Cycles: don't count volume boundaries as transparent bounces.

This is more important now that we will have tigther volume bounds that
we hit multiple times. It also avoids some noise due to RR previously
affecting these surfaces, which shouldn't have been the case and should
eventually be fixed for transparent BSDFs as well.

For non-volume scenes I found no performance impact on NVIDIA or AMD.
For volume scenes the noise decrease and fixed artifacts are worth the
little extra render time, when there is any.

8 files changed, 124 insertions, 45 deletions

diff --git a/intern/cycles/kernel/kernel_path.h b/intern/cycles/kernel/kernel_path.h
index aef350b0658..b0f53aef2d5 100644
--- a/intern/cycles/kernel/kernel_path.h
+++ b/intern/cycles/kernel/kernel_path.h
@@ -445,8 +445,15 @@ ccl_device void kernel_path_indirect(KernelGlobals *kg,
 			break;
 		}
 
-		/* Setup and evaluate shader. */
+		/* Setup shader data. */
 		shader_setup_from_ray(kg, sd, &isect, ray);
+
+		/* Skip most work for volume bounding surface. */
+#ifdef __VOLUME__
+		if(!(sd->flag & SD_HAS_ONLY_VOLUME)) {
+#endif
+
+		/* Evaluate shader. */
 		shader_eval_surface(kg, sd, state, state->flag);
 		shader_prepare_closures(sd, state);
 
@@ -523,6 +530,10 @@ ccl_device void kernel_path_indirect(KernelGlobals *kg,
 		}
 #endif  /* defined(__EMISSION__) */
 
+#ifdef __VOLUME__
+		}
+#endif
+
 		if(!kernel_path_surface_bounce(kg, sd, &throughput, state, &L->state, ray))
 			break;
 	}
@@ -605,8 +616,15 @@ ccl_device_forceinline void kernel_path_integrate(
 			break;
 		}
 
-		/* Setup and evaluate shader. */
+		/* Setup shader data. */
 		shader_setup_from_ray(kg, &sd, &isect, ray);
+
+		/* Skip most work for volume bounding surface. */
+#ifdef __VOLUME__
+		if(!(sd.flag & SD_HAS_ONLY_VOLUME)) {
+#endif
+
+		/* Evaluate shader. */
 		shader_eval_surface(kg, &sd, state, state->flag);
 		shader_prepare_closures(&sd, state);
 
@@ -669,6 +687,10 @@ ccl_device_forceinline void kernel_path_integrate(
 		/* direct lighting */
 		kernel_path_surface_connect_light(kg, &sd, emission_sd, throughput, state, L);
 
+#ifdef __VOLUME__
+		}
+#endif
+
 		/* compute direct lighting and next bounce */
 		if(!kernel_path_surface_bounce(kg, &sd, &throughput, state, &L->state, ray))
 			break;
diff --git a/intern/cycles/kernel/kernel_path_branched.h b/intern/cycles/kernel/kernel_path_branched.h
index 441a06eeba3..d43d418db29 100644
--- a/intern/cycles/kernel/kernel_path_branched.h
+++ b/intern/cycles/kernel/kernel_path_branched.h
@@ -480,6 +480,12 @@ ccl_device void kernel_branched_path_integrate(KernelGlobals *kg,
 
 		/* Setup and evaluate shader. */
 		shader_setup_from_ray(kg, &sd, &isect, &ray);
+
+		/* Skip most work for volume bounding surface. */
+#ifdef __VOLUME__
+		if(!(sd.flag & SD_HAS_ONLY_VOLUME)) {
+#endif
+
 		shader_eval_surface(kg, &sd, &state, state.flag);
 		shader_merge_closures(&sd);
 
@@ -533,37 +539,46 @@ ccl_device void kernel_branched_path_integrate(KernelGlobals *kg,
 		}
 #endif  /* __SUBSURFACE__ */
 
-		if(!(sd.flag & SD_HAS_ONLY_VOLUME)) {
-			PathState hit_state = state;
+		PathState hit_state = state;
 
 #ifdef __EMISSION__
-			/* direct light */
-			if(kernel_data.integrator.use_direct_light) {
-				int all = (kernel_data.integrator.sample_all_lights_direct) ||
-				          (state.flag & PATH_RAY_SHADOW_CATCHER);
-				kernel_branched_path_surface_connect_light(kg,
-					&sd, emission_sd, &hit_state, throughput, 1.0f, L, all);
-			}
+		/* direct light */
+		if(kernel_data.integrator.use_direct_light) {
+			int all = (kernel_data.integrator.sample_all_lights_direct) ||
+					  (state.flag & PATH_RAY_SHADOW_CATCHER);
+			kernel_branched_path_surface_connect_light(kg,
+				&sd, emission_sd, &hit_state, throughput, 1.0f, L, all);
+		}
 #endif  /* __EMISSION__ */
 
