Cycles: Split BVH nodes storage into inner and leaf nodes

This way we can get rid of inefficient memory usage caused by BVH boundbox
part being unused by leaf nodes but still being allocated for them. Doing
such split allows to save 6 of float4 values for QBVH per leaf node and 3
of float4 values for regular BVH per leaf node.

This translates into following memory save using 01.01.01.G rendered
without hair:

                   Device memory size   Device memory peak   Global memory peak
Before the patch:  4957                 5051                 7668
With the patch:    4467                 4562                 7332

The measurements are done against current master. Still need to run speed tests
and it's hard to predict if it's faster or not: on the one hand leaf nodes are
now much more coherent in cache, on the other hand they're not so much coherent
with regular nodes anymore.

Reviewers: brecht, juicyfruit

Subscribers: venomgfx, eyecandy

Differential Revision: https://developer.blender.org/D1236

15 files changed, 151 insertions, 95 deletions

diff --git a/intern/cycles/bvh/bvh.cpp b/intern/cycles/bvh/bvh.cpp
index f2777c4a15b..d1c3feed963 100644
--- a/intern/cycles/bvh/bvh.cpp
+++ b/intern/cycles/bvh/bvh.cpp
@@ -28,6 +28,7 @@
 #include "util_cache.h"
 #include "util_debug.h"
 #include "util_foreach.h"
+#include "util_logging.h"
 #include "util_map.h"
 #include "util_progress.h"
 #include "util_system.h"
@@ -111,8 +112,7 @@ bool BVH::cache_read(CacheData& key)
 		     value.read(pack.prim_type) &&
 		     value.read(pack.prim_visibility) &&
 		     value.read(pack.prim_index) &&
-		     value.read(pack.prim_object) &&
-		     value.read(pack.is_leaf)))
+		     value.read(pack.prim_object)))
 		{
 			/* Clear the pack if load failed. */
 			pack.root_index = 0;
@@ -124,7 +124,6 @@ bool BVH::cache_read(CacheData& key)
 			pack.prim_visibility.clear();
 			pack.prim_index.clear();
 			pack.prim_object.clear();
-			pack.is_leaf.clear();
 			return false;
 		}
 		return true;
@@ -147,7 +146,6 @@ void BVH::cache_write(CacheData& key)
 	value.add(pack.prim_visibility);
 	value.add(pack.prim_index);
 	value.add(pack.prim_object);
-	value.add(pack.is_leaf);
 
 	Cache::global.insert(key, value);
 
@@ -322,13 +320,14 @@ void BVH::pack_primitives()
 
 /* Pack Instances */
 
-void BVH::pack_instances(size_t nodes_size)
+void BVH::pack_instances(size_t nodes_size, size_t leaf_nodes_size)
 {
 	/* The BVH's for instances are built separately, but for traversal all
 	 * BVH's are stored in global arrays. This function merges them into the
 	 * top level BVH, adjusting indexes and offsets where appropriate. */
 	bool use_qbvh = params.use_qbvh;
 	size_t nsize = (use_qbvh)? BVH_QNODE_SIZE: BVH_NODE_SIZE;
+	size_t nsize_leaf = (use_qbvh)? BVH_QNODE_LEAF_SIZE: BVH_NODE_LEAF_SIZE;
 
 	/* adjust primitive index to point to the triangle in the global array, for
 	 * meshes with transform applied and already in the top level BVH */
@@ -343,6 +342,7 @@ void BVH::pack_instances(size_t nodes_size)
 	/* track offsets of instanced BVH data in global array */
 	size_t prim_offset = pack.prim_index.size();
 	size_t nodes_offset = nodes_size;
+	size_t nodes_leaf_offset = leaf_nodes_size;
 
 	/* clear array that gives the node indexes for instanced objects */
 	pack.object_node.clear();
@@ -354,6 +354,7 @@ void BVH::pack_instances(size_t nodes_size)
 	size_t pack_prim_index_offset = prim_index_size;
 	size_t pack_tri_woop_offset = tri_woop_size;
 	size_t pack_nodes_offset = nodes_size;
+	size_t pack_leaf_nodes_offset = leaf_nodes_size;
 	size_t object_offset = 0;
 
 	map<Mesh*, int> mesh_map;
@@ -367,6 +368,7 @@ void BVH::pack_instances(size_t nodes_size)
 				prim_index_size += bvh->pack.prim_index.size();
 				tri_woop_size += bvh->pack.tri_woop.size();
 				nodes_size += bvh->pack.nodes.size();
+				leaf_nodes_size += bvh->pack.leaf_nodes.size();
 
 				mesh_map[mesh] = 1;
 			}
@@ -381,6 +383,7 @@ void BVH::pack_instances(size_t nodes_size)
 	pack.prim_visibility.resize(prim_index_size);
 	pack.tri_woop.resize(tri_woop_size);
 	pack.nodes.resize(nodes_size);
+	pack.leaf_nodes.resize(leaf_nodes_size);
 	pack.object_node.resize(objects.size());
 
