Cycles: Implement unaligned nodes BVH traversal

This commit implements traversal of unaligned BVH nodes.

QBVH traversal is fully SIMD optimized and calculates orientation
for all 4 children at a time, regular BVH might probably be optimized
a bit more.

1 files changed, 54 insertions, 19 deletions

diff --git a/intern/cycles/kernel/geom/geom_qbvh_volume_all.h b/intern/cycles/kernel/geom/geom_qbvh_volume_all.h
index 89950f17c64..8a31775fae3 100644
--- a/intern/cycles/kernel/geom/geom_qbvh_volume_all.h
+++ b/intern/cycles/kernel/geom/geom_qbvh_volume_all.h
@@ -26,6 +26,12 @@
  *
  */
 
+#if BVH_FEATURE(BVH_HAIR)
+#  define NODE_INTERSECT qbvh_node_intersect
+#else
+#  define NODE_INTERSECT qbvh_aligned_node_intersect
+#endif
+
 ccl_device uint BVH_FUNCTION_FULL_NAME(QBVH)(KernelGlobals *kg,
                                              const Ray *ray,
                                              Intersection *isect_array,
@@ -72,13 +78,17 @@ ccl_device uint BVH_FUNCTION_FULL_NAME(QBVH)(KernelGlobals *kg,
 #endif
 
 	ssef tnear(0.0f), tfar(isect_t);
+#if BVH_FEATURE(BVH_HAIR)
+	sse3f dir4(ssef(dir.x), ssef(dir.y), ssef(dir.z));
+#endif
 	sse3f idir4(ssef(idir.x), ssef(idir.y), ssef(idir.z));
 
 #ifdef __KERNEL_AVX2__
 	float3 P_idir = P*idir;
-	sse3f P_idir4 = sse3f(P_idir.x, P_idir.y, P_idir.z);
-#else
-	sse3f org = sse3f(ssef(P.x), ssef(P.y), ssef(P.z));
+	sse3f P_idir4(P_idir.x, P_idir.y, P_idir.z);
+#endif
+#if BVH_FEATURE(BVH_HAIR) || !defined(__KERNEL_AVX2__)
+	sse3f org4(ssef(P.x), ssef(P.y), ssef(P.z));
 #endif
 
 	/* Offsets to select the side that becomes the lower or upper bound. */
@@ -108,22 +118,35 @@ ccl_device uint BVH_FUNCTION_FULL_NAME(QBVH)(KernelGlobals *kg,
 #endif
 
 				ssef dist;
-				int traverseChild = qbvh_node_intersect(kg,
-				                                        tnear,
-				                                        tfar,
+				int traverseChild = NODE_INTERSECT(kg,
+				                                   tnear,
+				                                   tfar,
 #ifdef __KERNEL_AVX2__
-				                                        P_idir4,
-#else
-				                                        org,
+				                                   P_idir4,
+#endif
+#if BVH_FEATURE(BVH_HAIR) || !defined(__KERNEL_AVX2__)
+				                                   org4,
+#endif
+#if BVH_FEATURE(BVH_HAIR)
+				                                   dir4,
 #endif
-				                                        idir4,
-				                                        near_x, near_y, near_z,
-				                                        far_x, far_y, far_z,
-				                                        nodeAddr,
-				                                        &dist);
+				                                   idir4,
+				                                   near_x, near_y, near_z,
+				                                   far_x, far_y, far_z,
+				                                   nodeAddr,
+				                                   &dist);
 
 				if(traverseChild != 0) {
-					float4 cnodes = kernel_tex_fetch(__bvh_nodes, nodeAddr+7);
+					float4 cnodes;
+#if BVH_FEATURE(BVH_HAIR)
+					if(__float_as_uint(inodes.x) & PATH_RAY_NODE_UNALIGNED) {
+						cnodes = kernel_tex_fetch(__bvh_nodes, nodeAddr+13);
+					}
+					else
+#endif
+					{
+						cnodes = kernel_tex_fetch(__bvh_nodes, nodeAddr+7);
+					}
 
 					/* One child is hit, continue with that child. */
 					int r = __bscf(traverseChild);
@@ -330,12 +353,17 @@ ccl_device uint BVH_FUNCTION_FULL_NAME(QBVH)(KernelGlobals *kg,
 						if(idir.z >= 0.0f) { near_z = 4; far_z = 5; } else { near_z = 5; far_z = 4; }
 						tfar = ssef(isect_t);
 						idir4 = sse3f(ssef(idir.x), ssef(idir.y), ssef(idir.z));
+#  if BVH_FEATURE(BVH_HAIR)
+						dir4 = sse3f(ssef(dir.x), ssef(dir.y), ssef(dir.z));
+#  endif
 #  ifdef __KERNEL_AVX2__
 						P_idir = P*idir;
 						P_idir4 = sse3f(P_idir.x, P_idir.y, P_idir.z);
-#  else
-						org = sse3f(ssef(P.x), ssef(P.y), ssef(P.z));
 #  endif
+#  if BVH_FEATURE(BVH_HAIR) || !defined(__KERNEL_AVX2__)
+						org4 = sse3f(ssef(P.x), ssef(P.y), ssef(P.z));
+#  endif
+
 						triangle_intersect_precalc(dir, &isect_precalc);
 						num_hits_in_instance = 0;
 						isect_array->t = isect_t;
@@ -389,13 +417,18 @@ ccl_device uint BVH_FUNCTION_FULL_NAME(QBVH)(KernelGlobals *kg,
 			if(idir.y >= 0.0f) { near_y = 2; far_y = 3; } else { near_y = 3; far_y = 2; }
 			if(idir.z >= 0.0f) { near_z = 4; far_z = 5; } else { near_z = 5; far_z = 4; }
 			tfar = ssef(isect_t);
+#  if BVH_FEATURE(BVH_HAIR)
+			dir4 = sse3f(ssef(dir.x), ssef(dir.y), ssef(dir.z));
+#  endif
 			idir4 = sse3f(ssef(idir.x), ssef(idir.y), ssef(idir.z));
 #  ifdef __KERNEL_AVX2__
 			P_idir = P*idir;
 			P_idir4 = sse3f(P_idir.x, P_idir.y, P_idir.z);
-#  else
-			org = sse3f(ssef(P.x), ssef(P.y), ssef(P.z));
 #  endif
+#  if BVH_FEATURE(BVH_HAIR) || !defined(__KERNEL_AVX2__)
+			org4 = sse3f(ssef(P.x), ssef(P.y), ssef(P.z));
+#  endif
+
 			triangle_intersect_precalc(dir, &isect_precalc);
 			isect_t = tmax;
 			isect_array->t = isect_t;
@@ -409,3 +442,5 @@ ccl_device uint BVH_FUNCTION_FULL_NAME(QBVH)(KernelGlobals *kg,
 
 	return num_hits;
 }
+
+#undef NODE_INTERSECT