-			/* indirect light */
-			kernel_branched_path_surface_indirect_light(kg,
-				&sd, &indirect_sd, emission_sd, throughput, 1.0f, &hit_state, L);
+		/* indirect light */
+		kernel_branched_path_surface_indirect_light(kg,
+			&sd, &indirect_sd, emission_sd, throughput, 1.0f, &hit_state, L);
 
-			/* continue in case of transparency */
-			throughput *= shader_bsdf_transparency(kg, &sd);
+		/* continue in case of transparency */
+		throughput *= shader_bsdf_transparency(kg, &sd);
 
-			if(is_zero(throughput))
-				break;
-		}
+		if(is_zero(throughput))
+			break;
 
 		/* Update Path State */
 		path_state_next(kg, &state, LABEL_TRANSPARENT);
 
+#ifdef __VOLUME__
+		}
+		else {
+			/* For volume bounding meshes we pass through without counting transparent
+			 * bounces, only sanity check in case self intersection gets us stuck. */
+			state.volume_bounds_bounce++;
+			if (state.volume_bounds_bounce > VOLUME_BOUNDS_MAX) {
+				break;
+			}
+		}
+#endif
+
 		ray.P = ray_offset(sd.P, -sd.Ng);
 		ray.t -= sd.ray_length; /* clipping works through transparent */
 
-
 #ifdef __RAY_DIFFERENTIALS__
 		ray.dP = sd.dP;
 		ray.dD.dx = -sd.dI.dx;
diff --git a/intern/cycles/kernel/kernel_path_state.h b/intern/cycles/kernel/kernel_path_state.h
index 15d81fcddf4..ff7d1307a6c 100644
--- a/intern/cycles/kernel/kernel_path_state.h
+++ b/intern/cycles/kernel/kernel_path_state.h
@@ -55,6 +55,7 @@ ccl_device_inline void path_state_init(KernelGlobals *kg,
 
 #ifdef __VOLUME__
 	state->volume_bounce = 0;
+	state->volume_bounds_bounce = 0;
 
 	if(kernel_data.integrator.use_volumes) {
 		/* Initialize volume stack with volume we are inside of. */
diff --git a/intern/cycles/kernel/kernel_path_surface.h b/intern/cycles/kernel/kernel_path_surface.h
index 7b566b01b04..bca346d5ee0 100644
--- a/intern/cycles/kernel/kernel_path_surface.h
+++ b/intern/cycles/kernel/kernel_path_surface.h
@@ -329,10 +329,12 @@ ccl_device bool kernel_path_surface_bounce(KernelGlobals *kg,
 	}
 #ifdef __VOLUME__
 	else if(sd->flag & SD_HAS_ONLY_VOLUME) {
-		/* no surface shader but have a volume shader? act transparent */
-
-		/* update path state, count as transparent */
-		path_state_next(kg, state, LABEL_TRANSPARENT);
+		/* For volume bounding meshes we pass through without counting transparent
+		 * bounces, only sanity check in case self intersection gets us stuck. */
+		state->volume_bounds_bounce++;
+		if (state->volume_bounds_bounce > VOLUME_BOUNDS_MAX) {
+			return false;
+		}
 
 		if(state->bounce == 0)
 			ray->t -= sd->ray_length; /* clipping works through transparent */
diff --git a/intern/cycles/kernel/kernel_types.h b/intern/cycles/kernel/kernel_types.h
index d967edca75d..2a437cdbdc6 100644
--- a/intern/cycles/kernel/kernel_types.h
+++ b/intern/cycles/kernel/kernel_types.h
@@ -49,6 +49,8 @@ CCL_NAMESPACE_BEGIN
 #define BSSRDF_MAX_BOUNCES			256
 #define LOCAL_MAX_HITS				4
 
