int
main (int argc, char *argv[])
{
MINT *a, *b, *c, *d;
short h;
mp_set_memory_functions (NULL, NULL, NULL);
a = itom (123);
b = xtom ("DEADBEEF");
c = itom (0);
d = itom (0);
move (a, b);
madd (a, b, c);
msub (a, b, c);
mult (a, b, c);
mdiv (b, a, c, d);
sdiv (b, 2, c, &h);
msqrt (a, c, d);
pow (b, a, a, c);
rpow (a, 3, c);
gcd (a, b, c);
mcmp (a, b);
if (argc > 1)
{
min (c);
mout (a);
}
mtox (b);
mfree(a);
exit (0);
}
void
qsub(void)
{
unsigned int *a, *ab, *b, *ba, *c;
save();
p2 = pop();
p1 = pop();
ab = mmul(p1->u.q.a, p2->u.q.b);
ba = mmul(p1->u.q.b, p2->u.q.a);
a = msub(ab, ba);
mfree(ab);
mfree(ba);
// zero?
if (MZERO(a)) {
mfree(a);
push(zero);
restore();
return;
}
b = mmul(p1->u.q.b, p2->u.q.b);
c = mgcd(a, b);
MSIGN(c) = MSIGN(b);
p1 = alloc();
p1->k = NUM;
p1->u.q.a = mdiv(a, c);
p1->u.q.b = mdiv(b, c);
mfree(a);
mfree(b);
mfree(c);
push(p1);
restore();
}
FN minvert(MINT *a, MINT *b, MINT *c)
{ MINT x, y, z, w, Anew, Aold;
int i = 0;
static MINT one;
static int oneinit = 1;
if (oneinit) {
oneinit = 0;
MSET(1,&one);
}
MINIT(&x);
MINIT(&y);
MINIT(&z);
MINIT(&w);
MINIT(&Aold);
MSET (1,&Anew);
mcopy(b, &x);
mcopy(a, &y);
/*
* Loop invariant:
*
* y = -1^i * Anew * a mod b
*/
while(mtest(&y) != 0)
{ mdiv(&x, &y, &w, &z);
mcopy(&Anew, &x);
mmult(&w, &Anew, &Anew);
madd(&Anew, &Aold, &Anew);
mmove(&x, &Aold);
mmove(&y, &x);
mmove(&z, &y);
i++;
}
if (mcmp(&one,&x)) {
mcopy(&one,c);
} else {
mmove(&Aold, c);
if( (i&01) == 0) msub(b, c, c);
}
MFREE(&x);
MFREE(&y);
MFREE(&z);
MFREE(&w);
MFREE(&Aold);
MFREE(&Anew);
}
void newton_fd(double (*funcpt)(double *,int),double *xi,int N,double *x) {
int it,i;
double *jac,*hess,*L,*xf;
double stop,result;
jac = (double*) malloc(sizeof(double) *N);
xf = (double*) malloc(sizeof(double) *N);
hess = (double*) malloc(sizeof(double) *N * N);
L = (double*) malloc(sizeof(double) *N * N);
it = 0;
stop = 1.0;
while (stop > 1e-05 && it < 100) {
//result = funcpt(xi,N,jac,hess);
jacobian_fd(funcpt,xi,N,jac);
hessian_fd(funcpt,xi,N,hess);
//mdisplay(jac,1,N);
//mdisplay(hess,N,N);
modelhess2(hess,N,L);
//mdisplay(L,N,N);
scale(jac,1,N,-1.0);
linsolve_lower(L,N,jac,x);
madd(x,xi,x,1,N);
msub(x,xi,xf,1,N);
stop = array_max_abs(xf,N);
for(i=0;i < N;++i) {
xi[i] = x[i];
}
it++;
//printf("result %lf \n",result);
printf("Values obtained %lf , %lf after %d Iterations \n",x[0],x[1],it);
}
free(jac);
free(hess);
free(xf);
free(L);
}
msqrt(MINT *a, MINT *b, MINT *r)
{ MINT x,y,z;
register alen,j;
MINIT(&x); MINIT(&y); MINIT(&z);
alen = a->len;
if (alen<0) mpfatal("msqrt: neg arg");
if (alen==0) {
mset(0,b);
mset(0,r);
return(0);
}
if(alen & 01) x.len = (1+alen)/2;
else x.len = 1 + alen/2;
valloc(x.val,x.len);
for (j=x.len; (--j)>=0;) x.val[j]=0;
if (alen & 01) x.val[x.len-1]=0400;
else x.val[x.len-1]=1;
for (;;) {
mdiv(a,&x,&y,&z);
madd(&x,&y,&y);
mshiftr(&y,1);
if (mcmp(&x,&y) <= 0) break;
mmove(&y,&x);
}
mcopy(&x,&y);
mmult(&x,&x,&x);
msub(a,&x,r);
MFREE(&x);
MMOVEFREE(&y,b);
MFREE(&z);
return(r->len);
}
void BVH4Intersector4Chunk<PrimitiveIntersector4>::intersect(sseb* valid_i, BVH4* bvh, Ray4& ray)
{
/* load ray */
const sseb valid0 = *valid_i;
const sse3f rdir = rcp_safe(ray.dir);
const sse3f org(ray.org), org_rdir = org * rdir;
ssef ray_tnear = select(valid0,ray.tnear,ssef(pos_inf));
ssef ray_tfar = select(valid0,ray.tfar ,ssef(neg_inf));
const ssef inf = ssef(pos_inf);
Precalculations pre(valid0,ray);
/* allocate stack and push root node */
ssef stack_near[stackSize];
NodeRef stack_node[stackSize];
stack_node[0] = BVH4::invalidNode;
stack_near[0] = inf;
stack_node[1] = bvh->root;
stack_near[1] = ray_tnear;
NodeRef* stackEnd = stack_node+stackSize;
NodeRef* __restrict__ sptr_node = stack_node + 2;
ssef* __restrict__ sptr_near = stack_near + 2;
while (1)
{
/* pop next node from stack */
assert(sptr_node > stack_node);
sptr_node--;
sptr_near--;
NodeRef curNode = *sptr_node;
if (unlikely(curNode == BVH4::invalidNode)) {
assert(sptr_node == stack_node);
break;
}
/* cull node if behind closest hit point */
ssef curDist = *sptr_near;
if (unlikely(none(ray_tfar > curDist)))
continue;
while (1)
{
/* test if this is a leaf node */
if (unlikely(curNode.isLeaf()))
break;
const sseb valid_node = ray_tfar > curDist;
STAT3(normal.trav_nodes,1,popcnt(valid_node),4);
const Node* __restrict__ const node = curNode.node();
/* pop of next node */
assert(sptr_node > stack_node);
sptr_node--;
sptr_near--;
curNode = *sptr_node;
curDist = *sptr_near;
#pragma unroll(4)
for (unsigned i=0; i<BVH4::N; i++)
{
const NodeRef child = node->children[i];
if (unlikely(child == BVH4::emptyNode)) break;
#if defined(__AVX2__)
const ssef lclipMinX = msub(node->lower_x[i],rdir.x,org_rdir.x);
const ssef lclipMinY = msub(node->lower_y[i],rdir.y,org_rdir.y);
const ssef lclipMinZ = msub(node->lower_z[i],rdir.z,org_rdir.z);
const ssef lclipMaxX = msub(node->upper_x[i],rdir.x,org_rdir.x);
const ssef lclipMaxY = msub(node->upper_y[i],rdir.y,org_rdir.y);
const ssef lclipMaxZ = msub(node->upper_z[i],rdir.z,org_rdir.z);
#else
const ssef lclipMinX = (node->lower_x[i] - org.x) * rdir.x;
const ssef lclipMinY = (node->lower_y[i] - org.y) * rdir.y;
const ssef lclipMinZ = (node->lower_z[i] - org.z) * rdir.z;
const ssef lclipMaxX = (node->upper_x[i] - org.x) * rdir.x;
const ssef lclipMaxY = (node->upper_y[i] - org.y) * rdir.y;
const ssef lclipMaxZ = (node->upper_z[i] - org.z) * rdir.z;
#endif
#if defined(__SSE4_1__)
const ssef lnearP = maxi(maxi(mini(lclipMinX, lclipMaxX), mini(lclipMinY, lclipMaxY)), mini(lclipMinZ, lclipMaxZ));
const ssef lfarP = mini(mini(maxi(lclipMinX, lclipMaxX), maxi(lclipMinY, lclipMaxY)), maxi(lclipMinZ, lclipMaxZ));
const sseb lhit = maxi(lnearP,ray_tnear) <= mini(lfarP,ray_tfar);
#else
const ssef lnearP = max(max(min(lclipMinX, lclipMaxX), min(lclipMinY, lclipMaxY)), min(lclipMinZ, lclipMaxZ));
const ssef lfarP = min(min(max(lclipMinX, lclipMaxX), max(lclipMinY, lclipMaxY)), max(lclipMinZ, lclipMaxZ));
const sseb lhit = max(lnearP,ray_tnear) <= min(lfarP,ray_tfar);
#endif
/* if we hit the child we choose to continue with that child if it
is closer than the current next child, or we push it onto the stack */
if (likely(any(lhit)))
{
assert(sptr_node < stackEnd);
const ssef childDist = select(lhit,lnearP,inf);
const NodeRef child = node->children[i];
assert(child != BVH4::emptyNode);
sptr_node++;
sptr_near++;
/* push cur node onto stack and continue with hit child */
if (any(childDist < curDist))
{
*(sptr_node-1) = curNode;
*(sptr_near-1) = curDist;
curDist = childDist;
curNode = child;
}
/* push hit child onto stack */
else {
*(sptr_node-1) = child;
*(sptr_near-1) = childDist;
}
}
}
}
/* return if stack is empty */
if (unlikely(curNode == BVH4::invalidNode)) {
assert(sptr_node == stack_node);
break;
}
/* intersect leaf */
const sseb valid_leaf = ray_tfar > curDist;
STAT3(normal.trav_leaves,1,popcnt(valid_leaf),4);
size_t items; const Primitive* prim = (Primitive*) curNode.leaf(items);
PrimitiveIntersector4::intersect(valid_leaf,pre,ray,prim,items,bvh->geometry);
ray_tfar = select(valid_leaf,ray.tfar,ray_tfar);
}
AVX_ZERO_UPPER();
}
static void test_heart() {
printf("run test heart\n");
assert(madd(5, 6) == 11);
assert(msub(1024, 512) == 512);
}
void BVH8Intersector8Hybrid<PrimitiveIntersector8>::occluded(bool8* valid_i, BVH8* bvh, Ray8& ray)
{
/* load ray */
const bool8 valid = *valid_i;
bool8 terminated = !valid;
Vec3f8 ray_org = ray.org, ray_dir = ray.dir;
float8 ray_tnear = ray.tnear, ray_tfar = ray.tfar;
const Vec3f8 rdir = rcp_safe(ray_dir);
const Vec3f8 org(ray_org), org_rdir = org * rdir;
ray_tnear = select(valid,ray_tnear,float8(pos_inf));
ray_tfar = select(valid,ray_tfar ,float8(neg_inf));
const float8 inf = float8(pos_inf);
Precalculations pre(valid,ray);
/* compute near/far per ray */
Vec3i8 nearXYZ;
nearXYZ.x = select(rdir.x >= 0.0f,int8(0*(int)sizeof(float8)),int8(1*(int)sizeof(float8)));
nearXYZ.y = select(rdir.y >= 0.0f,int8(2*(int)sizeof(float8)),int8(3*(int)sizeof(float8)));
nearXYZ.z = select(rdir.z >= 0.0f,int8(4*(int)sizeof(float8)),int8(5*(int)sizeof(float8)));
/* allocate stack and push root node */
float8 stack_near[stackSizeChunk];
NodeRef stack_node[stackSizeChunk];
stack_node[0] = BVH8::invalidNode;
stack_near[0] = inf;
stack_node[1] = bvh->root;
stack_near[1] = ray_tnear;
NodeRef* stackEnd = stack_node+stackSizeChunk;
NodeRef* __restrict__ sptr_node = stack_node + 2;
float8* __restrict__ sptr_near = stack_near + 2;
while (1)
{
/* pop next node from stack */
assert(sptr_node > stack_node);
sptr_node--;
sptr_near--;
NodeRef cur = *sptr_node;
if (unlikely(cur == BVH8::invalidNode)) {
assert(sptr_node == stack_node);
break;
}
/* cull node if behind closest hit point */
float8 curDist = *sptr_near;
const bool8 active = curDist < ray_tfar;
if (unlikely(none(active)))
continue;
/* switch to single ray traversal */
#if !defined(__WIN32__) || defined(__X86_64__)
size_t bits = movemask(active);
if (unlikely(__popcnt(bits) <= SWITCH_THRESHOLD)) {
for (size_t i=__bsf(bits); bits!=0; bits=__btc(bits,i), i=__bsf(bits)) {
if (occluded1(bvh,cur,i,pre,ray,ray_org,ray_dir,rdir,ray_tnear,ray_tfar,nearXYZ))
terminated[i] = -1;
}
if (all(terminated)) break;
ray_tfar = select(terminated,float8(neg_inf),ray_tfar);
continue;
}
#endif
while (1)
{
/* test if this is a leaf node */
if (unlikely(cur.isLeaf()))
break;
const bool8 valid_node = ray_tfar > curDist;
STAT3(shadow.trav_nodes,1,popcnt(valid_node),8);
const Node* __restrict__ const node = (Node*)cur.node();
/* pop of next node */
assert(sptr_node > stack_node);
sptr_node--;
sptr_near--;
cur = *sptr_node;
curDist = *sptr_near;
for (unsigned i=0; i<BVH8::N; i++)
{
const NodeRef child = node->children[i];
if (unlikely(child == BVH8::emptyNode)) break;
#if defined(__AVX2__)
const float8 lclipMinX = msub(node->lower_x[i],rdir.x,org_rdir.x);
const float8 lclipMinY = msub(node->lower_y[i],rdir.y,org_rdir.y);
const float8 lclipMinZ = msub(node->lower_z[i],rdir.z,org_rdir.z);
const float8 lclipMaxX = msub(node->upper_x[i],rdir.x,org_rdir.x);
const float8 lclipMaxY = msub(node->upper_y[i],rdir.y,org_rdir.y);
const float8 lclipMaxZ = msub(node->upper_z[i],rdir.z,org_rdir.z);
const float8 lnearP = maxi(maxi(mini(lclipMinX, lclipMaxX), mini(lclipMinY, lclipMaxY)), mini(lclipMinZ, lclipMaxZ));
const float8 lfarP = mini(mini(maxi(lclipMinX, lclipMaxX), maxi(lclipMinY, lclipMaxY)), maxi(lclipMinZ, lclipMaxZ));
const bool8 lhit = maxi(lnearP,ray_tnear) <= mini(lfarP,ray_tfar);
#else
const float8 lclipMinX = (node->lower_x[i] - org.x) * rdir.x;
const float8 lclipMinY = (node->lower_y[i] - org.y) * rdir.y;
const float8 lclipMinZ = (node->lower_z[i] - org.z) * rdir.z;
const float8 lclipMaxX = (node->upper_x[i] - org.x) * rdir.x;
const float8 lclipMaxY = (node->upper_y[i] - org.y) * rdir.y;
const float8 lclipMaxZ = (node->upper_z[i] - org.z) * rdir.z;
const float8 lnearP = max(max(min(lclipMinX, lclipMaxX), min(lclipMinY, lclipMaxY)), min(lclipMinZ, lclipMaxZ));
const float8 lfarP = min(min(max(lclipMinX, lclipMaxX), max(lclipMinY, lclipMaxY)), max(lclipMinZ, lclipMaxZ));
const bool8 lhit = max(lnearP,ray_tnear) <= min(lfarP,ray_tfar);
#endif
/* if we hit the child we choose to continue with that child if it
is closer than the current next child, or we push it onto the stack */
if (likely(any(lhit)))
{
assert(sptr_node < stackEnd);
assert(child != BVH8::emptyNode);
const float8 childDist = select(lhit,lnearP,inf);
sptr_node++;
sptr_near++;
/* push cur node onto stack and continue with hit child */
if (any(childDist < curDist))
{
*(sptr_node-1) = cur;
*(sptr_near-1) = curDist;
curDist = childDist;
cur = child;
}
/* push hit child onto stack */
else {
*(sptr_node-1) = child;
*(sptr_near-1) = childDist;
}
}
}
}
/* return if stack is empty */
if (unlikely(cur == BVH8::invalidNode)) {
assert(sptr_node == stack_node);
break;
}
/* intersect leaf */
assert(cur != BVH8::emptyNode);
const bool8 valid_leaf = ray_tfar > curDist;
STAT3(shadow.trav_leaves,1,popcnt(valid_leaf),8);
size_t items; const Triangle* prim = (Triangle*) cur.leaf(items);
terminated |= PrimitiveIntersector8::occluded(!terminated,pre,ray,prim,items,bvh->scene);
if (all(terminated)) break;
ray_tfar = select(terminated,float8(neg_inf),ray_tfar);
}
store8i(valid & terminated,&ray.geomID,0);
AVX_ZERO_UPPER();
}
void BVH4iIntersector1<TriangleIntersector>::intersect(const BVH4iIntersector1* This, Ray& ray)
{
AVX_ZERO_UPPER();
STAT3(normal.travs,1,1,1);
/*! stack state */
const BVH4i* bvh = This->bvh;
StackItem stack[1+3*BVH4i::maxDepth]; //!< stack of nodes
StackItem* stackPtr = stack+1; //!< current stack pointer
stack[0].ptr = bvh->root;
stack[0].dist = neg_inf;
/*! offsets to select the side that becomes the lower or upper bound */
const size_t nearX = ray.dir.x >= 0.0f ? 0*sizeof(ssef_m) : 1*sizeof(ssef_m);
const size_t nearY = ray.dir.y >= 0.0f ? 2*sizeof(ssef_m) : 3*sizeof(ssef_m);
const size_t nearZ = ray.dir.z >= 0.0f ? 4*sizeof(ssef_m) : 5*sizeof(ssef_m);
/*! load the ray into SIMD registers */
const sse3f norg(-ray.org.x,-ray.org.y,-ray.org.z);
const Vector3f ray_rdir = rcp_safe(ray.dir);
const sse3f rdir(ray_rdir.x,ray_rdir.y,ray_rdir.z);
const Vector3f ray_org_rdir = ray.org*ray_rdir;
const sse3f org_rdir(ray_org_rdir.x,ray_org_rdir.y,ray_org_rdir.z);
const ssef rayNear(ray.tnear);
ssef rayFar(ray.tfar);
const void* nodePtr = bvh->nodePtr();