 	int *pack_prim_index = (pack.prim_index.size())? &pack.prim_index[0]: NULL;
@@ -389,6 +392,7 @@ void BVH::pack_instances(size_t nodes_size)
 	uint *pack_prim_visibility = (pack.prim_visibility.size())? &pack.prim_visibility[0]: NULL;
 	float4 *pack_tri_woop = (pack.tri_woop.size())? &pack.tri_woop[0]: NULL;
 	int4 *pack_nodes = (pack.nodes.size())? &pack.nodes[0]: NULL;
+	int4 *pack_leaf_nodes = (pack.leaf_nodes.size())? &pack.leaf_nodes[0]: NULL;
 
 	/* merge */
 	foreach(Object *ob, objects) {
@@ -414,12 +418,13 @@ void BVH::pack_instances(size_t nodes_size)
 		BVH *bvh = mesh->bvh;
 
 		int noffset = nodes_offset/nsize;
+		int noffset_leaf = nodes_leaf_offset/nsize_leaf;
 		int mesh_tri_offset = mesh->tri_offset;
 		int mesh_curve_offset = mesh->curve_offset;
 
 		/* fill in node indexes for instances */
-		if((bvh->pack.is_leaf.size() != 0) && bvh->pack.is_leaf[0])
-			pack.object_node[object_offset++] = -noffset-1;
+		if(bvh->pack.root_index == -1)
+			pack.object_node[object_offset++] = -noffset_leaf-1;
 		else
 			pack.object_node[object_offset++] = noffset;
 
@@ -453,6 +458,18 @@ void BVH::pack_instances(size_t nodes_size)
 		}
 
 		/* merge nodes */
+		if(bvh->pack.leaf_nodes.size()) {
+			int4 *leaf_nodes_offset = &bvh->pack.leaf_nodes[0];
+			size_t leaf_nodes_offset_size = bvh->pack.leaf_nodes.size();
+			for(size_t i = 0, j = 0; i < leaf_nodes_offset_size; i+=nsize_leaf, j++) {
+				int4 data = leaf_nodes_offset[i];
+				data.x += prim_offset;
+				data.y += prim_offset;
+				pack_leaf_nodes[pack_leaf_nodes_offset] = data;
+				pack_leaf_nodes_offset += nsize_leaf;
+			}
+		}
+
 		if(bvh->pack.nodes.size()) {
 			/* For QBVH we're packing a child bbox into 6 float4,
 			 * and for regular BVH they're packed into 3 float4.
@@ -460,7 +477,6 @@ void BVH::pack_instances(size_t nodes_size)
 			size_t nsize_bbox = (use_qbvh)? 6: 3;
 			int4 *bvh_nodes = &bvh->pack.nodes[0];
 			size_t bvh_nodes_size = bvh->pack.nodes.size(); 
-			bool *bvh_is_leaf = (bvh->pack.is_leaf.size() != 0) ? &bvh->pack.is_leaf[0] : NULL;
 
 			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));
@@ -468,18 +484,12 @@ void BVH::pack_instances(size_t nodes_size)
 				/* modify offsets into arrays */
 				int4 data = bvh_nodes[i + nsize_bbox];
 
-				if(bvh_is_leaf && bvh_is_leaf[j]) {
-					data.x += prim_offset;
-					data.y += prim_offset;
-				}
-				else {
-					data.x += (data.x < 0)? -noffset: noffset;
-					data.y += (data.y < 0)? -noffset: noffset;
+				data.x += (data.x < 0)? -noffset_leaf: noffset;
+				data.y += (data.y < 0)? -noffset_leaf: noffset;
 
-					if(use_qbvh) {
-						data.z += (data.z < 0)? -noffset: noffset;
-						data.w += (data.w < 0)? -noffset: noffset;
-					}
+				if(use_qbvh) {
+					data.z += (data.z < 0)? -noffset_leaf: noffset;
+					data.w += (data.w < 0)? -noffset_leaf: noffset;
 				}
 
 				pack_nodes[pack_nodes_offset + nsize_bbox] = data;
@@ -496,6 +506,7 @@ void BVH::pack_instances(size_t nodes_size)
 		}
 
 		nodes_offset += bvh->pack.nodes.size();
+		nodes_leaf_offset += bvh->pack.leaf_nodes.size();
 		prim_offset += bvh->pack.prim_index.size();
 	}
 }
@@ -509,20 +520,24 @@ RegularBVH::RegularBVH(const BVHParams& params_, const vector<Object*>& objects_
 
