth); // only select swaps that fulfill depth constraints
if (none(valid)) continue;
size_t n = select_min(valid,otherCost);
float cost = otherCost[n]-childArea[c]; //< increasing the original child bound is bad, decreasing good
/*! accept a swap when it reduces cost and is not swapping a node with itself */
if (cost < bestCost && n != c) {
bestCost = cost;
bestChild = (int)c;
bestOther = (int)n;
bestTarget = pos[n];
}
}
/*! if we did not find a swap that improves the SAH then do nothing */
if (bestCost >= 0) return 1+reduce_max(cdepth);
/*! perform the best found tree rotation */
Node* child = parent->child[bestChild]->node();
swap(parent,bestOther,child,bestTarget);
parent->lower_x[bestChild] = reduce_min(child->lower_x);
parent->lower_y[bestChild] = reduce_min(child->lower_y);
parent->lower_z[bestChild] = reduce_min(child->lower_z);
parent->upper_x[bestChild] = reduce_max(child->upper_x);
parent->upper_y[bestChild] = reduce_max(child->upper_y);
parent->upper_z[bestChild] = reduce_max(child->upper_z);
parent->lower_dx[bestChild] = reduce_min(child->lower_dx);
parent->lower_dy[bestChild] = reduce_min(child->lower_dy);
parent->lower_dz[bestChild] = reduce_min(child->lower_dz);
parent->upper_dx[bestChild] = reduce_max(child->upper_dx);
parent->upper_dy[bestChild] = reduce_max(child->upper_dy);
parent->upper_dz[bestChild] = reduce_max(child->upper_dz);
/*! This returned depth is conservative as the child that was
* pulled up in the tree could have been on the critical path. */
cdepth[bestOther]++; // bestOther was pushed down one level
return 1+reduce_max(cdepth);
}
size_t BVHNRotate<4>::rotate(NodeRef parentRef, size_t depth)
{
/*! nothing to rotate if we reached a leaf node. */
if (parentRef.isBarrier()) return 0;
if (parentRef.isLeaf()) return 0;
Node* parent = parentRef.node();
/*! rotate all children first */
vint4 cdepth;
for (size_t c=0; c<4; c++)
cdepth[c] = (int)rotate(parent->child(c),depth+1);
/* compute current areas of all children */
vfloat4 sizeX = parent->upper_x-parent->lower_x;
vfloat4 sizeY = parent->upper_y-parent->lower_y;
vfloat4 sizeZ = parent->upper_z-parent->lower_z;
vfloat4 childArea = sizeX*(sizeY + sizeZ) + sizeY*sizeZ;
/*! get node bounds */
BBox<vfloat4> child1_0,child1_1,child1_2,child1_3;
parent->bounds(child1_0,child1_1,child1_2,child1_3);
/*! Find best rotation. We pick a first child (child1) and a sub-child
(child2child) of a different second child (child2), and swap child1
and child2child. We perform the best such swap. */
float bestArea = 0;
size_t bestChild1 = -1, bestChild2 = -1, bestChild2Child = -1;
for (size_t c2=0; c2<4; c2++)
{
/*! ignore leaf nodes as we cannot descent into them */
if (parent->child(c2).isBarrier()) continue;
if (parent->child(c2).isLeaf()) continue;
Node* child2 = parent->child(c2).node();
/*! transpose child bounds */
BBox<vfloat4> child2c0,child2c1,child2c2,child2c3;
child2->bounds(child2c0,child2c1,child2c2,child2c3);
/*! put child1_0 at each child2 position */
float cost00 = halfArea3f(merge(child1_0,child2c1,child2c2,child2c3));
float cost01 = halfArea3f(merge(child2c0,child1_0,child2c2,child2c3));
float cost02 = halfArea3f(merge(child2c0,child2c1,child1_0,child2c3));
float cost03 = halfArea3f(merge(child2c0,child2c1,child2c2,child1_0));
vfloat4 cost0 = vfloat4(cost00,cost01,cost02,cost03);
vfloat4 min0 = vreduce_min(cost0);
int pos0 = (int)__bsf(movemask(min0 == cost0));
/*! put child1_1 at each child2 position */
float cost10 = halfArea3f(merge(child1_1,child2c1,child2c2,child2c3));
float cost11 = halfArea3f(merge(child2c0,child1_1,child2c2,child2c3));
float cost12 = halfArea3f(merge(child2c0,child2c1,child1_1,child2c3));
float cost13 = halfArea3f(merge(child2c0,child2c1,child2c2,child1_1));
vfloat4 cost1 = vfloat4(cost10,cost11,cost12,cost13);
vfloat4 min1 = vreduce_min(cost1);
int pos1 = (int)__bsf(movemask(min1 == cost1));
/*! put child1_2 at each child2 position */
float cost20 = halfArea3f(merge(child1_2,child2c1,child2c2,child2c3));
float cost21 = halfArea3f(merge(child2c0,child1_2,child2c2,child2c3));
float cost22 = halfArea3f(merge(child2c0,child2c1,child1_2,child2c3));
float cost23 = halfArea3f(merge(child2c0,child2c1,child2c2,child1_2));
vfloat4 cost2 = vfloat4(cost20,cost21,cost22,cost23);
vfloat4 min2 = vreduce_min(cost2);
int pos2 = (int)__bsf(movemask(min2 == cost2));
/*! put child1_3 at each child2 position */
float cost30 = halfArea3f(merge(child1_3,child2c1,child2c2,child2c3));
float cost31 = halfArea3f(merge(child2c0,child1_3,child2c2,child2c3));
float cost32 = halfArea3f(merge(child2c0,child2c1,child1_3,child2c3));
float cost33 = halfArea3f(merge(child2c0,child2c1,child2c2,child1_3));
vfloat4 cost3 = vfloat4(cost30,cost31,cost32,cost33);
vfloat4 min3 = vreduce_min(cost3);
int pos3 = (int)__bsf(movemask(min3 == cost3));
/*! find best other child */
vfloat4 area0123 = vfloat4(extract<0>(min0),extract<0>(min1),extract<0>(min2),extract<0>(min3)) - vfloat4(childArea[c2]);
int pos[4] = { pos0,pos1,pos2,pos3 };
const size_t mbd = BVH4::maxBuildDepth;
vbool4 valid = vint4(int(depth+1))+cdepth <= vint4(mbd); // only select swaps that fulfill depth constraints
valid &= vint4(c2) != vint4(step);
if (none(valid)) continue;
size_t c1 = select_min(valid,area0123);
float area = area0123[c1];
if (c1 == c2) continue; // can happen if bounds are NANs
/*! accept a swap when it reduces cost and is not swapping a node with itself */
if (area < bestArea) {
bestArea = area;
bestChild1 = c1;
bestChild2 = c2;
bestChild2Child = pos[c1];
}
}
/*! if we did not find a swap that improves the SAH then do nothing */
if (bestChild1 == size_t(-1)) return 1+reduce_max(cdepth);
/*! perform the best found tree rotation */
Node* child2 = parent->child(bestChild2).node();
BVH4::swap(parent,bestChild1,child2,bestChild2Child);
parent->set(bestChild2,child2->bounds());
BVH4::compact(parent);
BVH4::compact(child2);
/*! This returned depth is conservative as the child that was
* pulled up in the tree could have been on the critical path. */
cdepth[bestChild1]++; // bestChild1 was pushed down one level
return 1+reduce_max(cdepth);
}