const void* triPtr = bvh->triPtr();
/* pop loop */
while (true) pop:
{
/*! pop next node */
if (unlikely(stackPtr == stack)) break;
stackPtr--;
NodeRef cur = NodeRef(stackPtr->ptr);
/*! if popped node is too far, pop next one */
if (unlikely(stackPtr->dist > ray.tfar))
continue;
/* downtraversal loop */
while (true)
{
/*! stop if we found a leaf */
if (unlikely(cur.isLeaf())) break;
STAT3(normal.trav_nodes,1,1,1);
/*! single ray intersection with 4 boxes */
const Node* node = cur.node(nodePtr);
const size_t farX = nearX ^ 16, farY = nearY ^ 16, farZ = nearZ ^ 16;
#if defined (__AVX2__)
const ssef tNearX = msub(ssef((const char*)nodePtr+(size_t)cur+nearX), rdir.x, org_rdir.x);
const ssef tNearY = msub(ssef((const char*)nodePtr+(size_t)cur+nearY), rdir.y, org_rdir.y);
const ssef tNearZ = msub(ssef((const char*)nodePtr+(size_t)cur+nearZ), rdir.z, org_rdir.z);
const ssef tFarX = msub(ssef((const char*)nodePtr+(size_t)cur+farX ), rdir.x, org_rdir.x);
const ssef tFarY = msub(ssef((const char*)nodePtr+(size_t)cur+farY ), rdir.y, org_rdir.y);
const ssef tFarZ = msub(ssef((const char*)nodePtr+(size_t)cur+farZ ), rdir.z, org_rdir.z);
#else
const ssef tNearX = (norg.x + ssef((const char*)nodePtr+(size_t)cur+nearX)) * rdir.x;
const ssef tNearY = (norg.y + ssef((const char*)nodePtr+(size_t)cur+nearY)) * rdir.y;
const ssef tNearZ = (norg.z + ssef((const char*)nodePtr+(size_t)cur+nearZ)) * rdir.z;
const ssef tFarX = (norg.x + ssef((const char*)nodePtr+(size_t)cur+farX )) * rdir.x;
const ssef tFarY = (norg.y + ssef((const char*)nodePtr+(size_t)cur+farY )) * rdir.y;
const ssef tFarZ = (norg.z + ssef((const char*)nodePtr+(size_t)cur+farZ )) * rdir.z;
#endif
const ssef tNear = max(tNearX,tNearY,tNearZ,rayNear);
const ssef tFar = min(tFarX ,tFarY ,tFarZ ,rayFar);
size_t mask = movemask(tNear <= tFar);
/*! if no child is hit, pop next node */
if (unlikely(mask == 0))
goto pop;
/*! one child is hit, continue with that child */
size_t r = __bsf(mask); mask = __btc(mask,r);
if (likely(mask == 0)) {
cur = node->child(r);
continue;
}
/*! two children are hit, push far child, and continue with closer child */
NodeRef c0 = node->child(r); const float d0 = tNear[r];
r = __bsf(mask); mask = __btc(mask,r);
NodeRef c1 = node->child(r); const float d1 = tNear[r];
if (likely(mask == 0)) {
if (d0 < d1) { stackPtr->ptr = c1; stackPtr->dist = d1; stackPtr++; cur = c0; continue; }
else { stackPtr->ptr = c0; stackPtr->dist = d0; stackPtr++; cur = c1; continue; }
}
/*! Here starts the slow path for 3 or 4 hit children. We push
* all nodes onto the stack to sort them there. */
stackPtr->ptr = c0; stackPtr->dist = d0; stackPtr++;
stackPtr->ptr = c1; stackPtr->dist = d1; stackPtr++;
/*! three children are hit, push all onto stack and sort 3 stack items, continue with closest child */
r = __bsf(mask); mask = __btc(mask,r);
NodeRef c = node->child(r); float d = tNear[r]; stackPtr->ptr = c; stackPtr->dist = d; stackPtr++;
if (likely(mask == 0)) {
sort(stackPtr[-1],stackPtr[-2],stackPtr[-3]);
cur = (NodeRef) stackPtr[-1].ptr; stackPtr--;
continue;
}
/*! four children are hit, push all onto stack and sort 4 stack items, continue with closest child */
r = __bsf(mask); mask = __btc(mask,r);
c = node->child(r); d = tNear[r]; stackPtr->ptr = c; stackPtr->dist = d; stackPtr++;
sort(stackPtr[-1],stackPtr[-2],stackPtr[-3],stackPtr[-4]);
cur = (NodeRef) stackPtr[-1].ptr; stackPtr--;
}
/*! this is a leaf node */
STAT3(normal.trav_leaves,1,1,1);
size_t num; Triangle* tri = (Triangle*) cur.leaf(triPtr,num);
for (size_t i=0; i<num; i++)
TriangleIntersector::intersect(ray,tri[i],bvh->vertices);
rayFar = ray.tfar;
}
AVX_ZERO_UPPER();
}
void BVH8Intersector8Chunk<PrimitiveIntersector8>::intersect(avxb* valid_i, BVH8* bvh, Ray8& ray)
{
#if defined(__AVX__)
/* load ray */
const avxb valid0 = *valid_i;
const avx3f rdir = rcp_safe(ray.dir);
const avx3f org_rdir = ray.org * rdir;
avxf ray_tnear = select(valid0,ray.tnear,pos_inf);
avxf ray_tfar = select(valid0,ray.tfar ,neg_inf);
const avxf inf = avxf(pos_inf);
Precalculations pre(valid0,ray);
/* allocate stack and push root node */
avxf stack_near[3*BVH8::maxDepth+1];
NodeRef stack_node[3*BVH8::maxDepth+1];
stack_node[0] = BVH8::invalidNode;
stack_near[0] = inf;
stack_node[1] = bvh->root;
stack_near[1] = ray_tnear;
NodeRef* __restrict__ sptr_node = stack_node + 2;
avxf* __restrict__ sptr_near = stack_near + 2;
while (1)
{
/* pop next node from stack */
sptr_node--;
sptr_near--;
NodeRef cur = *sptr_node;
if (unlikely(cur == BVH8::invalidNode))
break;
/* cull node if behind closest hit point */
avxf curDist = *sptr_near;
if (unlikely(none(ray_tfar > curDist)))
continue;
while (1)
{
/* test if this is a leaf node */
if (unlikely(cur.isLeaf()))
break;
const avxb valid_node = ray_tfar > curDist;
STAT3(normal.trav_nodes,1,popcnt(valid_node),8);
const Node* __restrict__ const node = (BVH8::Node*)cur.node();
/* pop of next node */
sptr_node--;
sptr_near--;
cur = *sptr_node; // FIXME: this trick creates issues with stack depth
curDist = *sptr_near;
for (unsigned i=0; i<BVH8::N; i++)
{
const NodeRef child = node->children[i];
if (unlikely(child == BVH8::emptyNode)) break;
#if defined(__AVX2__)
const avxf lclipMinX = msub(node->lower_x[i],rdir.x,org_rdir.x);
const avxf lclipMinY = msub(node->lower_y[i],rdir.y,org_rdir.y);
const avxf lclipMinZ = msub(node->lower_z[i],rdir.z,org_rdir.z);
const avxf lclipMaxX = msub(node->upper_x[i],rdir.x,org_rdir.x);
const avxf lclipMaxY = msub(node->upper_y[i],rdir.y,org_rdir.y);
const avxf lclipMaxZ = msub(node->upper_z[i],rdir.z,org_rdir.z);
const avxf lnearP = maxi(maxi(mini(lclipMinX, lclipMaxX), mini(lclipMinY, lclipMaxY)), mini(lclipMinZ, lclipMaxZ));
const avxf lfarP = mini(mini(maxi(lclipMinX, lclipMaxX), maxi(lclipMinY, lclipMaxY)), maxi(lclipMinZ, lclipMaxZ));
const avxb lhit = maxi(lnearP,ray_tnear) <= mini(lfarP,ray_tfar);
#else
const avxf lclipMinX = node->lower_x[i] * rdir.x - org_rdir.x;
const avxf lclipMinY = node->lower_y[i] * rdir.y - org_rdir.y;
const avxf lclipMinZ = node->lower_z[i] * rdir.z - org_rdir.z;
const avxf lclipMaxX = node->upper_x[i] * rdir.x - org_rdir.x;
const avxf lclipMaxY = node->upper_y[i] * rdir.y - org_rdir.y;
const avxf lclipMaxZ = node->upper_z[i] * rdir.z - org_rdir.z;
const avxf lnearP = max(max(min(lclipMinX, lclipMaxX), min(lclipMinY, lclipMaxY)), min(lclipMinZ, lclipMaxZ));
const avxf lfarP = min(min(max(lclipMinX, lclipMaxX), max(lclipMinY, lclipMaxY)), max(lclipMinZ, lclipMaxZ));
const avxb lhit = max(lnearP,ray_tnear) <= min(lfarP,ray_tfar);
#endif
/* if we hit the child we choose to continue with that child if it
is closer than the current next child, or we push it onto the stack */
if (likely(any(lhit)))
{
const avxf childDist = select(lhit,lnearP,inf);
const NodeRef child = node->children[i];
/* push cur node onto stack and continue with hit child */
if (any(childDist < curDist))
{
*sptr_node = cur;
*sptr_near = curDist;
sptr_node++;
sptr_near++;
curDist = childDist;
cur = child;
}
/* push hit child onto stack*/
else {
*sptr_node = child;
*sptr_near = childDist;
sptr_node++;
sptr_near++;
}
assert(sptr_node - stack_node < BVH8::maxDepth);
}
}
}
/* return if stack is empty */
if (unlikely(cur == BVH8::invalidNode))
break;
/* intersect leaf */
assert(cur != BVH8::emptyNode);
const avxb valid_leaf = ray_tfar > curDist;
STAT3(normal.trav_leaves,1,popcnt(valid_leaf),8);
size_t items; const Triangle* tri = (Triangle*) cur.leaf(items);
PrimitiveIntersector8::intersect(valid_leaf,pre,ray,tri,items,bvh->geometry);
ray_tfar = select(valid_leaf,ray.tfar,ray_tfar);
}
AVX_ZERO_UPPER();
#endif
}
void BVH4iIntersector4Chunk<TriangleIntersector4>::occluded(sseb* valid_i, BVH4i* bvh, Ray4& ray)
{
/* load node and primitive array */
const Node * __restrict__ nodes = (Node *)bvh->nodePtr();
const Triangle * __restrict__ accel = (Triangle*)bvh->triPtr();
/* load ray */
const sseb valid = *valid_i;
sseb terminated = !valid;
const sse3f rdir = rcp_safe(ray.dir);
const sse3f org_rdir = ray.org * rdir;
ssef ray_tnear = select(valid,ray.tnear,pos_inf);
ssef ray_tfar = select(valid,ray.tfar ,neg_inf);
const ssef inf = ssef(pos_inf);
/* allocate stack and push root node */
ssef stack_near[3*BVH4i::maxDepth+1];
NodeRef stack_node[3*BVH4i::maxDepth+1];
stack_node[0] = BVH4i::invalidNode;
stack_near[0] = inf;
stack_node[1] = bvh->root;
stack_near[1] = ray_tnear;
NodeRef* __restrict__ sptr_node = stack_node + 2;
ssef* __restrict__ sptr_near = stack_near + 2;
while (1)
{
/* pop next node from stack */
sptr_node--;
sptr_near--;
NodeRef curNode = *sptr_node;
if (unlikely(curNode == BVH4i::invalidNode))
break;
/* cull node if behind closest hit point */
ssef curDist = *sptr_near;
if (unlikely(none(ray_tfar > curDist)))
continue;
while (1)
{
/* test if this is a leaf node */
if (unlikely(curNode.isLeaf()))
break;
const sseb valid_node = ray_tfar > curDist;
STAT3(shadow.trav_nodes,1,popcnt(valid_node),4);
const Node* __restrict__ const node = curNode.node(nodes);
/* pop of next node */
sptr_node--;
sptr_near--;
curNode = *sptr_node; // FIXME: this trick creates issues with stack depth
curDist = *sptr_near;
#pragma unroll(4)
for (unsigned i=0; i<4; i++)
{
const NodeRef child = node->children[i];
if (unlikely(child == BVH4i::emptyNode)) break;
#if defined(__AVX2__)
const ssef lclipMinX = msub(node->lower_x[i],rdir.x,org_rdir.x);
const ssef lclipMinY = msub(node->lower_y[i],rdir.y,org_rdir.y);
const ssef lclipMinZ = msub(node->lower_z[i],rdir.z,org_rdir.z);
const ssef lclipMaxX = msub(node->upper_x[i],rdir.x,org_rdir.x);
const ssef lclipMaxY = msub(node->upper_y[i],rdir.y,org_rdir.y);
const ssef lclipMaxZ = msub(node->upper_z[i],rdir.z,org_rdir.z);
const ssef lnearP = maxi(maxi(mini(lclipMinX, lclipMaxX), mini(lclipMinY, lclipMaxY)), mini(lclipMinZ, lclipMaxZ));
const ssef lfarP = mini(mini(maxi(lclipMinX, lclipMaxX), maxi(lclipMinY, lclipMaxY)), maxi(lclipMinZ, lclipMaxZ));
const sseb lhit = maxi(lnearP,ray_tnear) <= mini(lfarP,ray_tfar);
#else
const ssef lclipMinX = node->lower_x[i] * rdir.x - org_rdir.x;
const ssef lclipMinY = node->lower_y[i] * rdir.y - org_rdir.y;
const ssef lclipMinZ = node->lower_z[i] * rdir.z - org_rdir.z;
const ssef lclipMaxX = node->upper_x[i] * rdir.x - org_rdir.x;
const ssef lclipMaxY = node->upper_y[i] * rdir.y - org_rdir.y;
const ssef lclipMaxZ = node->upper_z[i] * rdir.z - org_rdir.z;
const ssef lnearP = max(max(min(lclipMinX, lclipMaxX), min(lclipMinY, lclipMaxY)), min(lclipMinZ, lclipMaxZ));
const ssef lfarP = min(min(max(lclipMinX, lclipMaxX), max(lclipMinY, lclipMaxY)), max(lclipMinZ, lclipMaxZ));
const sseb lhit = max(lnearP,ray_tnear) <= min(lfarP,ray_tfar);
#endif
/* if we hit the child we choose to continue with that child if it
is closer than the current next child, or we push it onto the stack */
if (likely(any(lhit)))
{
const ssef childDist = select(lhit,lnearP,inf);
sptr_node++;
sptr_near++;
/* push cur node onto stack and continue with hit child */
if (any(childDist < curDist))
{
*(sptr_node-1) = curNode;
*(sptr_near-1) = curDist;
curDist = childDist;
curNode = child;
}
/* push hit child onto stack*/
else {
*(sptr_node-1) = child;
*(sptr_near-1) = childDist;
}
assert(sptr_node - stack_node < BVH4i::maxDepth);
}
}
}
/* return if stack is empty */
if (unlikely(curNode == BVH4i::invalidNode))
break;
/* intersect leaf */
const sseb valid_leaf = ray_tfar > curDist;
STAT3(shadow.trav_leaves,1,popcnt(valid_leaf),4);
size_t items; const Triangle* tri = (Triangle*) curNode.leaf(accel, items);
terminated |= TriangleIntersector4::occluded(!terminated,ray,tri,items,bvh->geometry);
if (all(terminated)) break;
ray_tfar = select(terminated,neg_inf,ray_tfar);
}
store4i(valid & terminated,&ray.geomID,0);
AVX_ZERO_UPPER();
}
void BVH8iIntersector8Hybrid<TriangleIntersector8>::occluded(avxb* valid_i, BVH8i* bvh, Ray8& ray)
{
/* load ray */
const avxb valid = *valid_i;
avxb terminated = !valid;
avx3f ray_org = ray.org, ray_dir = ray.dir;
avxf ray_tnear = ray.tnear, ray_tfar = ray.tfar;
#if defined(__FIX_RAYS__)
const avxf float_range = 0.1f*FLT_MAX;
ray_org = clamp(ray_org,avx3f(-float_range),avx3f(+float_range));
ray_dir = clamp(ray_dir,avx3f(-float_range),avx3f(+float_range));
ray_tnear = max(ray_tnear,FLT_MIN);
ray_tfar = min(ray_tfar,float(inf));
#endif
const avx3f rdir = rcp_safe(ray_dir);
const avx3f org(ray_org), org_rdir = org * rdir;
ray_tnear = select(valid,ray_tnear,avxf(pos_inf));
ray_tfar = select(valid,ray_tfar ,avxf(neg_inf));
const avxf inf = avxf(pos_inf);
/* compute near/far per ray */
avx3i nearXYZ;
nearXYZ.x = select(rdir.x >= 0.0f,avxi(0*(int)sizeof(avxf)),avxi(1*(int)sizeof(avxf)));
nearXYZ.y = select(rdir.y >= 0.0f,avxi(2*(int)sizeof(avxf)),avxi(3*(int)sizeof(avxf)));
nearXYZ.z = select(rdir.z >= 0.0f,avxi(4*(int)sizeof(avxf)),avxi(5*(int)sizeof(avxf)));
/* allocate stack and push root node */
avxf stack_near[stackSizeChunk];
NodeRef stack_node[stackSizeChunk];
stack_node[0] = BVH4i::invalidNode;
stack_near[0] = inf;
stack_node[1] = bvh->root;
stack_near[1] = ray_tnear;
NodeRef* stackEnd = stack_node+stackSizeChunk;
NodeRef* __restrict__ sptr_node = stack_node + 2;
avxf* __restrict__ sptr_near = stack_near + 2;
const Node * __restrict__ nodes = (Node *)bvh->nodePtr();
const Triangle * __restrict__ accel = (Triangle*)bvh->triPtr();
while (1)
{
/* pop next node from stack */
assert(sptr_node > stack_node);