 void RegularBVH::pack_leaf(const BVHStackEntry& e, const LeafNode *leaf)
 {
+	float4 data[BVH_NODE_LEAF_SIZE];
+	memset(data, 0, sizeof(data));
 	if(leaf->num_triangles() == 1 && pack.prim_index[leaf->m_lo] == -1) {
 		/* object */
-		pack_node(e.idx, leaf->m_bounds, leaf->m_bounds, ~(leaf->m_lo), 0,
-		          leaf->m_visibility, leaf->m_visibility);
+		data[0].x = __int_as_float(~(leaf->m_lo));
+		data[0].y = __int_as_float(0);
 	}
 	else {
-		int prim_type = leaf->num_triangles() ? pack.prim_type[leaf->m_lo] : 0;
-		/* Triangle/curve primitive leaf.  */
-		pack_node(e.idx, leaf->m_bounds, leaf->m_bounds,
-		          leaf->m_lo, leaf->m_hi,
-		          leaf->m_visibility,
-		          prim_type);
+		/* triangle */
+		data[0].x = __int_as_float(leaf->m_lo);
+		data[0].y = __int_as_float(leaf->m_hi);
+	}
+	data[0].z = __uint_as_float(leaf->m_visibility);
+	if(leaf->num_triangles() != 0) {
+		data[0].w = __uint_as_float(pack.prim_type[leaf->m_lo]);
 	}
 
+	memcpy(&pack.leaf_nodes[e.idx * BVH_NODE_LEAF_SIZE], data, sizeof(float4)*BVH_NODE_LEAF_SIZE);
 }
 
 void RegularBVH::pack_inner(const BVHStackEntry& e, const BVHStackEntry& e0, const BVHStackEntry& e1)
@@ -545,31 +560,36 @@ void RegularBVH::pack_node(int idx, const BoundBox& b0, const BoundBox& b1, int
 
 void RegularBVH::pack_nodes(const BVHNode *root)
 {
-	size_t node_size = root->getSubtreeSize(BVH_STAT_NODE_COUNT);
+	size_t tot_node_size = root->getSubtreeSize(BVH_STAT_NODE_COUNT);
+	size_t leaf_node_size = root->getSubtreeSize(BVH_STAT_LEAF_COUNT);
+	size_t node_size = tot_node_size - leaf_node_size;
 
 	/* resize arrays */
 	pack.nodes.clear();
-	pack.is_leaf.clear();
-	pack.is_leaf.resize(node_size);
 
 	/* for top level BVH, first merge existing BVH's so we know the offsets */
-	if(params.top_level)
-		pack_instances(node_size*BVH_NODE_SIZE);
-	else
+	if(params.top_level) {
+		pack_instances(node_size*BVH_NODE_SIZE,
+		               leaf_node_size*BVH_NODE_LEAF_SIZE);
+	}
+	else {
 		pack.nodes.resize(node_size*BVH_NODE_SIZE);
+		pack.leaf_nodes.resize(leaf_node_size*BVH_NODE_LEAF_SIZE);
+	}
 
-	int nextNodeIdx = 0;
+	int nextNodeIdx = 0, nextLeafNodeIdx = 0;
 
 	vector<BVHStackEntry> stack;
 	stack.reserve(BVHParams::MAX_DEPTH*2);
-	stack.push_back(BVHStackEntry(root, nextNodeIdx++));
+	if(root->is_leaf())
+		stack.push_back(BVHStackEntry(root, nextLeafNodeIdx++));
+	else
+		stack.push_back(BVHStackEntry(root, nextNodeIdx++));
 
 	while(stack.size()) {
 		BVHStackEntry e = stack.back();
 		stack.pop_back();
 
-		pack.is_leaf[e.idx] = e.node->is_leaf();
-
 		if(e.node->is_leaf()) {
 			/* leaf node */
 			const LeafNode* leaf = reinterpret_cast<const LeafNode*>(e.node);
@@ -577,15 +597,17 @@ void RegularBVH::pack_nodes(const BVHNode *root)
 		}
 		else {
 			/* innner node */
-			stack.push_back(BVHStackEntry(e.node->get_child(0), nextNodeIdx++));
-			stack.push_back(BVHStackEntry(e.node->get_child(1), nextNodeIdx++));
+			int idx0 = (e.node->get_child(0)->is_leaf())? (nextLeafNodeIdx++) : (nextNodeIdx++);
+			int idx1 = (e.node->get_child(1)->is_leaf())? (nextLeafNodeIdx++) : (nextNodeIdx++);
+			stack.push_back(BVHStackEntry(e.node->get_child(0), idx0));
+			stack.push_back(BVHStackEntry(e.node->get_child(1), idx1));
 
 			pack_inner(e, stack[stack.size()-2], stack[stack.size()-1]);
 		}
 	}
 
 	/* root index to start traversal at, to handle case of single leaf node */
-	pack.root_index = (pack.is_leaf[0])? -1: 0;
+	pack.root_index = (root->is_leaf())? -1: 0;
 }
 
 void RegularBVH::refit_nodes()
@@ -594,17 +616,15 @@ void RegularBVH::refit_nodes()
 
