Page MenuHome
Differential D2828 Diff 9244 intern/cycles/kernel/bvh/bvh_nodes.h

Changeset View

Standalone View

View Options

intern/cycles/kernel/bvh/bvh_nodes.h

Show First 20 Lines • Show All 46 Lines • ▼ Show 20 Linesccl_device_forceinline int bvh_aligned_node_intersect(KernelGlobals *kg,
/* intersect ray against child nodes */ /* intersect ray against child nodes */
float c0lox = (node0.x - P.x) * idir.x; float c0lox = (node0.x - P.x) * idir.x;
float c0hix = (node0.z - P.x) * idir.x; float c0hix = (node0.z - P.x) * idir.x;
float c0loy = (node1.x - P.y) * idir.y; float c0loy = (node1.x - P.y) * idir.y;
float c0hiy = (node1.z - P.y) * idir.y; float c0hiy = (node1.z - P.y) * idir.y;
float c0loz = (node2.x - P.z) * idir.z; float c0loz = (node2.x - P.z) * idir.z;
float c0hiz = (node2.z - P.z) * idir.z; float c0hiz = (node2.z - P.z) * idir.z;
float c0min = max4(min(c0lox, c0hix), min(c0loy, c0hiy), min(c0loz, c0hiz), 0.0f); float c0min = max4(0.0f, min(c0lox, c0hix), min(c0loy, c0hiy), min(c0loz, c0hiz));
float c0max = min4(max(c0lox, c0hix), max(c0loy, c0hiy), max(c0loz, c0hiz), t); float c0max = min4(t, max(c0lox, c0hix), max(c0loy, c0hiy), max(c0loz, c0hiz));
float c1lox = (node0.y - P.x) * idir.x; float c1lox = (node0.y - P.x) * idir.x;
float c1hix = (node0.w - P.x) * idir.x; float c1hix = (node0.w - P.x) * idir.x;
float c1loy = (node1.y - P.y) * idir.y; float c1loy = (node1.y - P.y) * idir.y;
float c1hiy = (node1.w - P.y) * idir.y; float c1hiy = (node1.w - P.y) * idir.y;
float c1loz = (node2.y - P.z) * idir.z; float c1loz = (node2.y - P.z) * idir.z;
float c1hiz = (node2.w - P.z) * idir.z; float c1hiz = (node2.w - P.z) * idir.z;
float c1min = max4(min(c1lox, c1hix), min(c1loy, c1hiy), min(c1loz, c1hiz), 0.0f); float c1min = max4(0.0f, min(c1lox, c1hix), min(c1loy, c1hiy), min(c1loz, c1hiz));
float c1max = min4(max(c1lox, c1hix), max(c1loy, c1hiy), max(c1loz, c1hiz), t); float c1max = min4(t, max(c1lox, c1hix), max(c1loy, c1hiy), max(c1loz, c1hiz));
dist[0] = c0min; dist[0] = c0min;
dist[1] = c1min; dist[1] = c1min;
#ifdef __VISIBILITY_FLAG__ #ifdef __VISIBILITY_FLAG__
/* this visibility test gives a 5% performance hit, how to solve? */ /* this visibility test gives a 5% performance hit, how to solve? */
return (((c0max >= c0min) && (__float_as_uint(cnodes.x) & visibility))? 1: 0) | return (((c0max >= c0min) && (__float_as_uint(cnodes.x) & visibility))? 1: 0) |
(((c1max >= c1min) && (__float_as_uint(cnodes.y) & visibility))? 2: 0); (((c1max >= c1min) && (__float_as_uint(cnodes.y) & visibility))? 2: 0);
Show All 22 Linesccl_device_forceinline int bvh_aligned_node_intersect_robust(KernelGlobals *kg,
/* intersect ray against child nodes */ /* intersect ray against child nodes */
float c0lox = (node0.x - P.x) * idir.x; float c0lox = (node0.x - P.x) * idir.x;
float c0hix = (node0.z - P.x) * idir.x; float c0hix = (node0.z - P.x) * idir.x;
float c0loy = (node1.x - P.y) * idir.y; float c0loy = (node1.x - P.y) * idir.y;
float c0hiy = (node1.z - P.y) * idir.y; float c0hiy = (node1.z - P.y) * idir.y;
float c0loz = (node2.x - P.z) * idir.z; float c0loz = (node2.x - P.z) * idir.z;
float c0hiz = (node2.z - P.z) * idir.z; float c0hiz = (node2.z - P.z) * idir.z;
float c0min = max4(min(c0lox, c0hix), min(c0loy, c0hiy), min(c0loz, c0hiz), 0.0f); float c0min = max4(0.0f, min(c0lox, c0hix), min(c0loy, c0hiy), min(c0loz, c0hiz));
float c0max = min4(max(c0lox, c0hix), max(c0loy, c0hiy), max(c0loz, c0hiz), t); float c0max = min4(t, max(c0lox, c0hix), max(c0loy, c0hiy), max(c0loz, c0hiz));
float c1lox = (node0.y - P.x) * idir.x; float c1lox = (node0.y - P.x) * idir.x;