sptr_node--;
sptr_near--;
NodeRef curNode = *sptr_node;
if (unlikely(curNode == BVH4i::invalidNode)) {
assert(sptr_node == stack_node);
break;
}
/* cull node if behind closest hit point */
avxf curDist = *sptr_near;
const avxb active = curDist < ray_tfar;
if (unlikely(none(active)))
continue;
/* switch to single ray traversal */
#if !defined(__WIN32__) || defined(__X86_64__)
size_t bits = movemask(active);
if (unlikely(__popcnt(bits) <= SWITCH_THRESHOLD)) {
for (size_t i=__bsf(bits); bits!=0; bits=__btc(bits,i), i=__bsf(bits)) {
if (occluded1(bvh,curNode,i,ray,ray_org,ray_dir,rdir,ray_tnear,ray_tfar,nearXYZ))
terminated[i] = -1;
}
if (all(terminated)) break;
ray_tfar = select(terminated,avxf(neg_inf),ray_tfar);
continue;
}
#endif
while (1)
{
/* test if this is a leaf node */
if (unlikely(curNode.isLeaf()))
break;
const avxb valid_node = ray_tfar > curDist;
STAT3(shadow.trav_nodes,1,popcnt(valid_node),8);
const Node* __restrict__ const node = (Node*)curNode.node(nodes);
/* pop of next node */
assert(sptr_node > stack_node);
sptr_node--;
sptr_near--;
curNode = *sptr_node;
curDist = *sptr_near;
for (unsigned i=0; i<8; i++)
{
const NodeRef child = node->children[i];
if (unlikely(child == BVH4i::emptyNode)) break;
#if defined(__AVX2__)
const avxf lclipMinX = msub(node->lower_x[i],rdir.x,org_rdir.x);
const avxf lclipMinY = msub(node->lower_y[i],rdir.y,org_rdir.y);
const avxf lclipMinZ = msub(node->lower_z[i],rdir.z,org_rdir.z);
const avxf lclipMaxX = msub(node->upper_x[i],rdir.x,org_rdir.x);
const avxf lclipMaxY = msub(node->upper_y[i],rdir.y,org_rdir.y);
const avxf lclipMaxZ = msub(node->upper_z[i],rdir.z,org_rdir.z);
const avxf lnearP = maxi(maxi(mini(lclipMinX, lclipMaxX), mini(lclipMinY, lclipMaxY)), mini(lclipMinZ, lclipMaxZ));
const avxf lfarP = mini(mini(maxi(lclipMinX, lclipMaxX), maxi(lclipMinY, lclipMaxY)), maxi(lclipMinZ, lclipMaxZ));
const avxb lhit = maxi(lnearP,ray_tnear) <= mini(lfarP,ray_tfar);
#else
const avxf lclipMinX = (node->lower_x[i] - org.x) * rdir.x;
const avxf lclipMinY = (node->lower_y[i] - org.y) * rdir.y;
const avxf lclipMinZ = (node->lower_z[i] - org.z) * rdir.z;
const avxf lclipMaxX = (node->upper_x[i] - org.x) * rdir.x;
const avxf lclipMaxY = (node->upper_y[i] - org.y) * rdir.y;
const avxf lclipMaxZ = (node->upper_z[i] - org.z) * rdir.z;
const avxf lnearP = max(max(min(lclipMinX, lclipMaxX), min(lclipMinY, lclipMaxY)), min(lclipMinZ, lclipMaxZ));
const avxf lfarP = min(min(max(lclipMinX, lclipMaxX), max(lclipMinY, lclipMaxY)), max(lclipMinZ, lclipMaxZ));
const avxb lhit = max(lnearP,ray_tnear) <= min(lfarP,ray_tfar);
#endif
/* if we hit the child we choose to continue with that child if it
is closer than the current next child, or we push it onto the stack */
if (likely(any(lhit)))
{
assert(sptr_node < stackEnd);
assert(child != BVH4i::emptyNode);
const avxf childDist = select(lhit,lnearP,inf);
sptr_node++;
sptr_near++;
/* push cur node onto stack and continue with hit child */
if (any(childDist < curDist))
{
*(sptr_node-1) = curNode;
*(sptr_near-1) = curDist;
curDist = childDist;
curNode = child;
}
/* push hit child onto stack */
else {
*(sptr_node-1) = child;
*(sptr_near-1) = childDist;
}
}
}
}
/* return if stack is empty */
if (unlikely(curNode == BVH4i::invalidNode)) {
assert(sptr_node == stack_node);
break;
}
/* intersect leaf */
const avxb valid_leaf = ray_tfar > curDist;
STAT3(shadow.trav_leaves,1,popcnt(valid_leaf),8);
size_t items; const Triangle* prim = (Triangle*) curNode.leaf(accel,items);
terminated |= TriangleIntersector8::occluded(!terminated,ray,prim,items,bvh->geometry);
if (all(terminated)) break;
ray_tfar = select(terminated,avxf(neg_inf),ray_tfar);
}
store8i(valid & terminated,&ray.geomID,0);
AVX_ZERO_UPPER();
}
static int
mprimef(unsigned int *n, unsigned int *q, int k)
{
int i, j;
unsigned int *t, *x, *y;
// generate x
t = mcopy(n);
while (1) {
for (i = 0; i < MLENGTH(t); i++)
t[i] = rand();
x = mmod(t, n);
if (!MZERO(x) && !MEQUAL(x, 1))
break;
mfree(x);
}
mfree(t);
// exponentiate
y = mmodpow(x, q, n);
// done?
if (MEQUAL(y, 1)) {
mfree(x);
mfree(y);
return 1;
}
j = 0;
while (1) {
// y = n - 1?
t = msub(n, y);
if (MEQUAL(t, 1)) {
mfree(t);
mfree(x);
mfree(y);
return 1;
}
mfree(t);
if (++j == k) {
mfree(x);
mfree(y);
return 0;
}
// y = (y ^ 2) mod n
t = mmul(y, y);
mfree(y);
y = mmod(t, n);
mfree(t);
// y = 1?
if (MEQUAL(y, 1)) {
mfree(x);
mfree(y);
return 0;
}
}
}
__forceinline bool BVH4Intersector4Hybrid<PrimitiveIntersector4>::occluded1(const BVH4* bvh, NodeRef root, size_t k, Ray4& ray,
const sse3f& ray_org, const sse3f& ray_dir, const sse3f& ray_rdir,
const ssef& ray_tnear, const ssef& ray_tfar)
{
/*! stack state */
NodeRef stack[stackSizeSingle]; //!< stack of nodes that still need to get traversed
NodeRef* stackPtr = stack+1; //!< current stack pointer
NodeRef* stackEnd = stack+stackSizeSingle;
stack[0] = root;
/*! offsets to select the side that becomes the lower or upper bound */
const size_t nearX = ray_dir.x[k] >= 0.0f ? 0*sizeof(ssef) : 1*sizeof(ssef);
const size_t nearY = ray_dir.y[k] >= 0.0f ? 2*sizeof(ssef) : 3*sizeof(ssef);
const size_t nearZ = ray_dir.z[k] >= 0.0f ? 4*sizeof(ssef) : 5*sizeof(ssef);
/*! load the ray into SIMD registers */
const sse3f org (ray_org .x[k],ray_org .y[k],ray_org .z[k]);
const sse3f rdir(ray_rdir.x[k],ray_rdir.y[k],ray_rdir.z[k]);
const sse3f norg = -org, org_rdir(org*rdir);
const ssef rayNear(ray_tnear[k]), rayFar(ray_tfar[k]);
/* pop loop */
while (true) pop:
{
/*! pop next node */
if (unlikely(stackPtr == stack)) break;
stackPtr--;
NodeRef cur = (NodeRef) *stackPtr;
/* downtraversal loop */
while (true)
{
/*! stop if we found a leaf */
if (unlikely(cur.isLeaf())) break;
STAT3(shadow.trav_nodes,1,1,1);
/*! single ray intersection with 4 boxes */
const Node* node = cur.node();
const size_t farX = nearX ^ 16, farY = nearY ^ 16, farZ = nearZ ^ 16;
#if defined (__AVX2__)
const ssef tNearX = msub(load4f((const char*)node+nearX), rdir.x, org_rdir.x);
const ssef tNearY = msub(load4f((const char*)node+nearY), rdir.y, org_rdir.y);
const ssef tNearZ = msub(load4f((const char*)node+nearZ), rdir.z, org_rdir.z);
const ssef tFarX = msub(load4f((const char*)node+farX ), rdir.x, org_rdir.x);
const ssef tFarY = msub(load4f((const char*)node+farY ), rdir.y, org_rdir.y);
const ssef tFarZ = msub(load4f((const char*)node+farZ ), rdir.z, org_rdir.z);
#else
const ssef tNearX = (norg.x + load4f((const char*)node+nearX)) * rdir.x;
const ssef tNearY = (norg.y + load4f((const char*)node+nearY)) * rdir.y;
const ssef tNearZ = (norg.z + load4f((const char*)node+nearZ)) * rdir.z;
const ssef tFarX = (norg.x + load4f((const char*)node+farX )) * rdir.x;
const ssef tFarY = (norg.y + load4f((const char*)node+farY )) * rdir.y;
const ssef tFarZ = (norg.z + load4f((const char*)node+farZ )) * rdir.z;
#endif
#if defined(__SSE4_1__)
const ssef tNear = maxi(maxi(tNearX,tNearY),maxi(tNearZ,rayNear));
const ssef tFar = mini(mini(tFarX ,tFarY ),mini(tFarZ ,rayFar ));
const sseb vmask = cast(tNear) > cast(tFar);
size_t mask = movemask(vmask)^0xf;
#else
const ssef tNear = max(tNearX,tNearY,tNearZ,rayNear);
const ssef tFar = min(tFarX ,tFarY ,tFarZ ,rayFar);
const sseb vmask = tNear <= tFar;
size_t mask = movemask(vmask);
#endif
/*! if no child is hit, pop next node */
if (unlikely(mask == 0))
goto pop;
/*! one child is hit, continue with that child */
size_t r = __bscf(mask);
if (likely(mask == 0)) {
cur = node->child(r);
assert(cur != BVH4::emptyNode);
continue;
}
/*! two children are hit, push far child, and continue with closer child */
NodeRef c0 = node->child(r); const float d0 = tNear[r];
r = __bscf(mask);
NodeRef c1 = node->child(r); const float d1 = tNear[r];
assert(c0 != BVH4::emptyNode);
assert(c1 != BVH4::emptyNode);
if (likely(mask == 0)) {
assert(stackPtr < stackEnd);
if (d0 < d1) { *stackPtr = c1; stackPtr++; cur = c0; continue; }
else { *stackPtr = c0; stackPtr++; cur = c1; continue; }
}
assert(stackPtr < stackEnd);
*stackPtr = c0; stackPtr++;
assert(stackPtr < stackEnd);
*stackPtr = c1; stackPtr++;
/*! three children are hit */
r = __bscf(mask);
cur = node->child(r);
assert(cur != BVH4::emptyNode);
if (likely(mask == 0)) continue;
assert(stackPtr < stackEnd);
*stackPtr = cur; stackPtr++;
/*! four children are hit */
cur = node->child(3);
assert(cur != BVH4::emptyNode);
}
/*! this is a leaf node */
STAT3(shadow.trav_leaves,1,1,1);
size_t num; Primitive* prim = (Primitive*) cur.leaf(num);
if (PrimitiveIntersector4::occluded(ray,k,prim,num,bvh->geometry)) {
ray.geomID[k] = 0;
return true;
}
}
return false;
}
__forceinline void BVH4Intersector4Hybrid<PrimitiveIntersector4>::intersect1(const BVH4* bvh, NodeRef root, size_t k, Ray4& ray,
const sse3f& ray_org, const sse3f& ray_dir, const sse3f& ray_rdir,
const ssef& ray_tnear, const ssef& ray_tfar)
{
/*! stack state */
StackItem stack[stackSizeSingle]; //!< stack of nodes
StackItem* stackPtr = stack+1; //!< current stack pointer
StackItem* stackEnd = stack+stackSizeSingle;
stack[0].ptr = root;
stack[0].dist = neg_inf;
/*! offsets to select the side that becomes the lower or upper bound */
const size_t nearX = ray_dir.x[k] >= 0.0f ? 0*sizeof(ssef) : 1*sizeof(ssef);
const size_t nearY = ray_dir.y[k] >= 0.0f ? 2*sizeof(ssef) : 3*sizeof(ssef);
const size_t nearZ = ray_dir.z[k] >= 0.0f ? 4*sizeof(ssef) : 5*sizeof(ssef);
/*! load the ray into SIMD registers */
const sse3f org (ray_org .x[k],ray_org .y[k],ray_org .z[k]);
const sse3f rdir(ray_rdir.x[k],ray_rdir.y[k],ray_rdir.z[k]);
const sse3f norg = -org, org_rdir(org*rdir);
ssef rayNear(ray_tnear[k]), rayFar(ray_tfar[k]);
/* pop loop */
while (true) pop:
{
/*! pop next node */
if (unlikely(stackPtr == stack)) break;
stackPtr--;
NodeRef cur = NodeRef(stackPtr->ptr);
/*! if popped node is too far, pop next one */
if (unlikely(stackPtr->dist > ray.tfar[k]))
continue;
/* downtraversal loop */
while (true)
{
/*! stop if we found a leaf */
if (unlikely(cur.isLeaf())) break;
STAT3(normal.trav_nodes,1,1,1);
/*! single ray intersection with 4 boxes */
const Node* node = cur.node();
const size_t farX = nearX ^ 16, farY = nearY ^ 16, farZ = nearZ ^ 16;
#if defined (__AVX2__)
const ssef tNearX = msub(load4f((const char*)node+nearX), rdir.x, org_rdir.x);
const ssef tNearY = msub(load4f((const char*)node+nearY), rdir.y, org_rdir.y);
const ssef tNearZ = msub(load4f((const char*)node+nearZ), rdir.z, org_rdir.z);
const ssef tFarX = msub(load4f((const char*)node+farX ), rdir.x, org_rdir.x);
const ssef tFarY = msub(load4f((const char*)node+farY ), rdir.y, org_rdir.y);
const ssef tFarZ = msub(load4f((const char*)node+farZ ), rdir.z, org_rdir.z);
#else
const ssef tNearX = (norg.x + load4f((const char*)node+nearX)) * rdir.x;
const ssef tNearY = (norg.y + load4f((const char*)node+nearY)) * rdir.y;
const ssef tNearZ = (norg.z + load4f((const char*)node+nearZ)) * rdir.z;
const ssef tFarX = (norg.x + load4f((const char*)node+farX )) * rdir.x;
const ssef tFarY = (norg.y + load4f((const char*)node+farY )) * rdir.y;
const ssef tFarZ = (norg.z + load4f((const char*)node+farZ )) * rdir.z;
#endif
#if defined(__SSE4_1__)
const ssef tNear = maxi(maxi(tNearX,tNearY),maxi(tNearZ,rayNear));
const ssef tFar = mini(mini(tFarX ,tFarY ),mini(tFarZ ,rayFar ));
const sseb vmask = cast(tNear) > cast(tFar);
size_t mask = movemask(vmask)^0xf;
#else
const ssef tNear = max(tNearX,tNearY,tNearZ,rayNear);
const ssef tFar = min(tFarX ,tFarY ,tFarZ ,rayFar);
const sseb vmask = tNear <= tFar;
size_t mask = movemask(vmask);
#endif
/*! if no child is hit, pop next node */
if (unlikely(mask == 0))
goto pop;
/*! one child is hit, continue with that child */
size_t r = __bscf(mask);
if (likely(mask == 0)) {
cur = node->child(r);
assert(cur != BVH4::emptyNode);
continue;
}
/*! two children are hit, push far child, and continue with closer child */
NodeRef c0 = node->child(r); const float d0 = tNear[r];
r = __bscf(mask);
NodeRef c1 = node->child(r); const float d1 = tNear[r];
assert(c0 != BVH4::emptyNode);
assert(c1 != BVH4::emptyNode);
if (likely(mask == 0)) {
assert(stackPtr < stackEnd);
if (d0 < d1) { stackPtr->ptr = c1; stackPtr->dist = d1; stackPtr++; cur = c0; continue; }
else { stackPtr->ptr = c0; stackPtr->dist = d0; stackPtr++; cur = c1; continue; }
}
/*! Here starts the slow path for 3 or 4 hit children. We push
* all nodes onto the stack to sort them there. */
assert(stackPtr < stackEnd);
stackPtr->ptr = c0; stackPtr->dist = d0; stackPtr++;
assert(stackPtr < stackEnd);
stackPtr->ptr = c1; stackPtr->dist = d1; stackPtr++;
/*! three children are hit, push all onto stack and sort 3 stack items, continue with closest child */
assert(stackPtr < stackEnd);
r = __bscf(mask);
NodeRef c = node->child(r); float d = tNear[r]; stackPtr->ptr = c; stackPtr->dist = d; stackPtr++;
assert(c != BVH4::emptyNode);
if (likely(mask == 0)) {
sort(stackPtr[-1],stackPtr[-2],stackPtr[-3]);
cur = (NodeRef) stackPtr[-1].ptr; stackPtr--;
continue;
}
/*! four children are hit, push all onto stack and sort 4 stack items, continue with closest child */
assert(stackPtr < stackEnd);
r = __bscf(mask);
c = node->child(r); d = tNear[r]; stackPtr->ptr = c; stackPtr->dist = d; stackPtr++;
assert(c != BVH4::emptyNode);
sort(stackPtr[-1],stackPtr[-2],stackPtr[-3],stackPtr[-4]);
cur = (NodeRef) stackPtr[-1].ptr; stackPtr--;
}
/*! this is a leaf node */
STAT3(normal.trav_leaves,1,1,1);