 	BoundBox bbox = BoundBox::empty;
 	uint visibility = 0;
-	refit_node(0, (pack.is_leaf[0])? true: false, bbox, visibility);
+	refit_node(0, (pack.root_index == -1)? true: false, bbox, visibility);
 }
 
 void RegularBVH::refit_node(int idx, bool leaf, BoundBox& bbox, uint& visibility)
 {
-	int4 *data = &pack.nodes[idx*BVH_NODE_SIZE];
-
-	int c0 = data[3].x;
-	int c1 = data[3].y;
-
 	if(leaf) {
+		int4 *data = &pack.leaf_nodes[idx*BVH_NODE_LEAF_SIZE];
+		int c0 = data[0].x;
+		int c1 = data[0].y;
 		/* refit leaf node */
 		for(int prim = c0; prim < c1; prim++) {
 			int pidx = pack.prim_index[prim];
@@ -670,9 +690,20 @@ void RegularBVH::refit_node(int idx, bool leaf, BoundBox& bbox, uint& visibility
 			visibility |= ob->visibility;
 		}
 
-		pack_node(idx, bbox, bbox, c0, c1, visibility, data[3].w);
+		/* TODO(sergey): De-duplicate with pack_leaf(). */
+		float4 leaf_data[BVH_NODE_LEAF_SIZE];
+		leaf_data[0].x = __int_as_float(c0);
+		leaf_data[0].y = __int_as_float(c1);
+		leaf_data[0].z = __uint_as_float(visibility);
+		leaf_data[0].w = __uint_as_float(data[0].w);
+		memcpy(&pack.leaf_nodes[idx * BVH_NODE_LEAF_SIZE],
+		       leaf_data,
+		       sizeof(float4)*BVH_NODE_LEAF_SIZE);
 	}
 	else {
+		int4 *data = &pack.nodes[idx*BVH_NODE_SIZE];
+		int c0 = data[3].x;
+		int c1 = data[3].y;
 		/* refit inner node, set bbox from children */
 		BoundBox bbox0 = BoundBox::empty, bbox1 = BoundBox::empty;
 		uint visibility0 = 0, visibility1 = 0;
@@ -698,26 +729,24 @@ QBVH::QBVH(const BVHParams& params_, const vector<Object*>& objects_)
 
 void QBVH::pack_leaf(const BVHStackEntry& e, const LeafNode *leaf)
 {
-	float4 data[BVH_QNODE_SIZE];
-
+	float4 data[BVH_QNODE_LEAF_SIZE];
 	memset(data, 0, sizeof(data));
-
 	if(leaf->num_triangles() == 1 && pack.prim_index[leaf->m_lo] == -1) {
 		/* object */
-		data[6].x = __int_as_float(~(leaf->m_lo));
-		data[6].y = __int_as_float(0);
+		data[0].x = __int_as_float(~(leaf->m_lo));
+		data[0].y = __int_as_float(0);
 	}
 	else {
 		/* triangle */
-		data[6].x = __int_as_float(leaf->m_lo);
-		data[6].y = __int_as_float(leaf->m_hi);
+		data[0].x = __int_as_float(leaf->m_lo);
+		data[0].y = __int_as_float(leaf->m_hi);
 	}
-	data[6].z = __uint_as_float(leaf->m_visibility);
+	data[0].z = __uint_as_float(leaf->m_visibility);
 	if(leaf->num_triangles() != 0) {
-		data[6].w = __uint_as_float(pack.prim_type[leaf->m_lo]);
+		data[0].w = __uint_as_float(pack.prim_type[leaf->m_lo]);
 	}
 
-	memcpy(&pack.nodes[e.idx * BVH_QNODE_SIZE], data, sizeof(float4)*BVH_QNODE_SIZE);
+	memcpy(&pack.leaf_nodes[e.idx * BVH_QNODE_LEAF_SIZE], data, sizeof(float4)*BVH_QNODE_LEAF_SIZE);
 }
 
 void QBVH::pack_inner(const BVHStackEntry& e, const BVHStackEntry *en, int num)
@@ -761,31 +790,39 @@ void QBVH::pack_inner(const BVHStackEntry& e, const BVHStackEntry *en, int num)
 
 void QBVH::pack_nodes(const BVHNode *root)
 {
-	size_t node_size = root->getSubtreeSize(BVH_STAT_QNODE_COUNT);
+	size_t tot_node_size = root->getSubtreeSize(BVH_STAT_QNODE_COUNT);
+	size_t leaf_node_size = root->getSubtreeSize(BVH_STAT_LEAF_COUNT);
+	size_t node_size = tot_node_size - leaf_node_size;
 
 	/* resize arrays */
 	pack.nodes.clear();
-	pack.is_leaf.clear();
-	pack.is_leaf.resize(node_size);
+	pack.leaf_nodes.clear();
 
