Cycles: Use new SSE version of offset calculation for all QBVH flavors

Gives up to ~1% speedup again.

While it seems to be small, still nice since the code now is actually more
clean that it used to be before.

4 files changed, 30 insertions, 34 deletions

diff --git a/intern/cycles/kernel/bvh/qbvh_shadow_all.h b/intern/cycles/kernel/bvh/qbvh_shadow_all.h
index ae7aec2082f..5f4d06f12ea 100644
--- a/intern/cycles/kernel/bvh/qbvh_shadow_all.h
+++ b/intern/cycles/kernel/bvh/qbvh_shadow_all.h
@@ -92,10 +92,9 @@ ccl_device bool BVH_FUNCTION_FULL_NAME(QBVH)(KernelGlobals *kg,
 	/* Offsets to select the side that becomes the lower or upper bound. */
 	int near_x, near_y, near_z;
 	int far_x, far_y, far_z;
-
-	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; }
+	qbvh_near_far_idx_calc(idir,
+	                       &near_x, &near_y, &near_z,
+	                       &far_x, &far_y, &far_z);
 
 	IsectPrecalc isect_precalc;
 	triangle_intersect_precalc(dir, &isect_precalc);
@@ -392,9 +391,9 @@ ccl_device bool BVH_FUNCTION_FULL_NAME(QBVH)(KernelGlobals *kg,
 					num_hits_in_instance = 0;
 					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; }
+					qbvh_near_far_idx_calc(idir,
+					                       &near_x, &near_y, &near_z,
+					                       &far_x, &far_y, &far_z);
 					tfar = ssef(isect_t);
 #  if BVH_FEATURE(BVH_HAIR)
 					dir4 = sse3f(ssef(dir.x), ssef(dir.y), ssef(dir.z));
@@ -450,9 +449,9 @@ ccl_device bool BVH_FUNCTION_FULL_NAME(QBVH)(KernelGlobals *kg,
 			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; }
+			qbvh_near_far_idx_calc(idir,
+			                       &near_x, &near_y, &near_z,
+			                       &far_x, &far_y, &far_z);
 			tfar = ssef(isect_t);
 #  if BVH_FEATURE(BVH_HAIR)
 			dir4 = sse3f(ssef(dir.x), ssef(dir.y), ssef(dir.z));
diff --git a/intern/cycles/kernel/bvh/qbvh_subsurface.h b/intern/cycles/kernel/bvh/qbvh_subsurface.h
index 24aca96a298..ccd36df034a 100644
--- a/intern/cycles/kernel/bvh/qbvh_subsurface.h
+++ b/intern/cycles/kernel/bvh/qbvh_subsurface.h
@@ -101,10 +101,9 @@ ccl_device void BVH_FUNCTION_FULL_NAME(QBVH)(KernelGlobals *kg,
 	/* Offsets to select the side that becomes the lower or upper bound. */
 	int near_x, near_y, near_z;
 	int far_x, far_y, far_z;
-
-	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; }
+	qbvh_near_far_idx_calc(idir,
+	                       &near_x, &near_y, &near_z,
+	                       &far_x, &far_y, &far_z);
 
 	IsectPrecalc isect_precalc;
 	triangle_intersect_precalc(dir, &isect_precalc);
diff --git a/intern/cycles/kernel/bvh/qbvh_volume.h b/intern/cycles/kernel/bvh/qbvh_volume.h
index db9779351d2..424710b69f2 100644
--- a/intern/cycles/kernel/bvh/qbvh_volume.h
+++ b/intern/cycles/kernel/bvh/qbvh_volume.h
@@ -87,10 +87,9 @@ ccl_device bool BVH_FUNCTION_FULL_NAME(QBVH)(KernelGlobals *kg,
 	/* Offsets to select the side that becomes the lower or upper bound. */
 	int near_x, near_y, near_z;
 	int far_x, far_y, far_z;
-
-	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; }
+	qbvh_near_far_idx_calc(idir,
+	                       &near_x, &near_y, &near_z,
+	                       &far_x, &far_y, &far_z);
 
 	IsectPrecalc isect_precalc;
 	triangle_intersect_precalc(dir, &isect_precalc);
@@ -303,9 +302,9 @@ ccl_device bool BVH_FUNCTION_FULL_NAME(QBVH)(KernelGlobals *kg,
 						bvh_instance_push(kg, object, ray, &P, &dir, &idir, &isect->t);
 #  endif
 
-						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; }
+						qbvh_near_far_idx_calc(idir,
+						                       &near_x, &near_y, &near_z,
+						                       &far_x, &far_y, &far_z);
 						tfar = ssef(isect->t);
 #  if BVH_FEATURE(BVH_HAIR)
 						dir4 = sse3f(ssef(dir.x), ssef(dir.y), ssef(dir.z));
@@ -349,9 +348,9 @@ ccl_device bool BVH_FUNCTION_FULL_NAME(QBVH)(KernelGlobals *kg,
 			bvh_instance_pop(kg, object, ray, &P, &dir, &idir, &isect->t);
 #  endif
 
-			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; }
+			qbvh_near_far_idx_calc(idir,
+			                       &near_x, &near_y, &near_z,
+			                       &far_x, &far_y, &far_z);
 			tfar = ssef(isect->t);
 #  if BVH_FEATURE(BVH_HAIR)
 			dir4 = sse3f(ssef(dir.x), ssef(dir.y), ssef(dir.z));
diff --git a/intern/cycles/kernel/bvh/qbvh_volume_all.h b/intern/cycles/kernel/bvh/qbvh_volume_all.h
index 88f1f764e4c..eb48af6fc68 100644
--- a/intern/cycles/kernel/bvh/qbvh_volume_all.h
+++ b/intern/cycles/kernel/bvh/qbvh_volume_all.h
@@ -91,10 +91,9 @@ ccl_device uint BVH_FUNCTION_FULL_NAME(QBVH)(KernelGlobals *kg,
 	/* Offsets to select the side that becomes the lower or upper bound. */
 	int near_x, near_y, near_z;
 	int far_x, far_y, far_z;
-
-	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; }
+	qbvh_near_far_idx_calc(idir,
+	                       &near_x, &near_y, &near_z,
+	                       &far_x, &far_y, &far_z);
 
 	IsectPrecalc isect_precalc;
 	triangle_intersect_precalc(dir, &isect_precalc);
@@ -354,9 +353,9 @@ ccl_device uint BVH_FUNCTION_FULL_NAME(QBVH)(KernelGlobals *kg,
 						bvh_instance_push(kg, object, ray, &P, &dir, &idir, &isect_t);
 #  endif
 
-						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; }
+						qbvh_near_far_idx_calc(idir,
+						                       &near_x, &near_y, &near_z,
+						                       &far_x, &far_y, &far_z);
 						tfar = ssef(isect_t);
 						idir4 = sse3f(ssef(idir.x), ssef(idir.y), ssef(idir.z));
 #  if BVH_FEATURE(BVH_HAIR)
@@ -420,9 +419,9 @@ ccl_device uint BVH_FUNCTION_FULL_NAME(QBVH)(KernelGlobals *kg,
 			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; }
+			qbvh_near_far_idx_calc(idir,
+			                       &near_x, &near_y, &near_z,
+			                       &far_x, &far_y, &far_z);
 			tfar = ssef(isect_t);
 #  if BVH_FEATURE(BVH_HAIR)
 			dir4 = sse3f(ssef(dir.x), ssef(dir.y), ssef(dir.z));