void BVH4HairBuilder::computePrimRefsBezierCurves(const size_t threadID, const size_t numThreads)
{
DBG(PING);
const size_t numTotalGroups = scene->size();
/* count total number of virtual objects */
const size_t numBezierCurves = numPrimitives;
const size_t startID = (threadID+0)*numBezierCurves/numThreads;
const size_t endID = (threadID+1)*numBezierCurves/numThreads;
Bezier1i *__restrict__ const bptr = (Bezier1i*)this->prims;
// === find first group containing startID ===
unsigned int g=0, numSkipped = 0;
for (; g<numTotalGroups; g++) {
if (unlikely(scene->get(g) == NULL)) continue;
if (unlikely((scene->get(g)->type != BEZIER_CURVES))) continue;
if (unlikely(!scene->get(g)->isEnabled())) continue;
BezierCurves* geom = (BezierCurves*) scene->getBezierCurves(g);
const size_t numPrims = geom->numCurves;
if (numSkipped + numPrims > startID) break;
numSkipped += numPrims;
}
/* start with first group containing startID */
mic_f bounds_scene_min((float)pos_inf);
mic_f bounds_scene_max((float)neg_inf);
mic_f bounds_centroid_min((float)pos_inf);
mic_f bounds_centroid_max((float)neg_inf);
unsigned int num = 0;
unsigned int currentID = startID;
unsigned int offset = startID - numSkipped;
for (; g<numTotalGroups; g++)
{
if (unlikely(scene->get(g) == NULL)) continue;
if (unlikely((scene->get(g)->type != BEZIER_CURVES))) continue;
if (unlikely(!scene->get(g)->isEnabled())) continue;
BezierCurves* geom = (BezierCurves*) scene->getBezierCurves(g);
size_t N = geom->numCurves;
for (unsigned int i=offset; i<N && currentID < endID; i++, currentID++)
{
const mic2f b2 = geom->bounds_mic2f(i);
const mic_f bmin = b2.x;
const mic_f bmax = b2.y;
bounds_scene_min = min(bounds_scene_min,bmin);
bounds_scene_max = max(bounds_scene_max,bmax);
const mic_f centroid2 = bmin+bmax;
bounds_centroid_min = min(bounds_centroid_min,centroid2);
bounds_centroid_max = max(bounds_centroid_max,centroid2);
bptr[currentID].p = geom->fristVertexPtr(i);
bptr[currentID].geomID = g;
bptr[currentID].primID = i;
}
if (currentID == endID) break;
offset = 0;
}
/* update global bounds */
Centroid_Scene_AABB bounds;
store4f(&bounds.centroid2.lower,bounds_centroid_min);
store4f(&bounds.centroid2.upper,bounds_centroid_max);
store4f(&bounds.geometry.lower,bounds_scene_min);
store4f(&bounds.geometry.upper,bounds_scene_max);
global_bounds.extend_atomic(bounds);
}
void BVH4mbBuilder::computePrimRefsTrianglesMB(const size_t threadID, const size_t numThreads)
{
DBG(PING);
const size_t numGroups = scene->size();
const size_t startID = (threadID+0)*numPrimitives/numThreads;
const size_t endID = (threadID+1)*numPrimitives/numThreads;
PrimRef *__restrict__ const prims = this->prims;
// === find first group containing startID ===
unsigned int g=0, numSkipped = 0;
for (; g<numGroups; g++) {
if (unlikely(scene->get(g) == nullptr)) continue;
if (unlikely(scene->get(g)->type != Geometry::TRIANGLE_MESH)) continue;
const TriangleMesh* __restrict__ const mesh = scene->getTriangleMesh(g);
if (unlikely(!mesh->isEnabled())) continue;
if (unlikely(mesh->numTimeSteps == 1)) continue;
const size_t numTriangles = mesh->size();
if (numSkipped + numTriangles > startID) break;
numSkipped += numTriangles;
}
// === start with first group containing startID ===
mic_f bounds_scene_min((float)pos_inf);
mic_f bounds_scene_max((float)neg_inf);
mic_f bounds_centroid_min((float)pos_inf);
mic_f bounds_centroid_max((float)neg_inf);
unsigned int num = 0;
unsigned int currentID = startID;
unsigned int offset = startID - numSkipped;
__aligned(64) PrimRef local_prims[2];
size_t numLocalPrims = 0;
PrimRef *__restrict__ dest = &prims[currentID];
for (; g<numGroups; g++)
{
if (unlikely(scene->get(g) == nullptr)) continue;
if (unlikely(scene->get(g)->type != Geometry::TRIANGLE_MESH)) continue;
const TriangleMesh* __restrict__ const mesh = scene->getTriangleMesh(g);
if (unlikely(!mesh->isEnabled())) continue;
if (unlikely(mesh->numTimeSteps == 1)) continue;
for (unsigned int i=offset; i<mesh->size() && currentID < endID; i++, currentID++)
{
const TriangleMesh::Triangle& tri = mesh->triangle(i);
prefetch<PFHINT_L2>(&tri + L2_PREFETCH_ITEMS);
prefetch<PFHINT_L1>(&tri + L1_PREFETCH_ITEMS);
const mic3f v = mesh->getTriangleVertices<PFHINT_L2>(tri);
mic_f bmin = min(min(v[0],v[1]),v[2]);
mic_f bmax = max(max(v[0],v[1]),v[2]);
bounds_scene_min = min(bounds_scene_min,bmin);
bounds_scene_max = max(bounds_scene_max,bmax);
const mic_f centroid2 = bmin+bmax;
bounds_centroid_min = min(bounds_centroid_min,centroid2);
bounds_centroid_max = max(bounds_centroid_max,centroid2);
store4f(&local_prims[numLocalPrims].lower,bmin);
store4f(&local_prims[numLocalPrims].upper,bmax);
local_prims[numLocalPrims].lower.a = g;
local_prims[numLocalPrims].upper.a = i;
numLocalPrims++;
if (unlikely(((size_t)dest % 64) != 0) && numLocalPrims == 1)
{
*dest = local_prims[0];
dest++;
numLocalPrims--;
}
else
{
const mic_f twoAABBs = load16f(local_prims);
if (numLocalPrims == 2)
{
numLocalPrims = 0;
store16f_ngo(dest,twoAABBs);
dest+=2;
}
}
}
if (currentID == endID) break;
offset = 0;
}
/* is there anything left in the local queue? */
if (numLocalPrims % 2 != 0)
*dest = local_prims[0];
/* update global bounds */
Centroid_Scene_AABB bounds;
store4f(&bounds.centroid2.lower,bounds_centroid_min);
store4f(&bounds.centroid2.upper,bounds_centroid_max);
store4f(&bounds.geometry.lower,bounds_scene_min);
store4f(&bounds.geometry.upper,bounds_scene_max);
global_bounds.extend_atomic(bounds);
}