 	/* for top level BVH, first merge existing BVH's so we know the offsets */
-	if(params.top_level)
-		pack_instances(node_size*BVH_QNODE_SIZE);
-	else
+	if(params.top_level) {
+		pack_instances(node_size*BVH_QNODE_SIZE,
+		               leaf_node_size*BVH_QNODE_LEAF_SIZE);
+	}
+	else {
 		pack.nodes.resize(node_size*BVH_QNODE_SIZE);
+		pack.leaf_nodes.resize(leaf_node_size*BVH_QNODE_LEAF_SIZE);
+	}
 
-	int nextNodeIdx = 0;
+	int nextNodeIdx = 0, nextLeafNodeIdx = 0;
 
 	vector<BVHStackEntry> stack;
 	stack.reserve(BVHParams::MAX_DEPTH*2);
-	stack.push_back(BVHStackEntry(root, nextNodeIdx++));
+	if(root->is_leaf()) {
+		stack.push_back(BVHStackEntry(root, nextLeafNodeIdx++));
+	}
+	else {
+		stack.push_back(BVHStackEntry(root, nextNodeIdx++));
+	}
 
 	while(stack.size()) {
 		BVHStackEntry e = stack.back();
 		stack.pop_back();
 
-		pack.is_leaf[e.idx] = e.node->is_leaf();
-
 		if(e.node->is_leaf()) {
 			/* leaf node */
 			const LeafNode* leaf = reinterpret_cast<const LeafNode*>(e.node);
@@ -818,8 +855,16 @@ void QBVH::pack_nodes(const BVHNode *root)
 			}
 
 			/* push entries on the stack */
-			for(int i = 0; i < numnodes; i++)
-				stack.push_back(BVHStackEntry(nodes[i], nextNodeIdx++));
+			for(int i = 0; i < numnodes; i++) {
+				int idx;
+				if(nodes[i]->is_leaf()) {
+					idx = nextLeafNodeIdx++;
+				}
+				else {
+					idx = nextNodeIdx++;
+				}
+				stack.push_back(BVHStackEntry(nodes[i], idx));
+			}
 
 			/* set node */
 			pack_inner(e, &stack[stack.size()-numnodes], numnodes);
@@ -827,7 +872,7 @@ void QBVH::pack_nodes(const BVHNode *root)
 	}
 
 	/* root index to start traversal at, to handle case of single leaf node */
-	pack.root_index = (pack.is_leaf[0])? -1: 0;
+	pack.root_index = (root->is_leaf())? -1: 0;
 }
 
 void QBVH::refit_nodes()
@@ -836,14 +881,14 @@ void QBVH::refit_nodes()
 
 	BoundBox bbox = BoundBox::empty;
 	uint visibility = 0;
-	refit_node(0, (pack.is_leaf[0])? true: false, bbox, visibility);
+	refit_node(0, (pack.root_index == -1)? true: false, bbox, visibility);
 }
 
 void QBVH::refit_node(int idx, bool leaf, BoundBox& bbox, uint& visibility)
 {
-	int4 *data = &pack.nodes[idx*BVH_QNODE_SIZE];
-	int4 c = data[6];
 	if(leaf) {
+		int4 *data = &pack.leaf_nodes[idx*BVH_QNODE_LEAF_SIZE];
+		int4 c = data[0];
 		/* Refit leaf node. */
 		for(int prim = c.x; prim < c.y; prim++) {
 			int pidx = pack.prim_index[prim];
@@ -919,17 +964,18 @@ void QBVH::refit_node(int idx, bool leaf, BoundBox& bbox, uint& visibility)
 		 *
 		 * Same applies to the inner nodes case below.
 		 */
-		float4 leaf_data[BVH_QNODE_SIZE];
-		memset(leaf_data, 0, sizeof(leaf_data));
-		leaf_data[6].x = __int_as_float(c.x);
-		leaf_data[6].y = __int_as_float(c.y);
-		leaf_data[6].z = __uint_as_float(visibility);
-		leaf_data[6].w = __uint_as_float(c.w);
-		memcpy(&pack.nodes[idx * BVH_QNODE_SIZE],
+		float4 leaf_data[BVH_QNODE_LEAF_SIZE];
+		leaf_data[0].x = __int_as_float(c.x);
+		leaf_data[0].y = __int_as_float(c.y);
+		leaf_data[0].z = __uint_as_float(visibility);
+		leaf_data[0].w = __uint_as_float(c.w);
+		memcpy(&pack.leaf_nodes[idx * BVH_QNODE_LEAF_SIZE],
 		       leaf_data,
-		       sizeof(float4)*BVH_QNODE_SIZE);
+		       sizeof(float4)*BVH_QNODE_LEAF_SIZE);
 	}
 	else {
+		int4 *data = &pack.nodes[idx*BVH_QNODE_SIZE];
+		int4 c = data[6];
 		/* Refit inner node, set bbox from children. */
 		BoundBox child_bbox[4] = {BoundBox::empty,
 		                          BoundBox::empty,
diff --git a/intern/cycles/bvh/bvh.h b/intern/cycles/bvh/bvh.h
index 40f039541eb..669d2ccdcd5 100644
--- a/intern/cycles/bvh/bvh.h
+++ b/intern/cycles/bvh/bvh.h
@@ -36,7 +36,9 @@ class Object;
 class Progress;
 
