[Bf-blender-cvs] [5c6a14f] master: Cycles: More tweaks to make specialized BVH traversal matching

[Sergey Sharybin]

Commit: 5c6a14f4e520e2eca73ff3a2c4b6186ee4dab40a
Author: Sergey Sharybin
Date:   Mon Sep 19 15:29:37 2016 +0200
Branches: master
https://developer.blender.org/rB5c6a14f4e520e2eca73ff3a2c4b6186ee4dab40a

Cycles: More tweaks to make specialized BVH traversal matching

===================================================================

M	intern/cycles/kernel/bvh/qbvh_shadow_all.h
M	intern/cycles/kernel/bvh/qbvh_volume.h
M	intern/cycles/kernel/bvh/qbvh_volume_all.h

===================================================================

diff --git a/intern/cycles/kernel/bvh/qbvh_shadow_all.h b/intern/cycles/kernel/bvh/qbvh_shadow_all.h
index 03c678b..3136c49 100644
--- a/intern/cycles/kernel/bvh/qbvh_shadow_all.h
+++ b/intern/cycles/kernel/bvh/qbvh_shadow_all.h
@@ -40,6 +40,7 @@ ccl_device bool BVH_FUNCTION_FULL_NAME(QBVH)(KernelGlobals *kg,
                                              uint *num_hits)
 {
 	/* TODO(sergey):
+	*  - Test if pushing distance on the stack helps.
 	 * - Likely and unlikely for if() statements.
 	 * - Test restrict attribute for pointers.
 	 */
@@ -77,7 +78,7 @@ ccl_device bool BVH_FUNCTION_FULL_NAME(QBVH)(KernelGlobals *kg,
 	int num_hits_in_instance = 0;
 #endif
 
-	ssef tnear(0.0f), tfar(tmax);
+	ssef tnear(0.0f), tfar(isect_t);
 #if BVH_FEATURE(BVH_HAIR)
 	sse3f dir4(ssef(dir.x), ssef(dir.y), ssef(dir.z));
 #endif
@@ -125,12 +126,12 @@ ccl_device bool BVH_FUNCTION_FULL_NAME(QBVH)(KernelGlobals *kg,
 #ifdef __KERNEL_AVX2__
 				                                P_idir4,
 #endif
-#  if BVH_FEATURE(BVH_HAIR) || !defined(__KERNEL_AVX2__)
+#if BVH_FEATURE(BVH_HAIR) || !defined(__KERNEL_AVX2__)
 				                                org4,
-#  endif
-#  if BVH_FEATURE(BVH_HAIR)
+#endif
+#if BVH_FEATURE(BVH_HAIR)
 				                                dir4,
-#  endif
+#endif
 				                                idir4,
 				                                near_x, near_y, near_z,
 				                                far_x, far_y, far_z,
@@ -430,21 +431,18 @@ ccl_device bool BVH_FUNCTION_FULL_NAME(QBVH)(KernelGlobals *kg,
 			/* Instance pop. */
 			if(num_hits_in_instance) {
 				float t_fac;
-
 #  if BVH_FEATURE(BVH_MOTION)
 				bvh_instance_motion_pop_factor(kg, object, ray, &P, &dir, &idir, &t_fac, &ob_itfm);
 #  else
 				bvh_instance_pop_factor(kg, object, ray, &P, &dir, &idir, &t_fac);
 #  endif
-
-				/* scale isect->t to adjust for instancing */
+				/* Scale isect->t to adjust for instancing. */
 				for(int i = 0; i < num_hits_in_instance; i++) {
 					(isect_array-i-1)->t *= t_fac;
 				}
 			}
 			else {
 				float ignore_t = FLT_MAX;
-
 #  if BVH_FEATURE(BVH_MOTION)
 				bvh_instance_motion_pop(kg, object, ray, &P, &dir, &idir, &ignore_t, &ob_itfm);
 #  else
@@ -458,7 +456,7 @@ ccl_device bool BVH_FUNCTION_FULL_NAME(QBVH)(KernelGlobals *kg,
 			if(idir.x >= 0.0f) { near_x = 0; far_x = 1; } else { near_x = 1; far_x = 0; }
 			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(tmax);
+			tfar = ssef(isect_t);
 #  if BVH_FEATURE(BVH_HAIR)
 			dir4 = sse3f(ssef(dir.x), ssef(dir.y), ssef(dir.z));
 #  endif
diff --git a/intern/cycles/kernel/bvh/qbvh_volume.h b/intern/cycles/kernel/bvh/qbvh_volume.h
index 847a11d..aff9f55 100644
--- a/intern/cycles/kernel/bvh/qbvh_volume.h
+++ b/intern/cycles/kernel/bvh/qbvh_volume.h
@@ -103,8 +103,9 @@ ccl_device bool BVH_FUNCTION_FULL_NAME(QBVH)(KernelGlobals *kg,
 		do {
 			/* Traverse internal nodes. */
 			while(node_addr >= 0 && node_addr != ENTRYPOINT_SENTINEL) {
-#ifdef __VISIBILITY_FLAG__
 				float4 inodes = kernel_tex_fetch(__bvh_nodes, node_addr+0);
+
+#ifdef __VISIBILITY_FLAG__
 				if((__float_as_uint(inodes.x) & visibility) == 0) {
 					/* Pop. */
 					node_addr = traversal_stack[stack_ptr].addr;
diff --git a/intern/cycles/kernel/bvh/qbvh_volume_all.h b/intern/cycles/kernel/bvh/qbvh_volume_all.h
index 0fe4454..26da014 100644
--- a/intern/cycles/kernel/bvh/qbvh_volume_all.h
+++ b/intern/cycles/kernel/bvh/qbvh_volume_all.h
@@ -107,8 +107,9 @@ ccl_device uint BVH_FUNCTION_FULL_NAME(QBVH)(KernelGlobals *kg,
 		do {
 			/* Traverse internal nodes. */
 			while(node_addr >= 0 && node_addr != ENTRYPOINT_SENTINEL) {
-#ifdef __VISIBILITY_FLAG__
 				float4 inodes = kernel_tex_fetch(__bvh_nodes, node_addr+0);
+
+#ifdef __VISIBILITY_FLAG__
 				if((__float_as_uint(inodes.x) & visibility) == 0) {
 					/* Pop. */
 					node_addr = traversal_stack[stack_ptr].addr;
@@ -419,6 +420,9 @@ ccl_device uint BVH_FUNCTION_FULL_NAME(QBVH)(KernelGlobals *kg,
 #  endif
 			}
 
+			isect_t = tmax;
+			isect_array->t = isect_t;
+
 			if(idir.x >= 0.0f) { near_x = 0; far_x = 1; } else { near_x = 1; far_x = 0; }
 			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; }
@@ -436,8 +440,6 @@ ccl_device uint BVH_FUNCTION_FULL_NAME(QBVH)(KernelGlobals *kg,
 #  endif
 
 			triangle_intersect_precalc(dir, &isect_precalc);
-			isect_t = tmax;
-			isect_array->t = isect_t;
 
 			object = OBJECT_NONE;
 			node_addr = traversal_stack[stack_ptr].addr;