[Bf-blender-cvs] [0e393a3] cycles_bvh: Cycles: Support visibility check for inner nodes of QBVH
Sergey Sharybin
noreply at git.blender.org
Fri Jun 10 16:32:25 CEST 2016
Commit: 0e393a37a2cd4850b855aa2b9c171156d27911ae
Author: Sergey Sharybin
Date: Fri Jun 10 10:14:05 2016 +0200
Branches: cycles_bvh
https://developer.blender.org/rB0e393a37a2cd4850b855aa2b9c171156d27911ae
Cycles: Support visibility check for inner nodes of QBVH
It was initially unsupported because initial idea of checking visibility
of all children was slowing scenes down a lot. Now the idea has changed
and we only perform visibility check of current node. This avoids huge
slowdown (from tests here it seems to be withing 1-2%, but more tests
would never hurt) and gives nice speedup of ray traversal for complex
scenes which utilized ray visibility.
Here's timing of koro.blend:
Without visibility check With visibility check
Original file 4min 20sec 4min 23sec
Camera rays only 1min 43 sec 55sec
Unfortunately, this doesn't come for free and requires extra data in
BVH node, which increases memory usage of BVH nodes by 15%. This we
can solve with some future trickery of avoiding __tri_storage created
for curve segments.
===================================================================
M intern/cycles/bvh/bvh.cpp
M intern/cycles/bvh/bvh.h
M intern/cycles/kernel/geom/geom.h
M intern/cycles/kernel/geom/geom_qbvh.h
M intern/cycles/kernel/geom/geom_qbvh_shadow.h
M intern/cycles/kernel/geom/geom_qbvh_subsurface.h
M intern/cycles/kernel/geom/geom_qbvh_traversal.h
M intern/cycles/kernel/geom/geom_qbvh_volume.h
M intern/cycles/kernel/geom/geom_qbvh_volume_all.h
===================================================================
diff --git a/intern/cycles/bvh/bvh.cpp b/intern/cycles/bvh/bvh.cpp
index fa2b9ae..311d729 100644
--- a/intern/cycles/bvh/bvh.cpp
+++ b/intern/cycles/bvh/bvh.cpp
@@ -338,12 +338,14 @@ void BVH::pack_instances(size_t nodes_size, size_t leaf_nodes_size)
/* For QBVH we're packing a child bbox into 6 float4,
* and for regular BVH they're packed into 3 float4.
*/
- size_t nsize_bbox = (use_qbvh)? 6: 3;
+ const size_t nsize_bbox = (use_qbvh)? 7: 3;
int4 *bvh_nodes = &bvh->pack.nodes[0];
- size_t bvh_nodes_size = bvh->pack.nodes.size();
+ size_t bvh_nodes_size = bvh->pack.nodes.size();
for(size_t i = 0, j = 0; i < bvh_nodes_size; i+=nsize, j++) {
- memcpy(pack_nodes + pack_nodes_offset, bvh_nodes + i, nsize_bbox*sizeof(int4));
+ memcpy(pack_nodes + pack_nodes_offset,
+ bvh_nodes + i,
+ nsize_bbox*sizeof(int4));
/* modify offsets into arrays */
int4 data = bvh_nodes[i + nsize_bbox];
@@ -617,34 +619,35 @@ void QBVH::pack_inner(const BVHStackEntry& e, const BVHStackEntry *en, int num)
{
float4 data[BVH_QNODE_SIZE];
+ data[0].x = __uint_as_float(e.node->m_visibility);
for(int i = 0; i < num; i++) {
float3 bb_min = en[i].node->m_bounds.min;
float3 bb_max = en[i].node->m_bounds.max;
- data[0][i] = bb_min.x;
- data[1][i] = bb_max.x;
- data[2][i] = bb_min.y;
- data[3][i] = bb_max.y;
- data[4][i] = bb_min.z;
- data[5][i] = bb_max.z;
+ data[1][i] = bb_min.x;
+ data[2][i] = bb_max.x;
+ data[3][i] = bb_min.y;
+ data[4][i] = bb_max.y;
+ data[5][i] = bb_min.z;
+ data[6][i] = bb_max.z;
- data[6][i] = __int_as_float(en[i].encodeIdx());
+ data[7][i] = __int_as_float(en[i].encodeIdx());
}
for(int i = num; i < 4; i++) {
/* We store BB which would never be recorded as intersection
* so kernel might safely assume there are always 4 child nodes.