 #define BVH_NODE_SIZE	4
+#define BVH_NODE_LEAF_SIZE	1
 #define BVH_QNODE_SIZE	7
+#define BVH_QNODE_LEAF_SIZE	1
 #define BVH_ALIGN		4096
 #define TRI_NODE_SIZE	3
 
@@ -47,7 +49,9 @@ class Progress;
 struct PackedBVH {
 	/* BVH nodes storage, one node is 4x int4, and contains two bounding boxes,
 	 * and child, triangle or object indexes depending on the node type */
-	array<int4> nodes; 
+	array<int4> nodes;
+	/* BVH leaf nodes storage. */
+	array<int4> leaf_nodes;
 	/* object index to BVH node index mapping for instances */
 	array<int> object_node; 
 	/* precomputed triangle intersection data, one triangle is 4x float4 */
@@ -61,9 +65,6 @@ struct PackedBVH {
 	array<int> prim_index;
 	/* mapping from BVH primitive index, to the object id of that primitive. */
 	array<int> prim_object;
-	/* quick array to lookup if a node is a leaf, not used for traversal, only
-	 * for instance BVH merging  */
-	array<bool> is_leaf;
 
 	/* index of the root node. */
 	int root_index;
@@ -108,7 +109,7 @@ protected:
 	void pack_triangle(int idx, float4 woop[3]);
 
 	/* merge instance BVH's */
-	void pack_instances(size_t nodes_size);
+	void pack_instances(size_t nodes_size, size_t leaf_nodes_size);
 
 	/* for subclasses to implement */
 	virtual void pack_nodes(const BVHNode *root) = 0;
diff --git a/intern/cycles/kernel/geom/geom.h b/intern/cycles/kernel/geom/geom.h
index bf0d86e6206..5ab900d47aa 100644
--- a/intern/cycles/kernel/geom/geom.h
+++ b/intern/cycles/kernel/geom/geom.h
@@ -22,7 +22,9 @@
 #define BVH_STACK_SIZE 192
 #define BVH_QSTACK_SIZE 384
 #define BVH_NODE_SIZE 4
+#define BVH_NODE_LEAF_SIZE 1
 #define BVH_QNODE_SIZE 7
+#define BVH_QNODE_LEAF_SIZE 1
 #define TRI_NODE_SIZE 3
 
 /* silly workaround for float extended precision that happens when compiling
diff --git a/intern/cycles/kernel/geom/geom_bvh_shadow.h b/intern/cycles/kernel/geom/geom_bvh_shadow.h
index d2609fc029a..e4cba99dc96 100644
--- a/intern/cycles/kernel/geom/geom_bvh_shadow.h
+++ b/intern/cycles/kernel/geom/geom_bvh_shadow.h
@@ -200,7 +200,7 @@ ccl_device bool BVH_FUNCTION_FULL_NAME(BVH)(KernelGlobals *kg,
 
 			/* if node is leaf, fetch triangle list */
 			if(nodeAddr < 0) {
-				float4 leaf = kernel_tex_fetch(__bvh_nodes, (-nodeAddr-1)*BVH_NODE_SIZE+3);
+				float4 leaf = kernel_tex_fetch(__bvh_leaf_nodes, (-nodeAddr-1)*BVH_NODE_LEAF_SIZE);
 				int primAddr = __float_as_int(leaf.x);
 
 #if BVH_FEATURE(BVH_INSTANCING)
diff --git a/intern/cycles/kernel/geom/geom_bvh_subsurface.h b/intern/cycles/kernel/geom/geom_bvh_subsurface.h
index 73d392b2ea6..a73139f9c88 100644
--- a/intern/cycles/kernel/geom/geom_bvh_subsurface.h
+++ b/intern/cycles/kernel/geom/geom_bvh_subsurface.h
@@ -187,7 +187,7 @@ ccl_device uint BVH_FUNCTION_FULL_NAME(BVH)(KernelGlobals *kg,
 
 			/* if node is leaf, fetch triangle list */
 			if(nodeAddr < 0) {
-				float4 leaf = kernel_tex_fetch(__bvh_nodes, (-nodeAddr-1)*BVH_NODE_SIZE+3);
+				float4 leaf = kernel_tex_fetch(__bvh_leaf_nodes, (-nodeAddr-1)*BVH_NODE_LEAF_SIZE);
 				int primAddr = __float_as_int(leaf.x);
 
 #if BVH_FEATURE(BVH_INSTANCING)
diff --git a/intern/cycles/kernel/geom/geom_bvh_traversal.h b/intern/cycles/kernel/geom/geom_bvh_traversal.h
index c4d2a1bdc6a..87f8ccd01b1 100644
--- a/intern/cycles/kernel/geom/geom_bvh_traversal.h
+++ b/intern/cycles/kernel/geom/geom_bvh_traversal.h
@@ -248,7 +248,7 @@ ccl_device bool BVH_FUNCTION_FULL_NAME(BVH)(KernelGlobals *kg,
 