size_t num; Primitive* prim = (Primitive*) cur.leaf(num);
PrimitiveIntersector4::intersect(ray,k,prim,num,bvh->geometry);
rayFar = ray.tfar[k];
}
}
void BVH4Intersector4Hybrid<PrimitiveIntersector4>::intersect(sseb* valid_i, BVH4* bvh, Ray4& ray)
{
/* load ray */
const sseb valid0 = *valid_i;
sse3f ray_org = ray.org, ray_dir = ray.dir;
ssef ray_tnear = ray.tnear, ray_tfar = ray.tfar;
#if defined(__FIX_RAYS__)
const ssef float_range = 0.1f*FLT_MAX;
ray_org = clamp(ray_org,sse3f(-float_range),sse3f(+float_range));
ray_dir = clamp(ray_dir,sse3f(-float_range),sse3f(+float_range));
ray_tnear = max(ray_tnear,FLT_MIN);
ray_tfar = min(ray_tfar,float(inf));
#endif
const sse3f rdir = rcp_safe(ray_dir);
const sse3f org(ray_org), org_rdir = org * rdir;
ray_tnear = select(valid0,ray_tnear,ssef(pos_inf));
ray_tfar = select(valid0,ray_tfar ,ssef(neg_inf));
const ssef inf = ssef(pos_inf);
/* allocate stack and push root node */
ssef stack_near[stackSizeChunk];
NodeRef stack_node[stackSizeChunk];
stack_node[0] = BVH4::invalidNode;
stack_near[0] = inf;
stack_node[1] = bvh->root;
stack_near[1] = ray_tnear;
NodeRef* stackEnd = stack_node+stackSizeChunk;
NodeRef* __restrict__ sptr_node = stack_node + 2;
ssef* __restrict__ sptr_near = stack_near + 2;
while (1)
{
/* pop next node from stack */
assert(sptr_node > stack_node);
sptr_node--;
sptr_near--;
NodeRef curNode = *sptr_node;
if (unlikely(curNode == BVH4::invalidNode)) {
assert(sptr_node == stack_node);
break;
}
/* cull node if behind closest hit point */
ssef curDist = *sptr_near;
const sseb active = curDist < ray_tfar;
if (unlikely(none(active)))
continue;
/* switch to single ray traversal */
#if !defined(__WIN32__) || defined(__X86_64__)
size_t bits = movemask(active);
if (unlikely(__popcnt(bits) <= SWITCH_THRESHOLD)) {
for (size_t i=__bsf(bits); bits!=0; bits=__btc(bits,i), i=__bsf(bits)) {
intersect1(bvh,curNode,i,ray,ray_org,ray_dir,rdir,ray_tnear,ray_tfar);
}
ray_tfar = ray.tfar;
continue;
}
#endif
while (1)
{
/* test if this is a leaf node */
if (unlikely(curNode.isLeaf()))
break;
const sseb valid_node = ray_tfar > curDist;
STAT3(normal.trav_nodes,1,popcnt(valid_node),4);
const Node* __restrict__ const node = curNode.node();
/* pop of next node */
assert(sptr_node > stack_node);
sptr_node--;
sptr_near--;
curNode = *sptr_node;
curDist = *sptr_near;
#pragma unroll(4)
for (unsigned i=0; i<4; i++)
{
const NodeRef child = node->children[i];
if (unlikely(child == BVH4::emptyNode)) break;
#if defined(__AVX2__)
const ssef lclipMinX = msub(node->lower_x[i],rdir.x,org_rdir.x);
const ssef lclipMinY = msub(node->lower_y[i],rdir.y,org_rdir.y);
const ssef lclipMinZ = msub(node->lower_z[i],rdir.z,org_rdir.z);
const ssef lclipMaxX = msub(node->upper_x[i],rdir.x,org_rdir.x);
const ssef lclipMaxY = msub(node->upper_y[i],rdir.y,org_rdir.y);
const ssef lclipMaxZ = msub(node->upper_z[i],rdir.z,org_rdir.z);
#else
const ssef lclipMinX = (node->lower_x[i] - org.x) * rdir.x;
const ssef lclipMinY = (node->lower_y[i] - org.y) * rdir.y;
const ssef lclipMinZ = (node->lower_z[i] - org.z) * rdir.z;
const ssef lclipMaxX = (node->upper_x[i] - org.x) * rdir.x;
const ssef lclipMaxY = (node->upper_y[i] - org.y) * rdir.y;
const ssef lclipMaxZ = (node->upper_z[i] - org.z) * rdir.z;
#endif
#if defined(__SSE4_1__)
const ssef lnearP = maxi(maxi(mini(lclipMinX, lclipMaxX), mini(lclipMinY, lclipMaxY)), mini(lclipMinZ, lclipMaxZ));
const ssef lfarP = mini(mini(maxi(lclipMinX, lclipMaxX), maxi(lclipMinY, lclipMaxY)), maxi(lclipMinZ, lclipMaxZ));
const sseb lhit = maxi(lnearP,ray_tnear) <= mini(lfarP,ray_tfar);
#else
const ssef lnearP = max(max(min(lclipMinX, lclipMaxX), min(lclipMinY, lclipMaxY)), min(lclipMinZ, lclipMaxZ));
const ssef lfarP = min(min(max(lclipMinX, lclipMaxX), max(lclipMinY, lclipMaxY)), max(lclipMinZ, lclipMaxZ));
const sseb lhit = max(lnearP,ray_tnear) <= min(lfarP,ray_tfar);
#endif
/* if we hit the child we choose to continue with that child if it
is closer than the current next child, or we push it onto the stack */
if (likely(any(lhit)))
{
assert(sptr_node < stackEnd);
const ssef childDist = select(lhit,lnearP,inf);
const NodeRef child = node->children[i];
assert(child != BVH4::emptyNode);
sptr_node++;
sptr_near++;
/* push cur node onto stack and continue with hit child */
if (any(childDist < curDist))
{
*(sptr_node-1) = curNode;
*(sptr_near-1) = curDist;
curDist = childDist;
curNode = child;
}
/* push hit child onto stack */
else {
*(sptr_node-1) = child;
*(sptr_near-1) = childDist;
}
}
}
}
/* return if stack is empty */
if (unlikely(curNode == BVH4::invalidNode)) {
assert(sptr_node == stack_node);
break;
}
/* intersect leaf */
const sseb valid_leaf = ray_tfar > curDist;
STAT3(normal.trav_leaves,1,popcnt(valid_leaf),4);
size_t items; const Primitive* prim = (Primitive*) curNode.leaf(items);
PrimitiveIntersector4::intersect(valid_leaf,ray,prim,items,bvh->geometry);
ray_tfar = select(valid_leaf,ray.tfar,ray_tfar);
}
AVX_ZERO_UPPER();
}
__forceinline void intersectT(const BVH4* bvh, Ray& ray)
{
typedef typename TriangleIntersector::Triangle Triangle;
typedef StackItemT<size_t> StackItem;
typedef typename BVH4::NodeRef NodeRef;
typedef typename BVH4::Node Node;
/*! stack state */
StackItem stack[1+3*BVH4::maxDepth]; //!< stack of nodes
StackItem* stackPtr = stack+1; //!< current stack pointer
stack[0].ptr = bvh->root;
stack[0].dist = neg_inf;
/*! load the ray into SIMD registers */
const avxf pos_neg = avxf(ssef(+0.0f),ssef(-0.0f));
const avxf neg_pos = avxf(ssef(-0.0f),ssef(+0.0f));
const avxf flipSignX = swapX ? neg_pos : pos_neg;
const avxf flipSignY = swapY ? neg_pos : pos_neg;
const avxf flipSignZ = swapZ ? neg_pos : pos_neg;
const Vector3f ray_rdir = rcp_safe(ray.dir);
const avx3f norg(-ray.org.x,-ray.org.y,-ray.org.z);
const avx3f rdir(ray_rdir.x^flipSignX,ray_rdir.y^flipSignY,ray_rdir.z^flipSignZ);
const avx3f org_rdir(avx3f(ray.org.x,ray.org.y,ray.org.z)*rdir);
avxf rayNearFar(ssef(ray.tnear),-ssef(ray.tfar));
const void* nodePtr = bvh->nodePtr();
const void* triPtr = bvh->triPtr();
/* pop loop */
while (true) pop:
{
/*! pop next node */
if (unlikely(stackPtr == stack)) break;
stackPtr--;
NodeRef cur = NodeRef(stackPtr->ptr);
/*! if popped node is too far, pop next one */
if (unlikely(stackPtr->dist > ray.tfar))
continue;
/* downtraversal loop */
while (true)
{
/*! stop if we found a leaf */
if (unlikely(cur.isLeaf())) break;
STAT3(normal.trav_nodes,1,1,1);
/*! single ray intersection with 4 boxes */
const Node* node = cur.node(nodePtr);
#if defined (__AVX2__) || defined(__MIC__)
const avxf tLowerUpperX = msub(avxf::load(&node->lower_x), rdir.x, org_rdir.x);
const avxf tLowerUpperY = msub(avxf::load(&node->lower_y), rdir.y, org_rdir.y);
const avxf tLowerUpperZ = msub(avxf::load(&node->lower_z), rdir.z, org_rdir.z);
#else
const avxf tLowerUpperX = (norg.x + avxf::load(&node->lower_x)) * rdir.x;
const avxf tLowerUpperY = (norg.y + avxf::load(&node->lower_y)) * rdir.y;
const avxf tLowerUpperZ = (norg.z + avxf::load(&node->lower_z)) * rdir.z;
#endif
const avxf tNearFarX = swapX ? shuffle<1,0>(tLowerUpperX) : tLowerUpperX;
const avxf tNearFarY = swapY ? shuffle<1,0>(tLowerUpperY) : tLowerUpperY;
const avxf tNearFarZ = swapZ ? shuffle<1,0>(tLowerUpperZ) : tLowerUpperZ;
const avxf tNearFar = max(tNearFarX,tNearFarY,tNearFarZ,rayNearFar);
const ssef tNear = extract<0>(tNearFar);
const ssef tFar = extract<1>(tNearFar);
size_t mask = movemask(-tNear >= tFar);
/*! if no child is hit, pop next node */
if (unlikely(mask == 0))
goto pop;
/*! one child is hit, continue with that child */
size_t r = __bsf(mask); mask = __btc(mask,r);
if (likely(mask == 0)) {
cur = node->child(r);
continue;
}
/*! two children are hit, push far child, and continue with closer child */
NodeRef c0 = node->child(r); const float d0 = tNear[r];
r = __bsf(mask); mask = __btc(mask,r);
NodeRef c1 = node->child(r); const float d1 = tNear[r];
if (likely(mask == 0)) {
if (d0 < d1) { stackPtr->ptr = c1; stackPtr->dist = d1; stackPtr++; cur = c0; continue; }
else { stackPtr->ptr = c0; stackPtr->dist = d0; stackPtr++; cur = c1; continue; }
}
/*! Here starts the slow path for 3 or 4 hit children. We push
* all nodes onto the stack to sort them there. */
stackPtr->ptr = c0; stackPtr->dist = d0; stackPtr++;
stackPtr->ptr = c1; stackPtr->dist = d1; stackPtr++;
/*! three children are hit, push all onto stack and sort 3 stack items, continue with closest child */
r = __bsf(mask); mask = __btc(mask,r);
NodeRef c = node->child(r); float d = tNear[r]; stackPtr->ptr = c; stackPtr->dist = d; stackPtr++;
if (likely(mask == 0)) {
sort(stackPtr[-1],stackPtr[-2],stackPtr[-3]);
cur = (NodeRef) stackPtr[-1].ptr; stackPtr--;
continue;
}
/*! four children are hit, push all onto stack and sort 4 stack items, continue with closest child */
r = __bsf(mask); mask = __btc(mask,r);
c = node->child(r); d = tNear[r]; stackPtr->ptr = c; stackPtr->dist = d; stackPtr++;
sort(stackPtr[-1],stackPtr[-2],stackPtr[-3],stackPtr[-4]);
cur = (NodeRef) stackPtr[-1].ptr; stackPtr--;
}
/*! this is a leaf node */
STAT3(normal.trav_leaves,1,1,1);
size_t num; Triangle* tri = (Triangle*) cur.leaf(triPtr,num);
for (size_t i=0; i<num; i++)
TriangleIntersector::intersect(ray,tri[i],bvh->vertices);
rayNearFar = insert<1>(rayNearFar,-ssef(ray.tfar));
}
}
__forceinline bool occludedT(const BVH4* bvh, Ray& ray)
{
typedef typename TriangleIntersector::Triangle Triangle;
typedef StackItemT<size_t> StackItem;
typedef typename BVH4::NodeRef NodeRef;
typedef typename BVH4::Node Node;
/*! stack state */
NodeRef stack[1+3*BVH4::maxDepth]; //!< stack of nodes that still need to get traversed
NodeRef* stackPtr = stack+1; //!< current stack pointer
stack[0] = bvh->root;
/*! load the ray into SIMD registers */
const avxf pos_neg = avxf(ssef(+0.0f),ssef(-0.0f));
const avxf neg_pos = avxf(ssef(-0.0f),ssef(+0.0f));
const avxf flipSignX = swapX ? neg_pos : pos_neg;
const avxf flipSignY = swapY ? neg_pos : pos_neg;
const avxf flipSignZ = swapZ ? neg_pos : pos_neg;
const avx3f norg(-ray.org.x,-ray.org.y,-ray.org.z);
const Vector3f ray_rdir = rcp_safe(ray.dir);
const avx3f rdir(ray_rdir.x^flipSignX,ray_rdir.y^flipSignY,ray_rdir.z^flipSignZ);
const avx3f org_rdir(avx3f(ray.org.x,ray.org.y,ray.org.z)*rdir);
const avxf rayNearFar(ssef(ray.tnear),-ssef(ray.tfar));
const void* nodePtr = bvh->nodePtr();
const void* triPtr = bvh->triPtr();
/* pop loop */
while (true) pop:
{
/*! pop next node */
if (unlikely(stackPtr == stack)) break;
stackPtr--;
NodeRef cur = (NodeRef) *stackPtr;
/* downtraversal loop */
while (true)
{
/*! stop if we found a leaf */
if (unlikely(cur.isLeaf())) break;
STAT3(shadow.trav_nodes,1,1,1);
/*! single ray intersection with 4 boxes */
const Node* node = cur.node(nodePtr);
#if defined (__AVX2__) || defined(__MIC__)
const avxf tLowerUpperX = msub(avxf::load(&node->lower_x), rdir.x, org_rdir.x);
const avxf tLowerUpperY = msub(avxf::load(&node->lower_y), rdir.y, org_rdir.y);
const avxf tLowerUpperZ = msub(avxf::load(&node->lower_z), rdir.z, org_rdir.z);
#else
const avxf tLowerUpperX = (norg.x + avxf::load(&node->lower_x)) * rdir.x;
const avxf tLowerUpperY = (norg.y + avxf::load(&node->lower_y)) * rdir.y;
const avxf tLowerUpperZ = (norg.z + avxf::load(&node->lower_z)) * rdir.z;
#endif
const avxf tNearFarX = swapX ? shuffle<1,0>(tLowerUpperX) : tLowerUpperX;
const avxf tNearFarY = swapY ? shuffle<1,0>(tLowerUpperY) : tLowerUpperY;
const avxf tNearFarZ = swapZ ? shuffle<1,0>(tLowerUpperZ) : tLowerUpperZ;
const avxf tNearFar = max(tNearFarX,tNearFarY,tNearFarZ,rayNearFar);
const ssef tNear = extract<0>(tNearFar);
const ssef tFar = extract<1>(tNearFar);
size_t mask = movemask(-tNear >= tFar);
/*! if no child is hit, pop next node */
if (unlikely(mask == 0))
goto pop;
/*! one child is hit, continue with that child */
size_t r = __bsf(mask); mask = __btc(mask,r);
if (likely(mask == 0)) {
cur = node->child(r);
continue;
}
/*! two children are hit, push far child, and continue with closer child */
NodeRef c0 = node->child(r); const float d0 = tNear[r];
r = __bsf(mask); mask = __btc(mask,r);
NodeRef c1 = node->child(r); const float d1 = tNear[r];
if (likely(mask == 0)) {
if (d0 < d1) { *stackPtr = c1; stackPtr++; cur = c0; continue; }
else { *stackPtr = c0; stackPtr++; cur = c1; continue; }
}
*stackPtr = c0; stackPtr++;
*stackPtr = c1; stackPtr++;
/*! three children are hit */
r = __bsf(mask); mask = __btc(mask,r);
cur = node->child(r); *stackPtr = cur; stackPtr++;
if (likely(mask == 0)) {
stackPtr--;
continue;
}
/*! four children are hit */
cur = node->child(3);
}
/*! this is a leaf node */
STAT3(shadow.trav_leaves,1,1,1);
size_t num; Triangle* tri = (Triangle*) cur.leaf(triPtr,num);
for (size_t i=0; i<num; i++) {
if (TriangleIntersector::occluded(ray,tri[i],bvh->vertices)) {
AVX_ZERO_UPPER();
return true;
}
}
}
AVX_ZERO_UPPER();
return false;
}
unsigned int *
mfactor(unsigned int *n)
{
unsigned int *r, *root, *t, *two, *x, *y;
two = mint(2);
root = msqrt(n);
// y = 1;
y = mint(1);
// x = 2 isqrt(n) + 1
t = madd(root, root);
x = madd(t, y);
mfree(t);
// r = isqrt(n) ^ 2 - n
t = mmul(root, root);
r = msub(t, n);
mfree(t);
mfree(root);
while (1) {
if (MZERO(r)) {
// n = (x - y) / 2
t = msub(x, y);
n = mdiv(t, two);
mfree(t);
mfree(r);
mfree(x);
mfree(y);
mfree(two);
return n;
}
// r = r + x
t = madd(r, x);
mfree(r);
r = t;
// x = x + 2