*/
- data[0][i] = FLT_MAX;
- data[1][i] = -FLT_MAX;
+ data[1][i] = FLT_MAX;
+ data[2][i] = -FLT_MAX;
- data[2][i] = FLT_MAX;
- data[3][i] = -FLT_MAX;
+ data[3][i] = FLT_MAX;
+ data[4][i] = -FLT_MAX;
- data[4][i] = FLT_MAX;
- data[5][i] = -FLT_MAX;
+ data[5][i] = FLT_MAX;
+ data[6][i] = -FLT_MAX;
- data[6][i] = __int_as_float(0);
+ data[7][i] = __int_as_float(0);
}
memcpy(&pack.nodes[e.idx * BVH_QNODE_SIZE], data, sizeof(float4)*BVH_QNODE_SIZE);
@@ -839,7 +842,7 @@ void QBVH::refit_node(int idx, bool leaf, BoundBox& bbox, uint& visibility)
}
else {
int4 *data = &pack.nodes[idx*BVH_QNODE_SIZE];
- int4 c = data[6];
+ int4 c = data[7];
/* Refit inner node, set bbox from children. */
BoundBox child_bbox[4] = {BoundBox::empty,
BoundBox::empty,
@@ -859,16 +862,20 @@ void QBVH::refit_node(int idx, bool leaf, BoundBox& bbox, uint& visibility)
}
float4 inner_data[BVH_QNODE_SIZE];
+ inner_data[0] = make_float4(__int_as_float(data[0].x),
+ __int_as_float(data[1].y),
+ __int_as_float(data[2].z),
+ __int_as_float(data[3].w));
for(int i = 0; i < 4; ++i) {
float3 bb_min = child_bbox[i].min;
float3 bb_max = child_bbox[i].max;
- inner_data[0][i] = bb_min.x;
- inner_data[1][i] = bb_max.x;
- inner_data[2][i] = bb_min.y;
- inner_data[3][i] = bb_max.y;
- inner_data[4][i] = bb_min.z;
- inner_data[5][i] = bb_max.z;
- inner_data[6][i] = __int_as_float(c[i]);
+ inner_data[1][i] = bb_min.x;
+ inner_data[2][i] = bb_max.x;
+ inner_data[3][i] = bb_min.y;
+ inner_data[4][i] = bb_max.y;
+ inner_data[5][i] = bb_min.z;
+ inner_data[6][i] = bb_max.z;
+ inner_data[7][i] = __int_as_float(c[i]);
}
memcpy(&pack.nodes[idx * BVH_QNODE_SIZE],
inner_data,
diff --git a/intern/cycles/bvh/bvh.h b/intern/cycles/bvh/bvh.h
index 6076c25..b1655f0 100644
--- a/intern/cycles/bvh/bvh.h
+++ b/intern/cycles/bvh/bvh.h
@@ -35,7 +35,7 @@ class Progress;
#define BVH_NODE_SIZE 4
#define BVH_NODE_LEAF_SIZE 1
-#define BVH_QNODE_SIZE 7
+#define BVH_QNODE_SIZE 8
#define BVH_QNODE_LEAF_SIZE 1
#define BVH_ALIGN 4096
#define TRI_NODE_SIZE 3
diff --git a/intern/cycles/kernel/geom/geom.h b/intern/cycles/kernel/geom/geom.h
index c94a538..3e93287 100644
--- a/intern/cycles/kernel/geom/geom.h
+++ b/intern/cycles/kernel/geom/geom.h
@@ -23,7 +23,7 @@
#define BVH_QSTACK_SIZE 384
#define BVH_NODE_SIZE 4
#define BVH_NODE_LEAF_SIZE 1
-#define BVH_QNODE_SIZE 7
+#define BVH_QNODE_SIZE 8
#define BVH_QNODE_LEAF_SIZE 1
#define TRI_NODE_SIZE 3
diff --git a/intern/cycles/kernel/geom/geom_qbvh.h b/intern/cycles/kernel/geom/geom_qbvh.h