 			/* if node is leaf, fetch triangle list */
 			if(nodeAddr < 0) {
-				float4 leaf = kernel_tex_fetch(__bvh_nodes, (-nodeAddr-1)*BVH_NODE_SIZE+3);
+				float4 leaf = kernel_tex_fetch(__bvh_leaf_nodes, (-nodeAddr-1)*BVH_NODE_LEAF_SIZE);
 				int primAddr = __float_as_int(leaf.x);
 
 #if BVH_FEATURE(BVH_INSTANCING)
diff --git a/intern/cycles/kernel/geom/geom_bvh_volume.h b/intern/cycles/kernel/geom/geom_bvh_volume.h
index 111f840cb95..41c784869f2 100644
--- a/intern/cycles/kernel/geom/geom_bvh_volume.h
+++ b/intern/cycles/kernel/geom/geom_bvh_volume.h
@@ -188,7 +188,7 @@ ccl_device bool BVH_FUNCTION_FULL_NAME(BVH)(KernelGlobals *kg,
 
 			/* if node is leaf, fetch triangle list */
 			if(nodeAddr < 0) {
-				float4 leaf = kernel_tex_fetch(__bvh_nodes, (-nodeAddr-1)*BVH_NODE_SIZE+3);
+				float4 leaf = kernel_tex_fetch(__bvh_leaf_nodes, (-nodeAddr-1)*BVH_NODE_LEAF_SIZE);
 				int primAddr = __float_as_int(leaf.x);
 
 #if BVH_FEATURE(BVH_INSTANCING)
diff --git a/intern/cycles/kernel/geom/geom_qbvh_shadow.h b/intern/cycles/kernel/geom/geom_qbvh_shadow.h
index 2adebb28c28..f79b2ed9f34 100644
--- a/intern/cycles/kernel/geom/geom_qbvh_shadow.h
+++ b/intern/cycles/kernel/geom/geom_qbvh_shadow.h
@@ -206,7 +206,7 @@ ccl_device bool BVH_FUNCTION_FULL_NAME(QBVH)(KernelGlobals *kg,
 
 			/* If node is leaf, fetch triangle list. */
 			if(nodeAddr < 0) {
-				float4 leaf = kernel_tex_fetch(__bvh_nodes, (-nodeAddr-1)*BVH_QNODE_SIZE+6);
+				float4 leaf = kernel_tex_fetch(__bvh_leaf_nodes, (-nodeAddr-1)*BVH_QNODE_LEAF_SIZE);
 #ifdef __VISIBILITY_FLAG__
 				if((__float_as_uint(leaf.z) & PATH_RAY_SHADOW) == 0) {
 					/* Pop. */
diff --git a/intern/cycles/kernel/geom/geom_qbvh_subsurface.h b/intern/cycles/kernel/geom/geom_qbvh_subsurface.h
index 654c5cf1493..d85e1a4691e 100644
--- a/intern/cycles/kernel/geom/geom_qbvh_subsurface.h
+++ b/intern/cycles/kernel/geom/geom_qbvh_subsurface.h
@@ -202,7 +202,7 @@ ccl_device uint BVH_FUNCTION_FULL_NAME(QBVH)(KernelGlobals *kg,
 
 			/* If node is leaf, fetch triangle list. */
 			if(nodeAddr < 0) {
-				float4 leaf = kernel_tex_fetch(__bvh_nodes, (-nodeAddr-1)*BVH_QNODE_SIZE+6);
+				float4 leaf = kernel_tex_fetch(__bvh_leaf_nodes, (-nodeAddr-1)*BVH_QNODE_LEAF_SIZE);
 				int primAddr = __float_as_int(leaf.x);
 
 #if BVH_FEATURE(BVH_INSTANCING)
diff --git a/intern/cycles/kernel/geom/geom_qbvh_traversal.h b/intern/cycles/kernel/geom/geom_qbvh_traversal.h
index 6c61296ad1f..cc20ebd8154 100644
--- a/intern/cycles/kernel/geom/geom_qbvh_traversal.h
+++ b/intern/cycles/kernel/geom/geom_qbvh_traversal.h
@@ -260,7 +260,7 @@ ccl_device bool BVH_FUNCTION_FULL_NAME(QBVH)(KernelGlobals *kg,
 
 			/* If node is leaf, fetch triangle list. */
 			if(nodeAddr < 0) {
-				float4 leaf = kernel_tex_fetch(__bvh_nodes, (-nodeAddr-1)*BVH_QNODE_SIZE+6);
+				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)))
diff --git a/intern/cycles/kernel/geom/geom_qbvh_volume.h b/intern/cycles/kernel/geom/geom_qbvh_volume.h
index 1c9ace68d53..e3a5dcffee5 100644
--- a/intern/cycles/kernel/geom/geom_qbvh_volume.h
+++ b/intern/cycles/kernel/geom/geom_qbvh_volume.h
@@ -208,7 +208,7 @@ ccl_device bool BVH_FUNCTION_FULL_NAME(QBVH)(KernelGlobals *kg,
 