+#define VOLUME_BOUNDS_MAX       1024
+
 #define BECKMANN_TABLE_SIZE		256
 
 #define SHADER_NONE				(~0)
@@ -1107,6 +1109,7 @@ typedef struct PathState {
 	/* volume rendering */
 #ifdef __VOLUME__
 	int volume_bounce;
+	int volume_bounds_bounce;
 	uint rng_congruential;
 	VolumeStack volume_stack[VOLUME_STACK_SIZE];
 #endif
@@ -1497,8 +1500,10 @@ enum RayState {
 	RAY_ACTIVE,
 	/* Denotes ray has completed processing all samples and is inactive. */
 	RAY_INACTIVE,
-	/* Denoted ray has exited path-iteration and needs to update output buffer. */
+	/* Denotes ray has exited path-iteration and needs to update output buffer. */
 	RAY_UPDATE_BUFFER,
+	/* Denotes ray needs to skip most surface shader work. */
+	RAY_HAS_ONLY_VOLUME,
 	/* Donotes ray has hit background */
 	RAY_HIT_BACKGROUND,
 	/* Denotes ray has to be regenerated */
diff --git a/intern/cycles/kernel/split/kernel_next_iteration_setup.h b/intern/cycles/kernel/split/kernel_next_iteration_setup.h
index 81024f0cf99..8092419e796 100644
--- a/intern/cycles/kernel/split/kernel_next_iteration_setup.h
+++ b/intern/cycles/kernel/split/kernel_next_iteration_setup.h
@@ -53,39 +53,52 @@ ccl_device_inline void kernel_split_branched_indirect_light_init(KernelGlobals *
 	ADD_RAY_FLAG(kernel_split_state.ray_state, ray_index, RAY_BRANCHED_LIGHT_INDIRECT);
 }
 
-ccl_device void kernel_split_branched_indirect_light_end(KernelGlobals *kg, int ray_index)
+ccl_device void kernel_split_branched_transparent_bounce(KernelGlobals *kg, int ray_index)
 {
-	kernel_split_branched_path_indirect_loop_end(kg, ray_index);
-
 	ccl_global float3 *throughput = &kernel_split_state.throughput[ray_index];
 	ShaderData *sd = kernel_split_sd(sd, ray_index);
 	ccl_global PathState *state = &kernel_split_state.path_state[ray_index];
 	ccl_global Ray *ray = &kernel_split_state.ray[ray_index];
 
-	/* continue in case of transparency */
-	*throughput *= shader_bsdf_transparency(kg, sd);
+#  ifdef __VOLUME__
+	if(!(sd->flag & SD_HAS_ONLY_VOLUME)) {
+#  endif
+		/* continue in case of transparency */
+		*throughput *= shader_bsdf_transparency(kg, sd);
+
+		if(is_zero(*throughput)) {
+			kernel_split_path_end(kg, ray_index);
+			return;
+		}
 
-	if(is_zero(*throughput)) {
-		kernel_split_path_end(kg, ray_index);
-	}
-	else {
 		/* Update Path State */
 		path_state_next(kg, state, LABEL_TRANSPARENT);
+#  ifdef __VOLUME__
+	}
+	else {
+		/* For volume bounding meshes we pass through without counting transparent
+		 * bounces, only sanity check in case self intersection gets us stuck. */
+		state->volume_bounds_bounce++;
+		if (state->volume_bounds_bounce > VOLUME_BOUNDS_MAX) {
+			kernel_split_path_end(kg, ray_index);
+			return;
+		}
+	}
+#  endif
 
-		ray->P = ray_offset(sd->P, -sd->Ng);
-		ray->t -= sd->ray_length; /* clipping works through transparent */
+	ray->P = ray_offset(sd->P, -sd->Ng);
+	ray->t -= sd->ray_length; /* clipping works through transparent */
 
 #  ifdef __RAY_DIFFERENTIALS__
-		ray->dP = sd->dP;
-		ray->dD.dx = -sd->dI.dx;
-		ray->dD.dy = -sd->dI.dy;
+	ray->dP = sd->dP;
+	ray->dD.dx = -sd->dI.dx;
+	ray->dD.dy = -sd->dI.dy;
 #  endif  /* __RAY_DIFFERENTIALS__ */
 
 #  ifdef __VOLUME__
-		/* enter/exit volume */
-		kernel_volume_stack_enter_exit(kg, sd, state->volume_stack);
+	/* enter/exit volume */
+	kernel_volume_stack_enter_exit(kg, sd, state->volume_stack);
 #  endif  /* __VOLUME__ */
-	}
 }
 #endif  /* __BRANCHED_PATH__ */
 