index 2a2d782..118bceb 100644
--- a/intern/cycles/kernel/geom/geom_qbvh.h
+++ b/intern/cycles/kernel/geom/geom_qbvh.h
@@ -69,7 +69,7 @@ ccl_device_inline int qbvh_node_intersect(KernelGlobals *__restrict kg,
const int nodeAddr,
ssef *__restrict dist)
{
- const int offset = nodeAddr*BVH_QNODE_SIZE;
+ const int offset = nodeAddr*BVH_QNODE_SIZE + 1;
#ifdef __KERNEL_AVX2__
const ssef tnear_x = msub(kernel_tex_fetch_ssef(__bvh_nodes, offset+near_x), idir.x, org_idir.x);
const ssef tnear_y = msub(kernel_tex_fetch_ssef(__bvh_nodes, offset+near_y), idir.y, org_idir.y);
@@ -120,7 +120,7 @@ ccl_device_inline int qbvh_node_intersect_robust(KernelGlobals *__restrict kg,
const float difl,
ssef *__restrict dist)
{
- const int offset = nodeAddr*BVH_QNODE_SIZE;
+ const int offset = nodeAddr*BVH_QNODE_SIZE + 1;
#ifdef __KERNEL_AVX2__
const ssef tnear_x = msub(kernel_tex_fetch_ssef(__bvh_nodes, offset+near_x), idir.x, P_idir.x);
const ssef tnear_y = msub(kernel_tex_fetch_ssef(__bvh_nodes, offset+near_y), idir.y, P_idir.y);
diff --git a/intern/cycles/kernel/geom/geom_qbvh_shadow.h b/intern/cycles/kernel/geom/geom_qbvh_shadow.h
index edb5b5c7..01babba 100644
--- a/intern/cycles/kernel/geom/geom_qbvh_shadow.h
+++ b/intern/cycles/kernel/geom/geom_qbvh_shadow.h
@@ -97,6 +97,17 @@ ccl_device bool BVH_FUNCTION_FULL_NAME(QBVH)(KernelGlobals *kg,
do {
/* Traverse internal nodes. */
while(nodeAddr >= 0 && nodeAddr != ENTRYPOINT_SENTINEL) {
+ float4 inodes = kernel_tex_fetch(__bvh_nodes, nodeAddr*BVH_QNODE_SIZE+0);
+
+#ifdef __VISIBILITY_FLAG__
+ if((__float_as_uint(inodes.x) & PATH_RAY_SHADOW) == 0) {
+ /* Pop. */
+ nodeAddr = traversalStack[stackPtr].addr;
+ --stackPtr;
+ continue;
+ }
+#endif
+
ssef dist;
int traverseChild = qbvh_node_intersect(kg,
tnear,
@@ -113,7 +124,8 @@ ccl_device bool BVH_FUNCTION_FULL_NAME(QBVH)(KernelGlobals *kg,
&dist);
if(traverseChild != 0) {
- float4 cnodes = kernel_tex_fetch(__bvh_nodes, nodeAddr*BVH_QNODE_SIZE+6);
+ float4 cnodes = kernel_tex_fetch(__bvh_nodes,
+ nodeAddr*BVH_QNODE_SIZE+7);
/* One child is hit, continue with that child. */
int r = __bscf(traverseChild);
diff --git a/intern/cycles/kernel/geom/geom_qbvh_subsurface.h b/intern/cycles/kernel/geom/geom_qbvh_subsurface.h
index 84512a8..bc717a0 100644
--- a/intern/cycles/kernel/geom/geom_qbvh_subsurface.h
+++ b/intern/cycles/kernel/geom/geom_qbvh_subsurface.h
@@ -123,7 +123,8 @@ ccl_device void BVH_FUNCTION_FULL_NAME(QBVH)(KernelGlobals *kg,
&dist);
if(traverseChild != 0) {