 			/* If node is leaf, fetch triangle list. */
 			if(nodeAddr < 0) {
-				float4 leaf = kernel_tex_fetch(__bvh_nodes, (-nodeAddr-1)*BVH_QNODE_SIZE+6);
+				float4 leaf = kernel_tex_fetch(__bvh_leaf_nodes, (-nodeAddr-1)*BVH_QNODE_LEAF_SIZE);
 				int primAddr = __float_as_int(leaf.x);
 
 #if BVH_FEATURE(BVH_INSTANCING)
diff --git a/intern/cycles/kernel/kernel_textures.h b/intern/cycles/kernel/kernel_textures.h
index 374dc6d1dd9..f545a056cc8 100644
--- a/intern/cycles/kernel/kernel_textures.h
+++ b/intern/cycles/kernel/kernel_textures.h
@@ -24,6 +24,7 @@
 
 /* bvh */
 KERNEL_TEX(float4, texture_float4, __bvh_nodes)
+KERNEL_TEX(float4, texture_float4, __bvh_leaf_nodes)
 KERNEL_TEX(float4, texture_float4, __tri_woop)
 KERNEL_TEX(uint, texture_uint, __prim_type)
 KERNEL_TEX(uint, texture_uint, __prim_visibility)
diff --git a/intern/cycles/kernel/svm/svm_image.h b/intern/cycles/kernel/svm/svm_image.h
index e927d46d404..ccc1e3df730 100644
--- a/intern/cycles/kernel/svm/svm_image.h
+++ b/intern/cycles/kernel/svm/svm_image.h
@@ -251,9 +251,9 @@ ccl_device float4 svm_image_texture(KernelGlobals *kg, int id, float x, float y,
 		case 95: r = kernel_tex_image_interp(__tex_image_095, x, y); break;
 		case 96: r = kernel_tex_image_interp(__tex_image_096, x, y); break;
 		case 97: r = kernel_tex_image_interp(__tex_image_097, x, y); break;
-		case 98: r = kernel_tex_image_interp(__tex_image_098, x, y); break;
 
 #if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 300)
+		case 98: r = kernel_tex_image_interp(__tex_image_098, x, y); break;
 		case 99: r = kernel_tex_image_interp(__tex_image_099, x, y); break;
 		case 100: r = kernel_tex_image_interp(__tex_image_100, x, y); break;
 		case 101: r = kernel_tex_image_interp(__tex_image_101, x, y); break;
diff --git a/intern/cycles/render/mesh.cpp b/intern/cycles/render/mesh.cpp
index b6f3cb3502b..19ca159d03f 100644
--- a/intern/cycles/render/mesh.cpp
+++ b/intern/cycles/render/mesh.cpp
@@ -1101,6 +1101,10 @@ void MeshManager::device_update_bvh(Device *device, DeviceScene *dscene, Scene *
 		dscene->bvh_nodes.reference((float4*)&pack.nodes[0], pack.nodes.size());
 		device->tex_alloc("__bvh_nodes", dscene->bvh_nodes);
 	}
+	if(pack.leaf_nodes.size()) {
+		dscene->bvh_leaf_nodes.reference((float4*)&pack.leaf_nodes[0], pack.leaf_nodes.size());
+		device->tex_alloc("__bvh_leaf_nodes", dscene->bvh_leaf_nodes);
+	}
 	if(pack.object_node.size()) {
 		dscene->object_node.reference((uint*)&pack.object_node[0], pack.object_node.size());
 		device->tex_alloc("__object_node", dscene->object_node);
@@ -1315,6 +1319,7 @@ void MeshManager::device_update(Device *device, DeviceScene *dscene, Scene *scen
 void MeshManager::device_free(Device *device, DeviceScene *dscene)
 {
 	device->tex_free(dscene->bvh_nodes);
+	device->tex_free(dscene->bvh_leaf_nodes);
 	device->tex_free(dscene->object_node);
 	device->tex_free(dscene->tri_woop);
 	device->tex_free(dscene->prim_type);
diff --git a/intern/cycles/render/scene.h b/intern/cycles/render/scene.h
index 53c3a95903c..851e5ac0b72 100644
--- a/intern/cycles/render/scene.h
+++ b/intern/cycles/render/scene.h
@@ -62,6 +62,7 @@ class DeviceScene {
 public:
 	/* BVH */
 	device_vector<float4> bvh_nodes;
+	device_vector<float4> bvh_leaf_nodes;
 	device_vector<uint> object_node;
 	device_vector<float4> tri_woop;
 	device_vector<uint> prim_type;