t = madd(x, two);
mfree(x);
x = t;
while (1) {
// r = r - y
t = msub(r, y);
mfree(r);
r = t;
// y = y + 2
t = madd(y, two);
mfree(y);
y = t;
if (MSIGN(r) == -1 || MZERO(r))
break;
}
}
}
__forceinline void BVH8iIntersector8Hybrid<TriangleIntersector8>::intersect1(const BVH8i* bvh, NodeRef root, const size_t k, Ray8& ray,const avx3f &ray_org, const avx3f &ray_dir, const avx3f &ray_rdir, const avxf &ray_tnear, const avxf &ray_tfar, const avx3i& nearXYZ)
{
/*! stack state */
StackItemInt64 stack[stackSizeSingle]; //!< stack of nodes
StackItemInt64* stackPtr = stack+1; //!< current stack pointer
StackItemInt64* stackEnd = stack+stackSizeSingle;
stack[0].ptr = root;
stack[0].dist = neg_inf;
/*! offsets to select the side that becomes the lower or upper bound */
const size_t nearX = nearXYZ.x[k];
const size_t nearY = nearXYZ.y[k];
const size_t nearZ = nearXYZ.z[k];
/*! load the ray into SIMD registers */
const avx3f org (ray_org .x[k],ray_org .y[k],ray_org .z[k]);
const avx3f rdir(ray_rdir.x[k],ray_rdir.y[k],ray_rdir.z[k]);
const avx3f org_rdir(org*rdir);
avxf rayNear(ray_tnear[k]), rayFar(ray_tfar[k]);
const Node * __restrict__ nodes = (Node *)bvh->nodePtr();
const Triangle * __restrict__ accel = (Triangle*)bvh->triPtr();
/* pop loop */
while (true) pop:
{
/*! pop next node */
if (unlikely(stackPtr == stack)) break;
stackPtr--;
NodeRef cur = NodeRef(stackPtr->ptr);
/*! if popped node is too far, pop next one */
if (unlikely(*(float*)&stackPtr->dist > ray.tfar[k]))
continue;
/* downtraversal loop */
while (true)
{
/*! stop if we found a leaf */
if (unlikely(cur.isLeaf())) break;
STAT3(normal.trav_nodes,1,1,1);
/*! single ray intersection with 4 boxes */
const Node* node = (Node*)cur.node(nodes);
const size_t farX = nearX ^ sizeof(avxf), farY = nearY ^ sizeof(avxf), farZ = nearZ ^ sizeof(avxf);
#if defined (__AVX2__)
const avxf tNearX = msub(load8f((const char*)node+nearX), rdir.x, org_rdir.x);
const avxf tNearY = msub(load8f((const char*)node+nearY), rdir.y, org_rdir.y);
const avxf tNearZ = msub(load8f((const char*)node+nearZ), rdir.z, org_rdir.z);
const avxf tFarX = msub(load8f((const char*)node+farX ), rdir.x, org_rdir.x);
const avxf tFarY = msub(load8f((const char*)node+farY ), rdir.y, org_rdir.y);
const avxf tFarZ = msub(load8f((const char*)node+farZ ), rdir.z, org_rdir.z);
#else
const avxf tNearX = (load8f((const char*)node+nearX) - org.x) * rdir.x;
const avxf tNearY = (load8f((const char*)node+nearY) - org.y) * rdir.y;
const avxf tNearZ = (load8f((const char*)node+nearZ) - org.z) * rdir.z;
const avxf tFarX = (load8f((const char*)node+farX ) - org.x) * rdir.x;
const avxf tFarY = (load8f((const char*)node+farY ) - org.y) * rdir.y;
const avxf tFarZ = (load8f((const char*)node+farZ ) - org.z) * rdir.z;
#endif
#if defined(__AVX2__)
const avxf tNear = maxi(maxi(tNearX,tNearY),maxi(tNearZ,rayNear));
const avxf tFar = mini(mini(tFarX ,tFarY ),mini(tFarZ ,rayFar ));
const avxb vmask = cast(tNear) > cast(tFar);
unsigned int mask = movemask(vmask)^0xff;
#else
const avxf tNear = max(tNearX,tNearY,tNearZ,rayNear);
const avxf tFar = min(tFarX ,tFarY ,tFarZ ,rayFar);
const avxb vmask = tNear <= tFar;
unsigned int mask = movemask(vmask);
#endif
/*! if no child is hit, pop next node */
if (unlikely(mask == 0))
goto pop;
/*! one child is hit, continue with that child */
size_t r = __bscf(mask);
if (likely(mask == 0)) {
cur = node->child(r);
assert(cur != BVH4i::emptyNode);
continue;
}
/*! two children are hit, push far child, and continue with closer child */
NodeRef c0 = node->child(r); const unsigned int d0 = ((unsigned int*)&tNear)[r];
r = __bscf(mask);
NodeRef c1 = node->child(r); const unsigned int d1 = ((unsigned int*)&tNear)[r];
assert(c0 != BVH4i::emptyNode);
assert(c1 != BVH4i::emptyNode);
if (likely(mask == 0)) {
assert(stackPtr < stackEnd);
if (d0 < d1) { stackPtr->ptr = c1; stackPtr->dist = d1; stackPtr++; cur = c0; continue; }
else { stackPtr->ptr = c0; stackPtr->dist = d0; stackPtr++; cur = c1; continue; }
}
/*! Here starts the slow path for 3 or 4 hit children. We push
* all nodes onto the stack to sort them there. */
assert(stackPtr < stackEnd);
stackPtr->ptr = c0; stackPtr->dist = d0; stackPtr++;
assert(stackPtr < stackEnd);
stackPtr->ptr = c1; stackPtr->dist = d1; stackPtr++;
/*! three children are hit, push all onto stack and sort 3 stack items, continue with closest child */
assert(stackPtr < stackEnd);
r = __bscf(mask);
NodeRef c = node->child(r); unsigned int d = ((unsigned int*)&tNear)[r]; stackPtr->ptr = c; stackPtr->dist = d; stackPtr++;
assert(c0 != BVH4i::emptyNode);
if (likely(mask == 0)) {
sort(stackPtr[-1],stackPtr[-2],stackPtr[-3]);
cur = (NodeRef) stackPtr[-1].ptr; stackPtr--;
continue;
}
/*! four children are hit, push all onto stack and sort 4 stack items, continue with closest child */
assert(stackPtr < stackEnd);
r = __bscf(mask);
c = node->child(r); d = ((unsigned int*)&tNear)[r]; stackPtr->ptr = c; stackPtr->dist = d; stackPtr++;
assert(c != BVH4i::emptyNode);
if (likely(mask == 0)) {
sort(stackPtr[-1],stackPtr[-2],stackPtr[-3],stackPtr[-4]);
cur = (NodeRef) stackPtr[-1].ptr; stackPtr--;
continue;
}
while(1)
{
r = __bscf(mask);
c = node->child(r); d = ((unsigned int*)&tNear)[r]; stackPtr->ptr = c; stackPtr->dist = d; stackPtr++;
if (unlikely(mask == 0)) break;
}
cur = (NodeRef) stackPtr[-1].ptr; stackPtr--;
}
/*! this is a leaf node */
STAT3(normal.trav_leaves,1,1,1);
size_t num; Triangle* prim = (Triangle*) cur.leaf(accel,num);
TriangleIntersector8::intersect(ray,k,prim,num,bvh->geometry);
rayFar = ray.tfar[k];
}
}
__forceinline bool BVH8iIntersector8Hybrid<TriangleIntersector8>::occluded1(const BVH8i* bvh, NodeRef root, const size_t k, Ray8& ray,const avx3f &ray_org, const avx3f &ray_dir, const avx3f &ray_rdir, const avxf &ray_tnear, const avxf &ray_tfar, const avx3i& nearXYZ)
{
/*! stack state */
NodeRef stack[stackSizeSingle]; //!< stack of nodes that still need to get traversed
NodeRef* stackPtr = stack+1; //!< current stack pointer
NodeRef* stackEnd = stack+stackSizeSingle;
stack[0] = root;
/*! offsets to select the side that becomes the lower or upper bound */
const size_t nearX = nearXYZ.x[k];
const size_t nearY = nearXYZ.y[k];
const size_t nearZ = nearXYZ.z[k];
/*! load the ray into SIMD registers */
const avx3f org (ray_org .x[k],ray_org .y[k],ray_org .z[k]);
const avx3f rdir(ray_rdir.x[k],ray_rdir.y[k],ray_rdir.z[k]);
const avx3f norg = -org, org_rdir(org*rdir);
const avxf rayNear(ray_tnear[k]), rayFar(ray_tfar[k]);
const Node * __restrict__ nodes = (Node *)bvh->nodePtr();
const Triangle * __restrict__ accel = (Triangle*)bvh->triPtr();
/* pop loop */
while (true) pop:
{
/*! pop next node */
if (unlikely(stackPtr == stack)) break;
stackPtr--;
NodeRef cur = (NodeRef) *stackPtr;
/* downtraversal loop */
while (true)
{
/*! stop if we found a leaf */
if (unlikely(cur.isLeaf())) break;
STAT3(shadow.trav_nodes,1,1,1);
/*! single ray intersection with 4 boxes */
const Node* node = (Node*)cur.node(nodes);
const size_t farX = nearX ^ sizeof(avxf), farY = nearY ^ sizeof(avxf), farZ = nearZ ^ sizeof(avxf);
#if defined (__AVX2__)
const avxf tNearX = msub(load8f((const char*)node+nearX), rdir.x, org_rdir.x);
const avxf tNearY = msub(load8f((const char*)node+nearY), rdir.y, org_rdir.y);
const avxf tNearZ = msub(load8f((const char*)node+nearZ), rdir.z, org_rdir.z);
const avxf tFarX = msub(load8f((const char*)node+farX ), rdir.x, org_rdir.x);
const avxf tFarY = msub(load8f((const char*)node+farY ), rdir.y, org_rdir.y);
const avxf tFarZ = msub(load8f((const char*)node+farZ ), rdir.z, org_rdir.z);
#else
const avxf tNearX = (norg.x + load8f((const char*)node+nearX)) * rdir.x;
const avxf tNearY = (norg.y + load8f((const char*)node+nearY)) * rdir.y;
const avxf tNearZ = (norg.z + load8f((const char*)node+nearZ)) * rdir.z;
const avxf tFarX = (norg.x + load8f((const char*)node+farX )) * rdir.x;
const avxf tFarY = (norg.y + load8f((const char*)node+farY )) * rdir.y;
const avxf tFarZ = (norg.z + load8f((const char*)node+farZ )) * rdir.z;
#endif
#if defined(__AVX2__)
const avxf tNear = maxi(maxi(tNearX,tNearY),maxi(tNearZ,rayNear));
const avxf tFar = mini(mini(tFarX ,tFarY ),mini(tFarZ ,rayFar ));
const avxb vmask = cast(tNear) > cast(tFar);
unsigned int mask = movemask(vmask)^0xff;
#else
const avxf tNear = max(tNearX,tNearY,tNearZ,rayNear);
const avxf tFar = min(tFarX ,tFarY ,tFarZ ,rayFar);
const avxb vmask = tNear <= tFar;
unsigned int mask = movemask(vmask);
#endif
/*! if no child is hit, pop next node */
if (unlikely(mask == 0))
goto pop;
/*! one child is hit, continue with that child */
size_t r = __bscf(mask);
if (likely(mask == 0)) {
cur = node->child(r);
assert(cur != BVH4i::emptyNode);
continue;
}
/*! two children are hit, push far child, and continue with closer child */
NodeRef c0 = node->child(r); const unsigned int d0 = ((unsigned int*)&tNear)[r];
r = __bscf(mask);
NodeRef c1 = node->child(r); const unsigned int d1 = ((unsigned int*)&tNear)[r];
assert(c0 != BVH4i::emptyNode);
assert(c1 != BVH4i::emptyNode);
if (likely(mask == 0)) {
assert(stackPtr < stackEnd);
if (d0 < d1) { *stackPtr = c1; stackPtr++; cur = c0; continue; }
else { *stackPtr = c0; stackPtr++; cur = c1; continue; }
}
assert(stackPtr < stackEnd);
*stackPtr = c0; stackPtr++;
assert(stackPtr < stackEnd);
*stackPtr = c1; stackPtr++;
/*! three children are hit */
r = __bscf(mask);
cur = node->child(r);
assert(cur != BVH4i::emptyNode);
if (likely(mask == 0)) continue;
while(1)
{
r = __bscf(mask);
NodeRef c = node->child(r); *stackPtr = c; stackPtr++;
if (unlikely(mask == 0)) break;
}
cur = (NodeRef) stackPtr[-1]; stackPtr--;
// assert(stackPtr < stackEnd);
// *stackPtr = cur; stackPtr++;
// /*! four children are hit */
// cur = node->child(3);
// assert(cur != BVH4i::emptyNode);
}
/*! this is a leaf node */
STAT3(shadow.trav_leaves,1,1,1);
size_t num; Triangle* prim = (Triangle*) cur.leaf(accel,num);
if (TriangleIntersector8::occluded(ray,k,prim,num,bvh->geometry)) {
ray.geomID[k] = 0;
return true;
}
}
return false;
}
void BVH4iIntersector16Hybrid<LeafIntersector,ENABLE_COMPRESSED_BVH4I_NODES>::intersect(mic_i* valid_i, BVH4i* bvh, Ray16& ray16)
{
/* near and node stack */
__aligned(64) mic_f stack_dist[3*BVH4i::maxDepth+1];
__aligned(64) NodeRef stack_node[3*BVH4i::maxDepth+1];
__aligned(64) NodeRef stack_node_single[3*BVH4i::maxDepth+1];
/* load ray */
const mic_m valid0 = *(mic_i*)valid_i != mic_i(0);
const mic3f rdir16 = rcp_safe(ray16.dir);
const mic3f org_rdir16 = ray16.org * rdir16;
mic_f ray_tnear = select(valid0,ray16.tnear,pos_inf);
mic_f ray_tfar = select(valid0,ray16.tfar ,neg_inf);
const mic_f inf = mic_f(pos_inf);
/* allocate stack and push root node */
stack_node[0] = BVH4i::invalidNode;
stack_dist[0] = inf;
stack_node[1] = bvh->root;
stack_dist[1] = ray_tnear;
NodeRef* __restrict__ sptr_node = stack_node + 2;
mic_f* __restrict__ sptr_dist = stack_dist + 2;
const Node * __restrict__ nodes = (Node *)bvh->nodePtr();
const Triangle1 * __restrict__ accel = (Triangle1*)bvh->triPtr();
while (1) pop:
{
/* pop next node from stack */
NodeRef curNode = *(sptr_node-1);
mic_f curDist = *(sptr_dist-1);
sptr_node--;
sptr_dist--;
const mic_m m_stackDist = ray_tfar > curDist;
/* stack emppty ? */
if (unlikely(curNode == BVH4i::invalidNode)) break;
/* cull node if behind closest hit point */
if (unlikely(none(m_stackDist))) continue;
///////////////////////////////////////////////////////////////////////////////////////////////////////////////
///////////////////////////////////////////////////////////////////////////////////////////////////////////////
///////////////////////////////////////////////////////////////////////////////////////////////////////////////
/* switch to single ray mode */
if (unlikely(countbits(m_stackDist) <= BVH4i::hybridSIMDUtilSwitchThreshold))
{
float *__restrict__ stack_dist_single = (float*)sptr_dist;
store16f(stack_dist_single,inf);
/* traverse single ray */
long rayIndex = -1;
while((rayIndex = bitscan64(rayIndex,m_stackDist)) != BITSCAN_NO_BIT_SET_64)
{
stack_node_single[0] = BVH4i::invalidNode;
stack_node_single[1] = curNode;
size_t sindex = 2;
const mic_f org_xyz = loadAOS4to16f(rayIndex,ray16.org.x,ray16.org.y,ray16.org.z);
const mic_f dir_xyz = loadAOS4to16f(rayIndex,ray16.dir.x,ray16.dir.y,ray16.dir.z);
const mic_f rdir_xyz = loadAOS4to16f(rayIndex,rdir16.x,rdir16.y,rdir16.z);
const mic_f org_rdir_xyz = org_xyz * rdir_xyz;
const mic_f min_dist_xyz = broadcast1to16f(&ray16.tnear[rayIndex]);
mic_f max_dist_xyz = broadcast1to16f(&ray16.tfar[rayIndex]);
const unsigned int leaf_mask = BVH4I_LEAF_MASK;
while (1)
{
NodeRef curNode = stack_node_single[sindex-1];
sindex--;
traverse_single_intersect<ENABLE_COMPRESSED_BVH4I_NODES>(curNode,
sindex,
rdir_xyz,
org_rdir_xyz,
min_dist_xyz,
max_dist_xyz,
stack_node_single,
stack_dist_single,
nodes,
leaf_mask);
/* return if stack is empty */
if (unlikely(curNode == BVH4i::invalidNode)) break;
/* intersect one ray against four triangles */
const bool hit = LeafIntersector::intersect(curNode,
rayIndex,
dir_xyz,
org_xyz,
min_dist_xyz,
max_dist_xyz,
ray16,
accel,
(Scene*)bvh->geometry);
if (hit)
compactStack(stack_node_single,stack_dist_single,sindex,max_dist_xyz);
}
}
ray_tfar = select(valid0,ray16.tfar ,neg_inf);
continue;
}
///////////////////////////////////////////////////////////////////////////////////////////////////////////////