@@ -121,6 +134,13 @@ ccl_device void kernel_next_iteration_setup(KernelGlobals *kg,
 
 	ccl_global char *ray_state = kernel_split_state.ray_state;
 
+#  ifdef __VOLUME__
+	/* Reactivate only volume rays here, most surface work was skipped. */
+	if(IS_STATE(ray_state, ray_index, RAY_HAS_ONLY_VOLUME)) {
+		ASSIGN_RAY_STATE(ray_state, ray_index, RAY_ACTIVE);
+	}
+#  endif
+
 	bool active = IS_STATE(ray_state, ray_index, RAY_ACTIVE);
 	if(active) {
 		ccl_global float3 *throughput = &kernel_split_state.throughput[ray_index];
@@ -138,6 +158,9 @@ ccl_device void kernel_next_iteration_setup(KernelGlobals *kg,
 			}
 #ifdef __BRANCHED_PATH__
 		}
+		else if(sd->flag & SD_HAS_ONLY_VOLUME) {
+			kernel_split_branched_transparent_bounce(kg, ray_index);
+		}
 		else {
 			kernel_split_branched_indirect_light_init(kg, ray_index);
 
@@ -151,7 +174,8 @@ ccl_device void kernel_next_iteration_setup(KernelGlobals *kg,
 				ASSIGN_RAY_STATE(ray_state, ray_index, RAY_REGENERATED);
 			}
 			else {
-				kernel_split_branched_indirect_light_end(kg, ray_index);
+				kernel_split_branched_path_indirect_loop_end(kg, ray_index);
+				kernel_split_branched_transparent_bounce(kg, ray_index);
 			}
 		}
 #endif  /* __BRANCHED_PATH__ */
@@ -196,7 +220,8 @@ ccl_device void kernel_next_iteration_setup(KernelGlobals *kg,
 			ASSIGN_RAY_STATE(ray_state, ray_index, RAY_REGENERATED);
 		}
 		else {
-			kernel_split_branched_indirect_light_end(kg, ray_index);
+			kernel_split_branched_path_indirect_loop_end(kg, ray_index);
+			kernel_split_branched_transparent_bounce(kg, ray_index);
 		}
 	}
 
diff --git a/intern/cycles/kernel/split/kernel_queue_enqueue.h b/intern/cycles/kernel/split/kernel_queue_enqueue.h
index 66ce2dfb6f1..df67fabab19 100644
--- a/intern/cycles/kernel/split/kernel_queue_enqueue.h
+++ b/intern/cycles/kernel/split/kernel_queue_enqueue.h
@@ -56,6 +56,7 @@ ccl_device void kernel_queue_enqueue(KernelGlobals *kg,
 		queue_number = QUEUE_HITBG_BUFF_UPDATE_TOREGEN_RAYS;
 	}
 	else if(IS_STATE(kernel_split_state.ray_state, ray_index, RAY_ACTIVE) ||
+	        IS_STATE(kernel_split_state.ray_state, ray_index, RAY_HAS_ONLY_VOLUME) ||
 	        IS_STATE(kernel_split_state.ray_state, ray_index, RAY_REGENERATED)) {
 		queue_number = QUEUE_ACTIVE_AND_REGENERATED_RAYS;
 	}
diff --git a/intern/cycles/kernel/split/kernel_shader_setup.h b/intern/cycles/kernel/split/kernel_shader_setup.h
index 9d428ee8139..ea3ec2ec83f 100644
--- a/intern/cycles/kernel/split/kernel_shader_setup.h
+++ b/intern/cycles/kernel/split/kernel_shader_setup.h
@@ -59,12 +59,20 @@ ccl_device void kernel_shader_setup(KernelGlobals *kg,
 	if(IS_STATE(kernel_split_state.ray_state, ray_index, RAY_ACTIVE)) {
 		Intersection isect = kernel_split_state.isect[ray_index];
 		Ray ray = kernel_split_state.ray[ray_index];
+		ShaderData *sd = kernel_split_sd(sd, ray_index);
 
 		shader_setup_from_ray(kg,
-		                      kernel_split_sd(sd, ray_index),
+		                      sd,
 		                      &isect,
 		                      &ray);
+
+#ifdef __VOLUME__
+		if(sd->flag & SD_HAS_ONLY_VOLUME) {
+			ASSIGN_RAY_STATE(kernel_split_state.ray_state, ray_index, RAY_HAS_ONLY_VOLUME);
+		}
+#endif
 	}
+
 }
 
 CCL_NAMESPACE_END