[Bf-blender-cvs] [11b0034] cycles_bvh: Cycles: Hook up intersection SSE2 code for shadow regular BVH
Sergey Sharybin
noreply at git.blender.org
Fri Jun 17 16:21:54 CEST 2016
Commit: 11b0034e356df0d9d02c4792a0909f9f8f6e7dd2
Author: Sergey Sharybin
Date: Fri Jun 17 13:23:29 2016 +0200
Branches: cycles_bvh
https://developer.blender.org/rB11b0034e356df0d9d02c4792a0909f9f8f6e7dd2
Cycles: Hook up intersection SSE2 code for shadow regular BVH
===================================================================
M intern/cycles/kernel/geom/geom_bvh_shadow.h
===================================================================
diff --git a/intern/cycles/kernel/geom/geom_bvh_shadow.h b/intern/cycles/kernel/geom/geom_bvh_shadow.h
index 0a46211..0d4e9cf 100644
--- a/intern/cycles/kernel/geom/geom_bvh_shadow.h
+++ b/intern/cycles/kernel/geom/geom_bvh_shadow.h
@@ -75,6 +75,7 @@ ccl_device bool BVH_FUNCTION_FULL_NAME(BVH)(KernelGlobals *kg,
const ssef pn = cast(ssei(0, 0, 0x80000000, 0x80000000));
ssef Psplat[3], idirsplat[3];
+ ssef tnear(0.0f), tfar(isect_t);
shuffle_swap_t shufflexyz[3];
Psplat[0] = ssef(P.x);
@@ -94,86 +95,32 @@ ccl_device bool BVH_FUNCTION_FULL_NAME(BVH)(KernelGlobals *kg,
do {
/* traverse internal nodes */
while(nodeAddr >= 0 && nodeAddr != ENTRYPOINT_SENTINEL) {
- bool traverseChild0, traverseChild1;
- int nodeAddrChild1;
-
-#if !defined(__KERNEL_SSE2__)
- /* Intersect two child bounding boxes, non-SSE version */
- float t = isect_t;
-
- /* fetch node data */
+ int nodeAddrChild1, traverse_mask;
+ float dist[2];
float4 cnodes = kernel_tex_fetch(__bvh_nodes, nodeAddr+0);
- float4 node0 = kernel_tex_fetch(__bvh_nodes, nodeAddr+1);
- float4 node1 = kernel_tex_fetch(__bvh_nodes, nodeAddr+2);
- float4 node2 = kernel_tex_fetch(__bvh_nodes, nodeAddr+3);
-
- /* intersect ray against child nodes */
- NO_EXTENDED_PRECISION float c0lox = (node0.x - P.x) * idir.x;
- NO_EXTENDED_PRECISION float c0hix = (node0.z - P.x) * idir.x;
- NO_EXTENDED_PRECISION float c0loy = (node1.x - P.y) * idir.y;
- NO_EXTENDED_PRECISION float c0hiy = (node1.z - P.y) * idir.y;
- NO_EXTENDED_PRECISION float c0loz = (node2.x - P.z) * idir.z;
- NO_EXTENDED_PRECISION float c0hiz = (node2.z - P.z) * idir.z;
- NO_EXTENDED_PRECISION float c0min = max4(min(c0lox, c0hix), min(c0loy, c0hiy), min(c0loz, c0hiz), 0.0f);
- NO_EXTENDED_PRECISION float c0max = min4(max(c0lox, c0hix), max(c0loy, c0hiy), max(c0loz, c0hiz), t);
-
- NO_EXTENDED_PRECISION float c1lox = (node0.y - P.x) * idir.x;
- NO_EXTENDED_PRECISION float c1hix = (node0.w - P.x) * idir.x;
- NO_EXTENDED_PRECISION float c1loy = (node1.y - P.y) * idir.y;
- NO_EXTENDED_PRECISION float c1hiy = (node1.w - P.y) * idir.y;
- NO_EXTENDED_PRECISION float c1loz = (node2.y - P.z) * idir.z;
- NO_EXTENDED_PRECISION float c1hiz = (node2.w - P.z) * idir.z;
- NO_EXTENDED_PRECISION float c1min = max4(min(c1lox, c1hix), min(c1loy, c1hiy), min(c1loz, c1hiz), 0.0f);
- NO_EXTENDED_PRECISION float c1max = min4(max(c1lox, c1hix), max(c1loy, c1hiy), max(c1loz, c1hiz), t);
-
- /* decide which nodes to traverse next */
-# ifdef __VISIBILITY_FLAG__
- /* this visibility test gives a 5% performance hit, how to solve? */
- traverseChild0 = (c0max >= c0min) && (__float_as_uint(cnodes.x) & PATH_RAY_SHADOW);