///////////////////////////////////////////////////////////////////////////////////////////////////////////////
///////////////////////////////////////////////////////////////////////////////////////////////////////////////
const unsigned int leaf_mask = BVH4I_LEAF_MASK;
const mic3f org = ray16.org;
const mic3f dir = ray16.dir;
while (1)
{
/* test if this is a leaf node */
if (unlikely(curNode.isLeaf(leaf_mask))) break;
STAT3(normal.trav_nodes,1,popcnt(ray_tfar > curDist),16);
const Node* __restrict__ const node = curNode.node(nodes);
prefetch<PFHINT_L1>((mic_f*)node + 0);
prefetch<PFHINT_L1>((mic_f*)node + 1);
/* pop of next node */
sptr_node--;
sptr_dist--;
curNode = *sptr_node;
curDist = *sptr_dist;
#pragma unroll(4)
for (unsigned int i=0; i<4; i++)
{
BVH4i::NodeRef child;
mic_f lclipMinX,lclipMinY,lclipMinZ;
mic_f lclipMaxX,lclipMaxY,lclipMaxZ;
if (!ENABLE_COMPRESSED_BVH4I_NODES)
{
child = node->lower[i].child;
lclipMinX = msub(node->lower[i].x,rdir16.x,org_rdir16.x);
lclipMinY = msub(node->lower[i].y,rdir16.y,org_rdir16.y);
lclipMinZ = msub(node->lower[i].z,rdir16.z,org_rdir16.z);
lclipMaxX = msub(node->upper[i].x,rdir16.x,org_rdir16.x);
lclipMaxY = msub(node->upper[i].y,rdir16.y,org_rdir16.y);
lclipMaxZ = msub(node->upper[i].z,rdir16.z,org_rdir16.z);
}
else
{
BVH4i::QuantizedNode* __restrict__ const compressed_node = (BVH4i::QuantizedNode*)node;
child = compressed_node->child(i);
const mic_f startXYZ = compressed_node->decompress_startXYZ();
const mic_f diffXYZ = compressed_node->decompress_diffXYZ();
const mic_f clower = compressed_node->decompress_lowerXYZ(startXYZ,diffXYZ);
const mic_f cupper = compressed_node->decompress_upperXYZ(startXYZ,diffXYZ);
lclipMinX = msub(mic_f(clower[4*i+0]),rdir16.x,org_rdir16.x);
lclipMinY = msub(mic_f(clower[4*i+1]),rdir16.y,org_rdir16.y);
lclipMinZ = msub(mic_f(clower[4*i+2]),rdir16.z,org_rdir16.z);
lclipMaxX = msub(mic_f(cupper[4*i+0]),rdir16.x,org_rdir16.x);
lclipMaxY = msub(mic_f(cupper[4*i+1]),rdir16.y,org_rdir16.y);
lclipMaxZ = msub(mic_f(cupper[4*i+2]),rdir16.z,org_rdir16.z);
}
if (unlikely(i >=2 && child == BVH4i::invalidNode)) break;
const mic_f lnearP = max(max(min(lclipMinX, lclipMaxX), min(lclipMinY, lclipMaxY)), min(lclipMinZ, lclipMaxZ));
const mic_f lfarP = min(min(max(lclipMinX, lclipMaxX), max(lclipMinY, lclipMaxY)), max(lclipMinZ, lclipMaxZ));
const mic_m lhit = max(lnearP,ray_tnear) <= min(lfarP,ray_tfar);
const mic_f childDist = select(lhit,lnearP,inf);
const mic_m m_child_dist = childDist < curDist;
/* if we hit the child we choose to continue with that child if it
is closer than the current next child, or we push it onto the stack */
if (likely(any(lhit)))
{
sptr_node++;
sptr_dist++;
/* push cur node onto stack and continue with hit child */
if (any(m_child_dist))
{
*(sptr_node-1) = curNode;
*(sptr_dist-1) = curDist;
curDist = childDist;
curNode = child;
}
/* push hit child onto stack*/
else
{
*(sptr_node-1) = child;
*(sptr_dist-1) = childDist;
const char* __restrict__ const pnode = (char*)child.node(nodes);
prefetch<PFHINT_L2>(pnode + 0);
prefetch<PFHINT_L2>(pnode + 64);
}
assert(sptr_node - stack_node < BVH4i::maxDepth);
}
}
#if SWITCH_ON_DOWN_TRAVERSAL == 1
const mic_m curUtil = ray_tfar > curDist;
if (unlikely(countbits(curUtil) <= BVH4i::hybridSIMDUtilSwitchThreshold))
{
*sptr_node++ = curNode;
*sptr_dist++ = curDist;
goto pop;
}
#endif
}
/* return if stack is empty */
if (unlikely(curNode == BVH4i::invalidNode)) break;
/* intersect leaf */
const mic_m m_valid_leaf = ray_tfar > curDist;
STAT3(normal.trav_leaves,1,popcnt(m_valid_leaf),16);
LeafIntersector::intersect16(curNode,m_valid_leaf,dir,org,ray16,accel,(Scene*)bvh->geometry);
ray_tfar = select(m_valid_leaf,ray16.tfar,ray_tfar);
}
void BVH8Intersector1<robust,PrimitiveIntersector>::occluded(const BVH8* bvh, Ray& ray)
{
/*! perform per ray precalculations required by the primitive intersector */
Precalculations pre(ray,bvh);
/*! stack state */
NodeRef stack[stackSize]; //!< stack of nodes that still need to get traversed
NodeRef* stackPtr = stack+1; //!< current stack pointer
NodeRef* stackEnd = stack+stackSize;
stack[0] = bvh->root;
/* filter out invalid rays */
#if defined(RTCORE_IGNORE_INVALID_RAYS)
if (!ray.valid()) return;
#endif
/* verify correct input */
assert(ray.tnear > -FLT_MIN);
//assert(!(types & BVH4::FLAG_NODE_MB) || (ray.time >= 0.0f && ray.time <= 1.0f));
/*! load the ray into SIMD registers */
const Vec3f8 norg(-ray.org.x,-ray.org.y,-ray.org.z);
const Vec3fa ray_rdir = rcp_safe(ray.dir);
const Vec3f8 rdir(ray_rdir.x,ray_rdir.y,ray_rdir.z);
const Vec3fa ray_org_rdir = ray.org*ray_rdir;
const Vec3f8 org_rdir(ray_org_rdir.x,ray_org_rdir.y,ray_org_rdir.z);
const float8 ray_near(ray.tnear);
float8 ray_far(ray.tfar);
/*! offsets to select the side that becomes the lower or upper bound */
const size_t nearX = ray_rdir.x >= 0 ? 0*sizeof(float8) : 1*sizeof(float8);
const size_t nearY = ray_rdir.y >= 0 ? 2*sizeof(float8) : 3*sizeof(float8);
const size_t nearZ = ray_rdir.z >= 0 ? 4*sizeof(float8) : 5*sizeof(float8);
/* pop loop */
while (true) pop:
{
/*! pop next node */
if (unlikely(stackPtr == stack)) break;
stackPtr--;
NodeRef cur = (NodeRef) *stackPtr;
/* downtraversal loop */
while (true)
{
/*! stop if we found a leaf */
if (unlikely(cur.isLeaf())) break;
STAT3(shadow.trav_nodes,1,1,1);
/*! single ray intersection with 4 boxes */
const Node* node = cur.node();
const size_t farX = nearX ^ sizeof(float8), farY = nearY ^ sizeof(float8), farZ = nearZ ^ sizeof(float8);
#if defined (__AVX2__)
const float8 tNearX = msub(load8f((const char*)node+nearX), rdir.x, org_rdir.x);
const float8 tNearY = msub(load8f((const char*)node+nearY), rdir.y, org_rdir.y);
const float8 tNearZ = msub(load8f((const char*)node+nearZ), rdir.z, org_rdir.z);
const float8 tFarX = msub(load8f((const char*)node+farX ), rdir.x, org_rdir.x);
const float8 tFarY = msub(load8f((const char*)node+farY ), rdir.y, org_rdir.y);
const float8 tFarZ = msub(load8f((const char*)node+farZ ), rdir.z, org_rdir.z);
#else
const float8 tNearX = (norg.x + load8f((const char*)node+nearX)) * rdir.x;
const float8 tNearY = (norg.y + load8f((const char*)node+nearY)) * rdir.y;
const float8 tNearZ = (norg.z + load8f((const char*)node+nearZ)) * rdir.z;
const float8 tFarX = (norg.x + load8f((const char*)node+farX )) * rdir.x;
const float8 tFarY = (norg.y + load8f((const char*)node+farY )) * rdir.y;
const float8 tFarZ = (norg.z + load8f((const char*)node+farZ )) * rdir.z;
#endif
#if defined(__AVX2__)
const float8 tNear = maxi(maxi(tNearX,tNearY),maxi(tNearZ,ray_near));
const float8 tFar = mini(mini(tFarX ,tFarY ),mini(tFarZ ,ray_far ));
const bool8 vmask = cast(tNear) > cast(tFar);
size_t mask = movemask(vmask)^0xff;
#else
const float8 tNear = max(tNearX,tNearY,tNearZ,ray_near);
const float8 tFar = min(tFarX ,tFarY ,tFarZ ,ray_far);
const bool8 vmask = tNear <= tFar;
size_t mask = movemask(vmask);
#endif
/*! if no child is hit, pop next node */
if (unlikely(mask == 0))
goto pop;
/*! one child is hit, continue with that child */
size_t r = __bscf(mask);
if (likely(mask == 0)) {
cur = node->child(r); cur.prefetch();
assert(cur != BVH8::emptyNode);
continue;
}
/*! two children are hit, push far child, and continue with closer child */
NodeRef c0 = node->child(r); c0.prefetch(); const unsigned int d0 = ((unsigned int*)&tNear)[r];
r = __bscf(mask);
NodeRef c1 = node->child(r); c1.prefetch(); const unsigned int d1 = ((unsigned int*)&tNear)[r];
assert(c0 != BVH8::emptyNode);
assert(c1 != BVH8::emptyNode);
if (likely(mask == 0)) {
assert(stackPtr < stackEnd);
if (d0 < d1) { *stackPtr = c1; stackPtr++; cur = c0; continue; }
else { *stackPtr = c0; stackPtr++; cur = c1; continue; }
}
assert(stackPtr < stackEnd);
*stackPtr = c0; stackPtr++;
assert(stackPtr < stackEnd);
*stackPtr = c1; stackPtr++;
/*! three children are hit */
r = __bscf(mask);
cur = node->child(r); cur.prefetch(); *stackPtr = cur; stackPtr++;
if (likely(mask == 0)) {
stackPtr--;
continue;
}
/*! process more than three children */
while(1)
{
r = __bscf(mask);
NodeRef c = node->child(r); c.prefetch(); *stackPtr = c; stackPtr++;
if (unlikely(mask == 0)) break;
}
cur = (NodeRef) stackPtr[-1]; stackPtr--;
}
/*! this is a leaf node */
assert(cur != BVH8::emptyNode);
STAT3(shadow.trav_leaves,1,1,1);
size_t num; Primitive* prim = (Primitive*) cur.leaf(num);
size_t lazy_node = 0;
if (PrimitiveIntersector::occluded(pre,ray,prim,num,bvh->scene,lazy_node)) {
ray.geomID = 0;
break;
}
if (unlikely(lazy_node)) {
*stackPtr = (NodeRef)lazy_node;
stackPtr++;
}
}
AVX_ZERO_UPPER();
}
unsigned int *
mgcd(unsigned int *u, unsigned int *v)
{
int i, k, n;
unsigned int *t;
if (MZERO(u)) {
t = mcopy(v);
MSIGN(t) = 1;
return t;
}
if (MZERO(v)) {
t = mcopy(u);
MSIGN(t) = 1;
return t;
}
u = mcopy(u);
v = mcopy(v);
MSIGN(u) = 1;
MSIGN(v) = 1;
k = 0;
while ((u[0] & 1) == 0 && (v[0] & 1) == 0) {
mshiftright(u);
mshiftright(v);
k++;
}
if (u[0] & 1) {
t = mcopy(v);
MSIGN(t) *= -1;
} else
t = mcopy(u);
while (1) {
while ((t[0] & 1) == 0)
mshiftright(t);
if (MSIGN(t) == 1) {
mfree(u);
u = mcopy(t);
} else {
mfree(v);
v = mcopy(t);
MSIGN(v) *= -1;
}
mfree(t);
t = msub(u, v);
if (MZERO(t)) {
mfree(t);
mfree(v);
n = (k / 32) + 1;
v = mnew(n);
MSIGN(v) = 1;
MLENGTH(v) = n;
for (i = 0; i < n; i++)
v[i] = 0;
mp_set_bit(v, k);
t = mmul(u, v);
mfree(u);
mfree(v);
return t;
}
}
}
NOMIPS16 int main()
{
v2sf a, b, c, d, e, f;
float f1, f2;
f1 = 1.2;
f2 = 3.4;
a = init (f1, f2);
b = (v2sf) {1.2, 3.4};
if (!__builtin_mips_upper_c_eq_ps (a, b) ||
!__builtin_mips_lower_c_eq_ps (a, b))
abort ();
a = (v2sf) {1.2, 2.3};
b = (v2sf) {5.3, 6.1};
b = move (a);
if (!__builtin_mips_upper_c_eq_ps (a, b) ||
!__builtin_mips_lower_c_eq_ps (a, b))
abort ();
a = (v2sf) {1.2, 2.3};
b = (v2sf) {5.3, 6.1};
c = add (a, b);
d = (v2sf) {6.5, 8.4};
if (!__builtin_mips_upper_c_eq_ps (c, d) ||
!__builtin_mips_lower_c_eq_ps (c, d))
abort ();
a = (v2sf) {1, 12};
b = (v2sf) {5, 6};
c = sub (a, b);
d = (v2sf) {-4, 6};
if (!__builtin_mips_upper_c_eq_ps (c, d) ||
!__builtin_mips_lower_c_eq_ps (c, d))
abort ();
a = (v2sf) {1, 12};
b = (v2sf) {5, 6};
c = mul (a, b);
d = (v2sf) {5, 72};
if (!__builtin_mips_upper_c_eq_ps (c, d) ||
!__builtin_mips_lower_c_eq_ps (c, d))
abort ();
a = (v2sf) {1, 12};
b = (v2sf) {5, 6};
c = (v2sf) {5, 6};
d = madd (a, b, c);
e = (v2sf) {10, 78};
if (!__builtin_mips_upper_c_eq_ps (d, e) ||
!__builtin_mips_lower_c_eq_ps (d, e))
abort ();
a = (v2sf) {1, 12};
b = (v2sf) {5, 6};
c = (v2sf) {5, 6};
d = msub (a, b, c);
e = (v2sf) {0, 66};
if (!__builtin_mips_upper_c_eq_ps (d, e) ||
!__builtin_mips_lower_c_eq_ps (d, e))
abort ();
a = (v2sf) {1, 12};
b = (v2sf) {5, 6};
c = (v2sf) {5, 6};
d = nmadd (a, b, c);
e = (v2sf) {-10, -78};
if (!__builtin_mips_upper_c_eq_ps (d, e) ||
!__builtin_mips_lower_c_eq_ps (d, e))
abort ();
a = (v2sf) {1, 12};
b = (v2sf) {5, 6};
c = (v2sf) {5, 6};
d = nmsub (a, b, c);
e = (v2sf) {0, -66};
if (!__builtin_mips_upper_c_eq_ps (d, e) ||
!__builtin_mips_lower_c_eq_ps (d, e))
abort ();
a = (v2sf) {98, 12};
b = neg (a);
c = (v2sf) {-98, -12};
if (!__builtin_mips_upper_c_eq_ps (b, c) ||
!__builtin_mips_lower_c_eq_ps (b, c))
abort ();
a = (v2sf) {1, 12};
b = (v2sf) {5, 6};
c = cond_move1 (a, b, 1000);
if (!__builtin_mips_upper_c_eq_ps (c, a) ||
!__builtin_mips_lower_c_eq_ps (c, a))
abort ();
a = (v2sf) {1, 12};
b = (v2sf) {5, 6};
c = cond_move2 (a, b, -1000);
if (!__builtin_mips_upper_c_eq_ps (c, b) ||
!__builtin_mips_lower_c_eq_ps (c, b))
abort ();
a = (v2sf) {1, 12};
b = (v2sf) {5, 6};
c = cond_move3 (a, b, 9.0);
if (!__builtin_mips_upper_c_eq_ps (c, a) ||
!__builtin_mips_lower_c_eq_ps (c, a))
abort ();
a = (v2sf) {1, 12};
b = (v2sf) {5, 6};
c = cond_move4 (a, b, -10.0);
if (!__builtin_mips_upper_c_eq_ps (c, b) ||
!__builtin_mips_lower_c_eq_ps (c, b))
abort ();
a = (v2sf) {5, 12};
b = (v2sf) {5, 6};
c = (v2sf) {33, 123};
d = (v2sf) {8, 78};
e = __builtin_mips_movt_c_eq_ps (a, b, c, d);
f = (v2sf) {8, 123};
if (!__builtin_mips_upper_c_eq_ps (e, f) ||
!__builtin_mips_lower_c_eq_ps (e, f))
abort ();
a = (v2sf) {5, 12};
b = (v2sf) {5, 6};
c = (v2sf) {33, 123};
d = (v2sf) {8, 78};
e = __builtin_mips_movf_c_eq_ps (a, b, c, d);
f = (v2sf) {33, 78};
if (!__builtin_mips_upper_c_eq_ps (e, f) ||
!__builtin_mips_lower_c_eq_ps (e, f))
abort ();
a = load();
b = (v2sf) {100, 200};
if (!__builtin_mips_upper_c_eq_ps (a, b) ||
!__builtin_mips_lower_c_eq_ps (a, b))
abort ();
a = (v2sf) {123, 321};
store (a);
b = load();
if (!__builtin_mips_upper_c_eq_ps (a, b) ||
!__builtin_mips_lower_c_eq_ps (a, b))
abort ();
printf ("Test Passes\n");
exit (0);
}
void BVH8Intersector1<robust,PrimitiveIntersector>::intersect(const BVH8* bvh, Ray& ray)
{
/*! perform per ray precalculations required by the primitive intersector */
Precalculations pre(ray,bvh);
/*! stack state */
StackItemT<NodeRef> stack[stackSize]; //!< stack of nodes
StackItemT<NodeRef>* stackPtr = stack+1; //!< current stack pointer
StackItemT<NodeRef>* stackEnd = stack+stackSize;
stack[0].ptr = bvh->root;
stack[0].dist = neg_inf;
/* filter out invalid rays */
#if defined(RTCORE_IGNORE_INVALID_RAYS)
if (!ray.valid()) return;
#endif
/* verify correct input */
assert(ray.tnear > -FLT_MIN);
//assert(!(types & BVH4::FLAG_NODE_MB) || (ray.time >= 0.0f && ray.time <= 1.0f));
/*! load the ray into SIMD registers */