float c1hix = (node0.w - P.x) * idir.x; float c1hix = (node0.w - P.x) * idir.x;
float c1loy = (node1.y - P.y) * idir.y; float c1loy = (node1.y - P.y) * idir.y;
float c1hiy = (node1.w - P.y) * idir.y; float c1hiy = (node1.w - P.y) * idir.y;
float c1loz = (node2.y - P.z) * idir.z; float c1loz = (node2.y - P.z) * idir.z;
float c1hiz = (node2.w - P.z) * idir.z; float c1hiz = (node2.w - P.z) * idir.z;
float c1min = max4(min(c1lox, c1hix), min(c1loy, c1hiy), min(c1loz, c1hiz), 0.0f); float c1min = max4(0.0f, min(c1lox, c1hix), min(c1loy, c1hiy), min(c1loz, c1hiz));
float c1max = min4(max(c1lox, c1hix), max(c1loy, c1hiy), max(c1loz, c1hiz), t); float c1max = min4(t, max(c1lox, c1hix), max(c1loy, c1hiy), max(c1loz, c1hiz));
if(difl != 0.0f) { if(difl != 0.0f) {
float hdiff = 1.0f + difl; float hdiff = 1.0f + difl;
float ldiff = 1.0f - difl; float ldiff = 1.0f - difl;
if(__float_as_int(cnodes.z) & PATH_RAY_CURVE) { if(__float_as_int(cnodes.z) & PATH_RAY_CURVE) {
c0min = max(ldiff * c0min, c0min - extmax); c0min = max(ldiff * c0min, c0min - extmax);
c0max = min(hdiff * c0max, c0max + extmax); c0max = min(hdiff * c0max, c0max + extmax);
} }
▲ Show 20 Lines • Show All 355 Lines • ▼ Show 20 Linesccl_device_forceinline int bvh_unaligned_node_intersect(KernelGlobals *kg,
ssef tnear_x = min(lower_x, upper_x); ssef tnear_x = min(lower_x, upper_x);
ssef tnear_y = min(lower_y, upper_y); ssef tnear_y = min(lower_y, upper_y);
ssef tnear_z = min(lower_z, upper_z); ssef tnear_z = min(lower_z, upper_z);
ssef tfar_x = max(lower_x, upper_x); ssef tfar_x = max(lower_x, upper_x);
ssef tfar_y = max(lower_y, upper_y); ssef tfar_y = max(lower_y, upper_y);
ssef tfar_z = max(lower_z, upper_z); ssef tfar_z = max(lower_z, upper_z);
const ssef tnear = max4(tnear_x, tnear_y, tnear_z, isect_near); const ssef tnear = max4(isect_near, tnear_x, tnear_y, tnear_z);
const ssef tfar = min4(tfar_x, tfar_y, tfar_z, isect_far); const ssef tfar = min4(isect_far, tfar_x, tfar_y, tfar_z);
sseb vmask = tnear <= tfar; sseb vmask = tnear <= tfar;
dist[0] = tnear.f[0]; dist[0] = tnear.f[0];
dist[1] = tnear.f[1]; dist[1] = tnear.f[1];
int mask = (int)movemask(vmask); int mask = (int)movemask(vmask);
# ifdef __VISIBILITY_FLAG__ # ifdef __VISIBILITY_FLAG__
/* this visibility test gives a 5% performance hit, how to solve? */ /* this visibility test gives a 5% performance hit, how to solve? */
▲ Show 20 Lines • Show All 44 Lines • ▼ Show 20 Linesccl_device_forceinline int bvh_unaligned_node_intersect_robust(KernelGlobals *kg,
ssef tnear_x = min(lower_x, upper_x); ssef tnear_x = min(lower_x, upper_x);
ssef tnear_y = min(lower_y, upper_y); ssef tnear_y = min(lower_y, upper_y);
ssef tnear_z = min(lower_z, upper_z); ssef tnear_z = min(lower_z, upper_z);
ssef tfar_x = max(lower_x, upper_x); ssef tfar_x = max(lower_x, upper_x);
ssef tfar_y = max(lower_y, upper_y); ssef tfar_y = max(lower_y, upper_y);
ssef tfar_z = max(lower_z, upper_z); ssef tfar_z = max(lower_z, upper_z);
const ssef tnear = max4(tnear_x, tnear_y, tnear_z, isect_near); const ssef tnear = max4(isect_near, tnear_x, tnear_y, tnear_z);
const ssef tfar = min4(tfar_x, tfar_y, tfar_z, isect_far); const ssef tfar = min4(isect_far, tfar_x, tfar_y, tfar_z);
sseb vmask; sseb vmask;
if(difl != 0.0f) { if(difl != 0.0f) {
const float round_down = 1.0f - difl; const float round_down = 1.0f - difl;
const float round_up = 1.0f + difl; const float round_up = 1.0f + difl;
vmask = round_down*tnear <= round_up*tfar; vmask = round_down*tnear <= round_up*tfar;
} }
else { else {
vmask = tnear <= tfar; vmask = tnear <= tfar;
▲ Show 20 Lines • Show All 99 Lines • Show Last 20 Lines