- traverseChild1 = (c1max >= c1min) && (__float_as_uint(cnodes.y) & PATH_RAY_SHADOW);
-# else
- traverseChild0 = (c0max >= c0min);
- traverseChild1 = (c1max >= c1min);
-# endif
+#if !defined(__KERNEL_SSE2__)
#else // __KERNEL_SSE2__
- /* Intersect two child bounding boxes, SSE3 version adapted from Embree */
-
- /* fetch node data */
- const ssef *bvh_nodes = (ssef*)kg->__bvh_nodes.data + nodeAddr;
- const float4 cnodes = ((float4*)bvh_nodes)[0];
-
- /* intersect ray against child nodes */
- const ssef tminmaxx = (shuffle_swap(bvh_nodes[1], shufflexyz[0]) - Psplat[0]) * idirsplat[0];
- const ssef tminmaxy = (shuffle_swap(bvh_nodes[2], shufflexyz[1]) - Psplat[1]) * idirsplat[1];
- const ssef tminmaxz = (shuffle_swap(bvh_nodes[3], shufflexyz[2]) - Psplat[2]) * idirsplat[2];
-
- /* calculate { c0min, c1min, -c0max, -c1max} */
- const ssef minmax = max(max(tminmaxx, tminmaxy), max(tminmaxz, tsplat));
- const ssef tminmax = minmax ^ pn;
- const sseb lrhit = tminmax <= shuffle<2, 3, 0, 1>(tminmax);
-
- /* decide which nodes to traverse next */
-# ifdef __VISIBILITY_FLAG__
- /* this visibility test gives a 5% performance hit, how to solve? */
- traverseChild0 = (movemask(lrhit) & 1) && (__float_as_uint(cnodes.x) & PATH_RAY_SHADOW);
- traverseChild1 = (movemask(lrhit) & 2) && (__float_as_uint(cnodes.y) & PATH_RAY_SHADOW);
-# else
- traverseChild0 = (movemask(lrhit) & 1);
- traverseChild1 = (movemask(lrhit) & 2);
-# endif
+ traverse_mask = bvh_node_intersect(kg,
+ P,
+ dir,
+ tnear,
+ tfar,
+ tsplat,
+ Psplat,
+ idirsplat,
+ shufflexyz,
+ PATH_RAY_SHADOW,
+ nodeAddr,
+ dist);
#endif // __KERNEL_SSE2__
nodeAddr = __float_as_int(cnodes.z);
nodeAddrChild1 = __float_as_int(cnodes.w);
- if(traverseChild0 && traverseChild1) {
- /* both children were intersected, push the farther one */
-#if !defined(__KERNEL_SSE2__)
- bool closestChild1 = (c1min < c0min);
-#else
- bool closestChild1 = tminmax[1] < tminmax[0];
-#endif
+ if(traverse_mask == 3) {
+ /* Both children were intersected, push the farther one. */
+ bool closestChild1 = (dist[1] < dist[0]);
if(closestChild1) {
int tmp = nodeAddr;
@@ -186,12 +133,12 @@ ccl_device bool BVH_FUNCTION_FULL_NAME(BVH)(KernelGlobals *kg,
traversalStack[stackPtr] = nodeAddrChild1;
}
else {
- /* one child was intersected */
- if(traverseChild1) {
+ /* One child was intersected. */
+ if(traverse_mask == 2) {
nodeAddr = nodeAddrChild1;
}
- else if(!traverseChild0) {
- /* neither child was intersected */
+ else if(traverse_mask == 0) {
+ /* Neither child was intersected. */
nodeAddr = traversalStack[stackPtr];
--stackPtr;
}
@@ -317,6 +264,7 @@ ccl_device bool BVH_FUNCTION_FULL_NAME(BVH)(KernelGlobals *kg,
Psplat[2] = ssef(P.z);
tsplat = ssef(0.0f, 0.0f, -isect_t, -isect_t);
+ tfar = ssef(isect_t);
gen_idirsplat_swap(pn, shuf_identity, shuf_swap, idir, idirsplat, shufflexyz);
# endif
@@ -369,6 +317,7 @@ ccl_device bool BVH_FUNCTION_FULL_NAME(BVH)(KernelGlobals *kg,
Psplat[2] = ssef(P.z);
tsplat = ssef(0.0f, 0.0f, -isect_t, -isect_t);
+ tfar = ssef(isect_t);
gen_idirsplat_swap(pn, shuf_identity, shuf_swap, idir, idirsplat, shufflexyz);
# endif
More information about the Bf-blender-cvs
mailing list