const Vec3f8 norg(-ray.org.x,-ray.org.y,-ray.org.z);
const Vec3fa ray_rdir = rcp_safe(ray.dir);
const Vec3f8 rdir(ray_rdir.x,ray_rdir.y,ray_rdir.z);
const Vec3fa ray_org_rdir = ray.org*ray_rdir;
const Vec3f8 org_rdir(ray_org_rdir.x,ray_org_rdir.y,ray_org_rdir.z);
const float8 ray_near(ray.tnear);
float8 ray_far(ray.tfar);
/*! offsets to select the side that becomes the lower or upper bound */
const size_t nearX = ray_rdir.x >= 0.0f ? 0*sizeof(float8) : 1*sizeof(float8);
const size_t nearY = ray_rdir.y >= 0.0f ? 2*sizeof(float8) : 3*sizeof(float8);
const size_t nearZ = ray_rdir.z >= 0.0f ? 4*sizeof(float8) : 5*sizeof(float8);
/* pop loop */
while (true) pop:
{
/*! pop next node */
if (unlikely(stackPtr == stack)) break;
stackPtr--;
NodeRef cur = NodeRef(stackPtr->ptr);
/*! if popped node is too far, pop next one */
if (unlikely(*(float*)&stackPtr->dist > ray.tfar))
continue;
/* downtraversal loop */
while (true)
{
/*! stop if we found a leaf */
if (unlikely(cur.isLeaf())) break;
STAT3(normal.trav_nodes,1,1,1);
/*! single ray intersection with 4 boxes */
const Node* node = cur.node();
const size_t farX = nearX ^ sizeof(float8), farY = nearY ^ sizeof(float8), farZ = nearZ ^ sizeof(float8);
#if defined (__AVX2__)
const float8 tNearX = msub(load8f((const char*)node+nearX), rdir.x, org_rdir.x);
const float8 tNearY = msub(load8f((const char*)node+nearY), rdir.y, org_rdir.y);
const float8 tNearZ = msub(load8f((const char*)node+nearZ), rdir.z, org_rdir.z);
const float8 tFarX = msub(load8f((const char*)node+farX ), rdir.x, org_rdir.x);
const float8 tFarY = msub(load8f((const char*)node+farY ), rdir.y, org_rdir.y);
const float8 tFarZ = msub(load8f((const char*)node+farZ ), rdir.z, org_rdir.z);
#else
const float8 tNearX = (norg.x + load8f((const char*)node+nearX)) * rdir.x;
const float8 tNearY = (norg.y + load8f((const char*)node+nearY)) * rdir.y;
const float8 tNearZ = (norg.z + load8f((const char*)node+nearZ)) * rdir.z;
const float8 tFarX = (norg.x + load8f((const char*)node+farX )) * rdir.x;
const float8 tFarY = (norg.y + load8f((const char*)node+farY )) * rdir.y;
const float8 tFarZ = (norg.z + load8f((const char*)node+farZ )) * rdir.z;
#endif
const float round_down = 1.0f-2.0f*float(ulp);
const float round_up = 1.0f+2.0f*float(ulp);
#if defined(__AVX2__)
const float8 tNear = maxi(maxi(tNearX,tNearY),maxi(tNearZ,ray_near));
const float8 tFar = mini(mini(tFarX ,tFarY ),mini(tFarZ ,ray_far ));
const bool8 vmask = robust ? (round_down*tNear > round_up*tFar) : cast(tNear) > cast(tFar);
size_t mask = movemask(vmask)^0xff;
#else
const float8 tNear = max(tNearX,tNearY,tNearZ,ray_near);
const float8 tFar = min(tFarX ,tFarY ,tFarZ ,ray_far);
const bool8 vmask = robust ? (round_down*tNear > round_up*tFar) : tNear <= tFar;
size_t mask = movemask(vmask);
#endif
/*! if no child is hit, pop next node */
if (unlikely(mask == 0))
goto pop;
/*! one child is hit, continue with that child */
size_t r = __bscf(mask);
if (likely(mask == 0)) {
cur = node->child(r); cur.prefetch();
assert(cur != BVH8::emptyNode);
continue;
}
/*! two children are hit, push far child, and continue with closer child */
NodeRef c0 = node->child(r); c0.prefetch(); const unsigned int d0 = ((unsigned int*)&tNear)[r];
r = __bscf(mask);
NodeRef c1 = node->child(r); c1.prefetch(); const unsigned int d1 = ((unsigned int*)&tNear)[r];
assert(c0 != BVH8::emptyNode);
assert(c1 != BVH8::emptyNode);
if (likely(mask == 0)) {
assert(stackPtr < stackEnd);
if (d0 < d1) { stackPtr->ptr = c1; stackPtr->dist = d1; stackPtr++; cur = c0; continue; }
else { stackPtr->ptr = c0; stackPtr->dist = d0; stackPtr++; cur = c1; continue; }
}
/*! Here starts the slow path for 3 or 4 hit children. We push
* all nodes onto the stack to sort them there. */
assert(stackPtr < stackEnd);
stackPtr->ptr = c0; stackPtr->dist = d0; stackPtr++;
assert(stackPtr < stackEnd);
stackPtr->ptr = c1; stackPtr->dist = d1; stackPtr++;
/*! three children are hit, push all onto stack and sort 3 stack items, continue with closest child */
assert(stackPtr < stackEnd);
r = __bscf(mask);
NodeRef c = node->child(r); c.prefetch(); unsigned int d = ((unsigned int*)&tNear)[r]; stackPtr->ptr = c; stackPtr->dist = d; stackPtr++;
assert(c != BVH8::emptyNode);
if (likely(mask == 0)) {
sort(stackPtr[-1],stackPtr[-2],stackPtr[-3]);
cur = (NodeRef) stackPtr[-1].ptr; stackPtr--;
continue;
}
/*! four children are hit, push all onto stack and sort 4 stack items, continue with closest child */
r = __bscf(mask);
c = node->child(r); c.prefetch(); d = *(unsigned int*)&tNear[r]; stackPtr->ptr = c; stackPtr->dist = d; stackPtr++;
if (likely(mask == 0)) {
sort(stackPtr[-1],stackPtr[-2],stackPtr[-3],stackPtr[-4]);
cur = (NodeRef) stackPtr[-1].ptr; stackPtr--;
continue;
}
/*! fallback case if more than 4 children are hit */
while (1)
{
r = __bscf(mask);
assert(stackPtr < stackEnd);
c = node->child(r); c.prefetch(); d = *(unsigned int*)&tNear[r]; stackPtr->ptr = c; stackPtr->dist = d; stackPtr++;
if (unlikely(mask == 0)) break;
}
cur = (NodeRef) stackPtr[-1].ptr; stackPtr--;
}
/*! this is a leaf node */
assert(cur != BVH8::emptyNode);
STAT3(normal.trav_leaves,1,1,1);
size_t num; Primitive* prim = (Primitive*) cur.leaf(num);
size_t lazy_node = 0;
PrimitiveIntersector::intersect(pre,ray,prim,num,bvh->scene,lazy_node);
ray_far = ray.tfar;
if (unlikely(lazy_node)) {
stackPtr->ptr = lazy_node;
stackPtr->dist = inf;
stackPtr++;
}
}
AVX_ZERO_UPPER();
}
void BVH4Intersector1<PrimitiveIntersector>::intersect(const BVH4* bvh, Ray& ray)
{
/*! stack state */
StackItemInt32<NodeRef> stack[stackSize]; //!< stack of nodes
StackItemInt32<NodeRef>* stackPtr = stack+1; //!< current stack pointer
StackItemInt32<NodeRef>* stackEnd = stack+stackSize;
stack[0].ptr = bvh->root;
stack[0].dist = neg_inf;
/*! offsets to select the side that becomes the lower or upper bound */
const size_t nearX = ray.dir.x >= 0.0f ? 0*sizeof(ssef) : 1*sizeof(ssef);
const size_t nearY = ray.dir.y >= 0.0f ? 2*sizeof(ssef) : 3*sizeof(ssef);
const size_t nearZ = ray.dir.z >= 0.0f ? 4*sizeof(ssef) : 5*sizeof(ssef);
#if 0 // FIXME: why is this slower
/*! load the ray */
Vec3fa ray_org = ray.org;
Vec3fa ray_dir = ray.dir;
ssef ray_near = max(ray.tnear,FLT_MIN); // we do not support negative tnear values in this kernel due to integer optimizations
ssef ray_far = ray.tfar;
#if defined(__FIX_RAYS__)
const float float_range = 0.1f*FLT_MAX;
ray_org = clamp(ray_org,Vec3fa(-float_range),Vec3fa(+float_range));
ray_dir = clamp(ray_dir,Vec3fa(-float_range),Vec3fa(+float_range));
ray_far = min(ray_far,float(inf));
#endif
const Vec3fa ray_rdir = rcp_safe(ray_dir);
const sse3f org(ray_org), dir(ray_dir);
const sse3f norg(-ray_org), rdir(ray_rdir), org_rdir(ray_org*ray_rdir);
#else
/*! load the ray into SIMD registers */
const sse3f norg(-ray.org.x,-ray.org.y,-ray.org.z);
const Vec3fa ray_rdir = rcp_safe(ray.dir);
const sse3f rdir(ray_rdir.x,ray_rdir.y,ray_rdir.z);
const Vec3fa ray_org_rdir = ray.org*ray_rdir;
const sse3f org_rdir(ray_org_rdir.x,ray_org_rdir.y,ray_org_rdir.z);
const ssef ray_near(ray.tnear);
ssef ray_far(ray.tfar);
#endif
/* pop loop */
while (true) pop:
{
/*! pop next node */
if (unlikely(stackPtr == stack)) break;
stackPtr--;
NodeRef cur = NodeRef(stackPtr->ptr);
/*! if popped node is too far, pop next one */
if (unlikely(*(float*)&stackPtr->dist > ray.tfar))
continue;
/* downtraversal loop */
while (true)
{
/*! stop if we found a leaf */
if (unlikely(cur.isLeaf())) break;
STAT3(normal.trav_nodes,1,1,1);
/*! single ray intersection with 4 boxes */
const Node* node = cur.node();
const size_t farX = nearX ^ 16, farY = nearY ^ 16, farZ = nearZ ^ 16;
#if defined (__AVX2__)
const ssef tNearX = msub(load4f((const char*)node+nearX), rdir.x, org_rdir.x);
const ssef tNearY = msub(load4f((const char*)node+nearY), rdir.y, org_rdir.y);
const ssef tNearZ = msub(load4f((const char*)node+nearZ), rdir.z, org_rdir.z);
const ssef tFarX = msub(load4f((const char*)node+farX ), rdir.x, org_rdir.x);
const ssef tFarY = msub(load4f((const char*)node+farY ), rdir.y, org_rdir.y);
const ssef tFarZ = msub(load4f((const char*)node+farZ ), rdir.z, org_rdir.z);
#else
const ssef tNearX = (norg.x + load4f((const char*)node+nearX)) * rdir.x;
const ssef tNearY = (norg.y + load4f((const char*)node+nearY)) * rdir.y;
const ssef tNearZ = (norg.z + load4f((const char*)node+nearZ)) * rdir.z;
const ssef tFarX = (norg.x + load4f((const char*)node+farX )) * rdir.x;
const ssef tFarY = (norg.y + load4f((const char*)node+farY )) * rdir.y;
const ssef tFarZ = (norg.z + load4f((const char*)node+farZ )) * rdir.z;
#endif
#if defined(__SSE4_1__)
const ssef tNear = maxi(maxi(tNearX,tNearY),maxi(tNearZ,ray_near));
const ssef tFar = mini(mini(tFarX ,tFarY ),mini(tFarZ ,ray_far ));
const sseb vmask = cast(tNear) > cast(tFar);
size_t mask = movemask(vmask)^0xf;
#else
const ssef tNear = max(tNearX,tNearY,tNearZ,ray_near);
const ssef tFar = min(tFarX ,tFarY ,tFarZ ,ray_far);
const sseb vmask = tNear <= tFar;
size_t mask = movemask(vmask);
#endif
/*! if no child is hit, pop next node */
if (unlikely(mask == 0))
goto pop;
/*! one child is hit, continue with that child */
size_t r = __bscf(mask);
if (likely(mask == 0)) {
cur = node->child(r);
assert(cur != BVH4::emptyNode);
continue;
}
/*! two children are hit, push far child, and continue with closer child */
NodeRef c0 = node->child(r); const unsigned int d0 = ((unsigned int*)&tNear)[r];
r = __bscf(mask);
NodeRef c1 = node->child(r); const unsigned int d1 = ((unsigned int*)&tNear)[r];
assert(c0 != BVH4::emptyNode);
assert(c1 != BVH4::emptyNode);
if (likely(mask == 0)) {
assert(stackPtr < stackEnd);
if (d0 < d1) { stackPtr->ptr = c1; stackPtr->dist = d1; stackPtr++; cur = c0; continue; }
else { stackPtr->ptr = c0; stackPtr->dist = d0; stackPtr++; cur = c1; continue; }
}
/*! Here starts the slow path for 3 or 4 hit children. We push
* all nodes onto the stack to sort them there. */
assert(stackPtr < stackEnd);
stackPtr->ptr = c0; stackPtr->dist = d0; stackPtr++;
assert(stackPtr < stackEnd);
stackPtr->ptr = c1; stackPtr->dist = d1; stackPtr++;
/*! three children are hit, push all onto stack and sort 3 stack items, continue with closest child */
assert(stackPtr < stackEnd);
r = __bscf(mask);
NodeRef c = node->child(r); unsigned int d = ((unsigned int*)&tNear)[r]; stackPtr->ptr = c; stackPtr->dist = d; stackPtr++;
assert(c != BVH4::emptyNode);
if (likely(mask == 0)) {
sort(stackPtr[-1],stackPtr[-2],stackPtr[-3]);
cur = (NodeRef) stackPtr[-1].ptr; stackPtr--;
continue;
}
/*! four children are hit, push all onto stack and sort 4 stack items, continue with closest child */
assert(stackPtr < stackEnd);
r = __bscf(mask);
c = node->child(r); d = *(unsigned int*)&tNear[r]; stackPtr->ptr = c; stackPtr->dist = d; stackPtr++;
assert(c != BVH4::emptyNode);
sort(stackPtr[-1],stackPtr[-2],stackPtr[-3],stackPtr[-4]);
cur = (NodeRef) stackPtr[-1].ptr; stackPtr--;
}
/*! this is a leaf node */
STAT3(normal.trav_leaves,1,1,1);
size_t num; Primitive* prim = (Primitive*) cur.leaf(num);
PrimitiveIntersector::intersect(ray,prim,num,bvh->geometry);
ray_far = ray.tfar;
}
}
bool BVH4iIntersector1<TriangleIntersector>::occluded(const BVH4iIntersector1* This, Ray& ray)
{
AVX_ZERO_UPPER();
STAT3(shadow.travs,1,1,1);
/*! stack state */
const BVH4i* bvh = This->bvh;
NodeRef stack[1+3*BVH4i::maxDepth]; //!< stack of nodes that still need to get traversed
NodeRef* stackPtr = stack+1; //!< current stack pointer
stack[0] = bvh->root;
/*! offsets to select the side that becomes the lower or upper bound */
const size_t nearX = ray.dir.x >= 0 ? 0*sizeof(ssef_m) : 1*sizeof(ssef_m);
const size_t nearY = ray.dir.y >= 0 ? 2*sizeof(ssef_m) : 3*sizeof(ssef_m);
const size_t nearZ = ray.dir.z >= 0 ? 4*sizeof(ssef_m) : 5*sizeof(ssef_m);
/*! load the ray into SIMD registers */
const sse3f norg(-ray.org.x,-ray.org.y,-ray.org.z);
const Vector3f ray_rdir = rcp_safe(ray.dir);
const sse3f rdir(ray_rdir.x,ray_rdir.y,ray_rdir.z);
const Vector3f ray_org_rdir = ray.org*ray_rdir;
const sse3f org_rdir(ray_org_rdir.x,ray_org_rdir.y,ray_org_rdir.z);
const ssef rayNear(ray.tnear);
const ssef rayFar(ray.tfar);
const void* nodePtr = bvh->nodePtr();
const void* triPtr = bvh->triPtr();
/* pop loop */
while (true) pop:
{
/*! pop next node */
if (unlikely(stackPtr == stack)) break;
stackPtr--;
NodeRef cur = (NodeRef) *stackPtr;
/* downtraversal loop */
while (true)
{
/*! stop if we found a leaf */
if (unlikely(cur.isLeaf())) break;
STAT3(shadow.trav_nodes,1,1,1);
/*! single ray intersection with 4 boxes */
const Node* node = cur.node(nodePtr);
const size_t farX = nearX ^ 16, farY = nearY ^ 16, farZ = nearZ ^ 16;
#if defined (__AVX2__)
const ssef tNearX = msub(ssef((const char*)nodePtr+(size_t)cur+nearX), rdir.x, org_rdir.x);
const ssef tNearY = msub(ssef((const char*)nodePtr+(size_t)cur+nearY), rdir.y, org_rdir.y);
const ssef tNearZ = msub(ssef((const char*)nodePtr+(size_t)cur+nearZ), rdir.z, org_rdir.z);
const ssef tFarX = msub(ssef((const char*)nodePtr+(size_t)cur+farX ), rdir.x, org_rdir.x);
const ssef tFarY = msub(ssef((const char*)nodePtr+(size_t)cur+farY ), rdir.y, org_rdir.y);
const ssef tFarZ = msub(ssef((const char*)nodePtr+(size_t)cur+farZ ), rdir.z, org_rdir.z);
#else
const ssef tNearX = (norg.x + ssef((const char*)nodePtr+(size_t)cur+nearX)) * rdir.x;
const ssef tNearY = (norg.y + ssef((const char*)nodePtr+(size_t)cur+nearY)) * rdir.y;