- float4 cnodes = kernel_tex_fetch(__bvh_nodes, nodeAddr*BVH_QNODE_SIZE+6);
+ float4 cnodes = kernel_tex_fetch(__bvh_nodes,
+ nodeAddr*BVH_QNODE_SIZE+7);
/* One child is hit, continue with that child. */
int r = __bscf(traverseChild);
diff --git a/intern/cycles/kernel/geom/geom_qbvh_traversal.h b/intern/cycles/kernel/geom/geom_qbvh_traversal.h
index 738d08a..098881a 100644
--- a/intern/cycles/kernel/geom/geom_qbvh_traversal.h
+++ b/intern/cycles/kernel/geom/geom_qbvh_traversal.h
@@ -106,7 +106,14 @@ ccl_device bool BVH_FUNCTION_FULL_NAME(QBVH)(KernelGlobals *kg,
do {
/* Traverse internal nodes. */
while(nodeAddr >= 0 && nodeAddr != ENTRYPOINT_SENTINEL) {
- if(UNLIKELY(nodeDist > isect->t)) {
+ float4 inodes = kernel_tex_fetch(__bvh_nodes,
+ nodeAddr*BVH_QNODE_SIZE+0);
+
+ if(UNLIKELY(nodeDist > isect->t)
+#ifdef __VISIBILITY_FLAG__
+ || (__float_as_uint(inodes.x) & visibility) == 0)
+#endif
+ {
/* Pop. */
nodeAddr = traversalStack[stackPtr].addr;
nodeDist = traversalStack[stackPtr].dist;
@@ -160,7 +167,8 @@ ccl_device bool BVH_FUNCTION_FULL_NAME(QBVH)(KernelGlobals *kg,
}
if(traverseChild != 0) {
- float4 cnodes = kernel_tex_fetch(__bvh_nodes, nodeAddr*BVH_QNODE_SIZE+6);
+ float4 cnodes = kernel_tex_fetch(__bvh_nodes,
+ nodeAddr*BVH_QNODE_SIZE+7);
/* One child is hit, continue with that child. */
int r = __bscf(traverseChild);
@@ -261,7 +269,8 @@ ccl_device bool BVH_FUNCTION_FULL_NAME(QBVH)(KernelGlobals *kg,
float4 leaf = kernel_tex_fetch(__bvh_leaf_nodes, (-nodeAddr-1)*BVH_QNODE_LEAF_SIZE);
#ifdef __VISIBILITY_FLAG__
- if(UNLIKELY((nodeDist > isect->t) || ((__float_as_uint(leaf.z) & visibility) == 0)))
+ if(UNLIKELY((nodeDist > isect->t) ||
+ ((__float_as_uint(leaf.z) & visibility) == 0)))
#else
if(UNLIKELY((nodeDist > isect->t)))
#endif
diff --git a/intern/cycles/kernel/geom/geom_qbvh_volume.h b/intern/cycles/kernel/geom/geom_qbvh_volume.h
index ab2e530..2aae903 100644
--- a/intern/cycles/kernel/geom/geom_qbvh_volume.h
+++ b/intern/cycles/kernel/geom/geom_qbvh_volume.h
@@ -93,6 +93,16 @@ ccl_device bool BVH_FUNCTION_FULL_NAME(QBVH)(KernelGlobals *kg,
do {
/* Traverse internal nodes. */
while(nodeAddr >= 0 && nodeAddr != ENTRYPOINT_SENTINEL) {
+#ifdef __VISIBILITY_FLAG__
+ float4 inodes = kernel_tex_fetch(__bvh_nodes, nodeAddr*BVH_QNODE_SIZE+0);
+ if((__float_as_uint(inodes.x) & visibility) == 0) {
+ /* Pop. */
+ nodeAddr = traversalStack[stackPtr].addr;
+ --stackPtr;
+ continue;
+
@@ Diff output truncated at 10240 characters. @@
More information about the Bf-blender-cvs
mailing list