void QuadMesh::interpolate(const RTCInterpolateArguments* const args)
{
unsigned int primID = args->primID;
float u = args->u;
float v = args->v;
RTCBufferType bufferType = args->bufferType;
unsigned int bufferSlot = args->bufferSlot;
float* P = args->P;
float* dPdu = args->dPdu;
float* dPdv = args->dPdv;
float* ddPdudu = args->ddPdudu;
float* ddPdvdv = args->ddPdvdv;
float* ddPdudv = args->ddPdudv;
unsigned int valueCount = args->valueCount;
/* calculate base pointer and stride */
assert((bufferType == RTC_BUFFER_TYPE_VERTEX && bufferSlot < numTimeSteps) ||
(bufferType == RTC_BUFFER_TYPE_VERTEX_ATTRIBUTE && bufferSlot <= vertexAttribs.size()));
const char* src = nullptr;
size_t stride = 0;
if (bufferType == RTC_BUFFER_TYPE_VERTEX_ATTRIBUTE) {
src = vertexAttribs[bufferSlot].getPtr();
stride = vertexAttribs[bufferSlot].getStride();
} else {
src = vertices[bufferSlot].getPtr();
stride = vertices[bufferSlot].getStride();
}
for (unsigned int i=0; i<valueCount; i+=4)
{
const vbool4 valid = vint4((int)i)+vint4(step) < vint4(int(valueCount));
const size_t ofs = i*sizeof(float);
const Quad& tri = quad(primID);
const vfloat4 p0 = vfloat4::loadu(valid,(float*)&src[tri.v[0]*stride+ofs]);
const vfloat4 p1 = vfloat4::loadu(valid,(float*)&src[tri.v[1]*stride+ofs]);
const vfloat4 p2 = vfloat4::loadu(valid,(float*)&src[tri.v[2]*stride+ofs]);
const vfloat4 p3 = vfloat4::loadu(valid,(float*)&src[tri.v[3]*stride+ofs]);
const vbool4 left = u+v <= 1.0f;
const vfloat4 Q0 = select(left,p0,p2);
const vfloat4 Q1 = select(left,p1,p3);
const vfloat4 Q2 = select(left,p3,p1);
const vfloat4 U = select(left,u,vfloat4(1.0f)-u);
const vfloat4 V = select(left,v,vfloat4(1.0f)-v);
const vfloat4 W = 1.0f-U-V;
if (P) {
vfloat4::storeu(valid,P+i,madd(W,Q0,madd(U,Q1,V*Q2)));
}
if (dPdu) {
assert(dPdu); vfloat4::storeu(valid,dPdu+i,select(left,Q1-Q0,Q0-Q1));
assert(dPdv); vfloat4::storeu(valid,dPdv+i,select(left,Q2-Q0,Q0-Q2));
}
if (ddPdudu) {
assert(ddPdudu); vfloat4::storeu(valid,ddPdudu+i,vfloat4(zero));
assert(ddPdvdv); vfloat4::storeu(valid,ddPdvdv+i,vfloat4(zero));
assert(ddPdudv); vfloat4::storeu(valid,ddPdudv+i,vfloat4(zero));
}
}
}
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);
}