const ssef tNearZ = (norg.z + ssef((const char*)nodePtr+(size_t)cur+nearZ)) * rdir.z;
const ssef tFarX = (norg.x + ssef((const char*)nodePtr+(size_t)cur+farX )) * rdir.x;
const ssef tFarY = (norg.y + ssef((const char*)nodePtr+(size_t)cur+farY )) * rdir.y;
const ssef tFarZ = (norg.z + ssef((const char*)nodePtr+(size_t)cur+farZ )) * rdir.z;
#endif
const ssef tNear = max(tNearX,tNearY,tNearZ,rayNear);
const ssef tFar = min(tFarX ,tFarY ,tFarZ ,rayFar);
size_t mask = movemask(tNear <= tFar);
/*! if no child is hit, pop next node */
if (unlikely(mask == 0))
goto pop;
/*! one child is hit, continue with that child */
size_t r = __bsf(mask); mask = __btc(mask,r);
if (likely(mask == 0)) {
cur = node->child(r);
continue;
}
/*! two children are hit, push far child, and continue with closer child */
NodeRef c0 = node->child(r); const float d0 = tNear[r];
r = __bsf(mask); mask = __btc(mask,r);
NodeRef c1 = node->child(r); const float d1 = tNear[r];
if (likely(mask == 0)) {
if (d0 < d1) { *stackPtr = c1; stackPtr++; cur = c0; continue; }
else { *stackPtr = c0; stackPtr++; cur = c1; continue; }
}
*stackPtr = c0; stackPtr++;
*stackPtr = c1; stackPtr++;
/*! three children are hit */
r = __bsf(mask); mask = __btc(mask,r);
cur = node->child(r); *stackPtr = cur; stackPtr++;
if (likely(mask == 0)) {
stackPtr--;
continue;
}
/*! four children are hit */
cur = node->child(3);
}
/*! this is a leaf node */
STAT3(shadow.trav_leaves,1,1,1);
size_t num; Triangle* tri = (Triangle*) cur.leaf(triPtr,num);
for (size_t i=0; i<num; i++) {
if (TriangleIntersector::occluded(ray,tri[i],bvh->vertices)) {
AVX_ZERO_UPPER();
return true;
}
}
}
AVX_ZERO_UPPER();
return false;
}
void BVH4mbIntersector16Hybrid<LeafIntersector>::intersect(int16* valid_i, BVH4mb* bvh, Ray16& ray16)
{
/* near and node stack */
__aligned(64) float16 stack_dist[3*BVH4i::maxDepth+1];
__aligned(64) NodeRef stack_node[3*BVH4i::maxDepth+1];
__aligned(64) NodeRef stack_node_single[3*BVH4i::maxDepth+1];
/* load ray */
const bool16 valid0 = *(int16*)valid_i != int16(0);
const Vec3f16 rdir16 = rcp_safe(ray16.dir);
const Vec3f16 org_rdir16 = ray16.org * rdir16;
float16 ray_tnear = select(valid0,ray16.tnear,pos_inf);
float16 ray_tfar = select(valid0,ray16.tfar ,neg_inf);
const float16 inf = float16(pos_inf);
/* allocate stack and push root node */
stack_node[0] = BVH4i::invalidNode;
stack_dist[0] = inf;
stack_node[1] = bvh->root;
stack_dist[1] = ray_tnear;
NodeRef* __restrict__ sptr_node = stack_node + 2;
float16* __restrict__ sptr_dist = stack_dist + 2;
const Node * __restrict__ nodes = (Node *)bvh->nodePtr();
const BVH4mb::Triangle01 * __restrict__ accel = (BVH4mb::Triangle01 *)bvh->triPtr();
while (1) pop:
{
/* pop next node from stack */
NodeRef curNode = *(sptr_node-1);
float16 curDist = *(sptr_dist-1);
sptr_node--;
sptr_dist--;
const bool16 m_stackDist = ray_tfar > curDist;
/* stack emppty ? */
if (unlikely(curNode == BVH4i::invalidNode)) break;
/* cull node if behind closest hit point */
if (unlikely(none(m_stackDist))) continue;
///////////////////////////////////////////////////////////////////////////////////////////////////////////////
///////////////////////////////////////////////////////////////////////////////////////////////////////////////
///////////////////////////////////////////////////////////////////////////////////////////////////////////////
/* switch to single ray mode */
if (unlikely(countbits(m_stackDist) <= BVH4i::hybridSIMDUtilSwitchThreshold))
{
float *__restrict__ stack_dist_single = (float*)sptr_dist;
store16f(stack_dist_single,inf);
/* traverse single ray */
long rayIndex = -1;
while((rayIndex = bitscan64(rayIndex,m_stackDist)) != BITSCAN_NO_BIT_SET_64)
{
stack_node_single[0] = BVH4i::invalidNode;
stack_node_single[1] = curNode;
size_t sindex = 2;
const float16 org_xyz = loadAOS4to16f(rayIndex,ray16.org.x,ray16.org.y,ray16.org.z);
const float16 dir_xyz = loadAOS4to16f(rayIndex,ray16.dir.x,ray16.dir.y,ray16.dir.z);
const float16 rdir_xyz = loadAOS4to16f(rayIndex,rdir16.x,rdir16.y,rdir16.z);
const float16 org_rdir_xyz = org_xyz * rdir_xyz;
const float16 min_dist_xyz = broadcast1to16f(&ray16.tnear[rayIndex]);
float16 max_dist_xyz = broadcast1to16f(&ray16.tfar[rayIndex]);
const float16 time = broadcast1to16f(&ray16.time[rayIndex]);
const unsigned int leaf_mask = BVH4I_LEAF_MASK;
while (1)
{
NodeRef curNode = stack_node_single[sindex-1];
sindex--;
traverse_single_intersect(curNode,
sindex,
rdir_xyz,
org_rdir_xyz,
min_dist_xyz,
max_dist_xyz,
time,
stack_node_single,
stack_dist_single,
nodes,
leaf_mask);
/* return if stack is empty */
if (unlikely(curNode == BVH4i::invalidNode)) break;
/* intersect one ray against four triangles */
const bool hit = LeafIntersector::intersect(curNode,
rayIndex,
dir_xyz,
org_xyz,
min_dist_xyz,
max_dist_xyz,
ray16,
accel,
(Scene*)bvh->geometry);
if (hit)
compactStack(stack_node_single,stack_dist_single,sindex,max_dist_xyz);
}
}
ray_tfar = select(valid0,ray16.tfar ,neg_inf);
continue;
}
///////////////////////////////////////////////////////////////////////////////////////////////////////////////
///////////////////////////////////////////////////////////////////////////////////////////////////////////////
///////////////////////////////////////////////////////////////////////////////////////////////////////////////
const unsigned int leaf_mask = BVH4I_LEAF_MASK;
const float16 time = ray16.time;
const float16 one_time = (float16::one() - time);
while (1)
{
/* test if this is a leaf node */
if (unlikely(curNode.isLeaf(leaf_mask))) break;
STAT3(normal.trav_nodes,1,popcnt(ray_tfar > curDist),16);
const Node* __restrict__ const node = curNode.node(nodes);
const BVH4mb::Node* __restrict__ const nodeMB = (BVH4mb::Node*)node;
/* pop of next node */
sptr_node--;
sptr_dist--;
curNode = *sptr_node;
curDist = *sptr_dist;
prefetch<PFHINT_L1>((char*)node + 0*64);
prefetch<PFHINT_L1>((char*)node + 1*64);
prefetch<PFHINT_L1>((char*)node + 2*64);
prefetch<PFHINT_L1>((char*)node + 3*64);
#pragma unroll(4)
for (unsigned int i=0; i<4; i++)
{
const NodeRef child = node->lower[i].child;
const float16 lower_x = one_time * nodeMB->lower[i].x + time * nodeMB->lower_t1[i].x;
const float16 lower_y = one_time * nodeMB->lower[i].y + time * nodeMB->lower_t1[i].y;
const float16 lower_z = one_time * nodeMB->lower[i].z + time * nodeMB->lower_t1[i].z;
const float16 upper_x = one_time * nodeMB->upper[i].x + time * nodeMB->upper_t1[i].x;
const float16 upper_y = one_time * nodeMB->upper[i].y + time * nodeMB->upper_t1[i].y;
const float16 upper_z = one_time * nodeMB->upper[i].z + time * nodeMB->upper_t1[i].z;
if (unlikely(i >=2 && child == BVH4i::invalidNode)) break;
const float16 lclipMinX = msub(lower_x,rdir16.x,org_rdir16.x);
const float16 lclipMinY = msub(lower_y,rdir16.y,org_rdir16.y);
const float16 lclipMinZ = msub(lower_z,rdir16.z,org_rdir16.z);
const float16 lclipMaxX = msub(upper_x,rdir16.x,org_rdir16.x);
const float16 lclipMaxY = msub(upper_y,rdir16.y,org_rdir16.y);
const float16 lclipMaxZ = msub(upper_z,rdir16.z,org_rdir16.z);
const float16 lnearP = max(max(min(lclipMinX, lclipMaxX), min(lclipMinY, lclipMaxY)), min(lclipMinZ, lclipMaxZ));
const float16 lfarP = min(min(max(lclipMinX, lclipMaxX), max(lclipMinY, lclipMaxY)), max(lclipMinZ, lclipMaxZ));
const bool16 lhit = max(lnearP,ray_tnear) <= min(lfarP,ray_tfar);
const float16 childDist = select(lhit,lnearP,inf);
const bool16 m_child_dist = childDist < curDist;
/* if we hit the child we choose to continue with that child if it
is closer than the current next child, or we push it onto the stack */
if (likely(any(lhit)))
{
sptr_node++;
sptr_dist++;
/* push cur node onto stack and continue with hit child */
if (any(m_child_dist))
{
*(sptr_node-1) = curNode;
*(sptr_dist-1) = curDist;
curDist = childDist;
curNode = child;
}
/* push hit child onto stack*/
else
{
*(sptr_node-1) = child;
*(sptr_dist-1) = childDist;
}
assert(sptr_node - stack_node < BVH4i::maxDepth);
}
}
#if SWITCH_ON_DOWN_TRAVERSAL == 1
const bool16 curUtil = ray_tfar > curDist;
if (unlikely(countbits(curUtil) <= BVH4i::hybridSIMDUtilSwitchThreshold))
{
*sptr_node++ = curNode;
*sptr_dist++ = curDist;
goto pop;
}
#endif
}
/* return if stack is empty */
if (unlikely(curNode == BVH4i::invalidNode)) break;
/* intersect leaf */
const bool16 m_valid_leaf = ray_tfar > curDist;
STAT3(normal.trav_leaves,1,popcnt(m_valid_leaf),16);
LeafIntersector::intersect16(curNode,
m_valid_leaf,
ray16.dir,
ray16.org,
ray16,
accel,
(Scene*)bvh->geometry);
ray_tfar = select(m_valid_leaf,ray16.tfar,ray_tfar);
}
void BVH4Intersector1<PrimitiveIntersector>::occluded(const BVH4* bvh, Ray& ray)
{
/*! stack state */
NodeRef stack[stackSize]; //!< stack of nodes that still need to get traversed
NodeRef* stackPtr = stack+1; //!< current stack pointer
NodeRef* stackEnd = stack+stackSize;
stack[0] = bvh->root;
/*! offsets to select the side that becomes the lower or upper bound */
const size_t nearX = ray.dir.x >= 0 ? 0*sizeof(ssef) : 1*sizeof(ssef);
const size_t nearY = ray.dir.y >= 0 ? 2*sizeof(ssef) : 3*sizeof(ssef);
const size_t nearZ = ray.dir.z >= 0 ? 4*sizeof(ssef) : 5*sizeof(ssef);
#if 0 // FIXME: why is this slower
/*! load the ray */
Vec3fa ray_org = ray.org;
Vec3fa ray_dir = ray.dir;
ssef ray_near = max(ray.tnear,FLT_MIN); // we do not support negative tnear values in this kernel due to integer optimizations
ssef ray_far = ray.tfar;
#if defined(__FIX_RAYS__)
const float float_range = 0.1f*FLT_MAX;
ray_org = clamp(ray_org,Vec3fa(-float_range),Vec3fa(+float_range));
ray_dir = clamp(ray_dir,Vec3fa(-float_range),Vec3fa(+float_range));
ray_far = min(ray_far,float(inf));
#endif
const Vec3fa ray_rdir = rcp_safe(ray_dir);
const sse3f org(ray_org), dir(ray_dir);
const sse3f norg(-ray_org), rdir(ray_rdir), org_rdir(ray_org*ray_rdir);
#else
/*! load the ray into SIMD registers */
const sse3f norg(-ray.org.x,-ray.org.y,-ray.org.z);
const Vec3fa ray_rdir = rcp_safe(ray.dir);
const sse3f rdir(ray_rdir.x,ray_rdir.y,ray_rdir.z);
const Vec3fa ray_org_rdir = ray.org*ray_rdir;
const sse3f org_rdir(ray_org_rdir.x,ray_org_rdir.y,ray_org_rdir.z);
const ssef ray_near(ray.tnear);
ssef ray_far(ray.tfar);
#endif
/* pop loop */
while (true) pop:
{
/*! pop next node */
if (unlikely(stackPtr == stack)) break;
stackPtr--;
NodeRef cur = (NodeRef) *stackPtr;
/* downtraversal loop */
while (true)
{
/*! stop if we found a leaf */
if (unlikely(cur.isLeaf())) break;
STAT3(shadow.trav_nodes,1,1,1);
/*! single ray intersection with 4 boxes */
const Node* node = cur.node();
const size_t farX = nearX ^ 16, farY = nearY ^ 16, farZ = nearZ ^ 16;
#if defined (__AVX2__)
const ssef tNearX = msub(load4f((const char*)node+nearX), rdir.x, org_rdir.x);
const ssef tNearY = msub(load4f((const char*)node+nearY), rdir.y, org_rdir.y);
const ssef tNearZ = msub(load4f((const char*)node+nearZ), rdir.z, org_rdir.z);
const ssef tFarX = msub(load4f((const char*)node+farX ), rdir.x, org_rdir.x);
const ssef tFarY = msub(load4f((const char*)node+farY ), rdir.y, org_rdir.y);
const ssef tFarZ = msub(load4f((const char*)node+farZ ), rdir.z, org_rdir.z);
#else
const ssef tNearX = (norg.x + load4f((const char*)node+nearX)) * rdir.x;
const ssef tNearY = (norg.y + load4f((const char*)node+nearY)) * rdir.y;
const ssef tNearZ = (norg.z + load4f((const char*)node+nearZ)) * rdir.z;
const ssef tFarX = (norg.x + load4f((const char*)node+farX )) * rdir.x;
const ssef tFarY = (norg.y + load4f((const char*)node+farY )) * rdir.y;
const ssef tFarZ = (norg.z + load4f((const char*)node+farZ )) * rdir.z;
#endif
#if defined(__SSE4_1__)
const ssef tNear = maxi(maxi(tNearX,tNearY),maxi(tNearZ,ray_near));
const ssef tFar = mini(mini(tFarX ,tFarY ),mini(tFarZ ,ray_far ));
const sseb vmask = cast(tNear) > cast(tFar);
size_t mask = movemask(vmask)^0xf;
#else
const ssef tNear = max(tNearX,tNearY,tNearZ,ray_near);
const ssef tFar = min(tFarX ,tFarY ,tFarZ ,ray_far);
const sseb vmask = tNear <= tFar;
size_t mask = movemask(vmask);
#endif
/*! if no child is hit, pop next node */
if (unlikely(mask == 0))
goto pop;
/*! one child is hit, continue with that child */
size_t r = __bscf(mask);
if (likely(mask == 0)) {
cur = node->child(r);
assert(cur != BVH4::emptyNode);
continue;
}
/*! two children are hit, push far child, and continue with closer child */
NodeRef c0 = node->child(r); const unsigned int d0 = ((unsigned int*)&tNear)[r];
r = __bscf(mask);
NodeRef c1 = node->child(r); const unsigned int d1 = ((unsigned int*)&tNear)[r];
assert(c0 != BVH4::emptyNode);
assert(c1 != BVH4::emptyNode);
if (likely(mask == 0)) {
assert(stackPtr < stackEnd);
if (d0 < d1) { *stackPtr = c1; stackPtr++; cur = c0; continue; }
else { *stackPtr = c0; stackPtr++; cur = c1; continue; }
}
assert(stackPtr < stackEnd);
*stackPtr = c0; stackPtr++;
assert(stackPtr < stackEnd);
*stackPtr = c1; stackPtr++;
/*! three children are hit */
r = __bscf(mask);
cur = node->child(r);
assert(cur != BVH4::emptyNode);
if (likely(mask == 0)) continue;
assert(stackPtr < stackEnd);
*stackPtr = cur; stackPtr++;
/*! four children are hit */
cur = node->child(3);
assert(cur != BVH4::emptyNode);
}
/*! this is a leaf node */
STAT3(shadow.trav_leaves,1,1,1);
size_t num; Primitive* prim = (Primitive*) cur.leaf(num);
if (PrimitiveIntersector::occluded(ray,prim,num,bvh->geometry)) {
ray.geomID = 0;
break;
}
}
AVX_ZERO_UPPER();
}