Cycles CUDA: reduce stack memory by reusing ShaderData.

57% less for path and 48% less for branched path.

14 files changed, 196 insertions, 169 deletions

diff --git a/intern/cycles/kernel/kernel_bake.h b/intern/cycles/kernel/kernel_bake.h
index 392cff9c281..77982ee548a 100644
--- a/intern/cycles/kernel/kernel_bake.h
+++ b/intern/cycles/kernel/kernel_bake.h
@@ -30,6 +30,9 @@ ccl_device void compute_light_pass(KernelGlobals *kg, ShaderData *sd, PathRadian
 	Ray ray;
 	float3 throughput = make_float3(1.0f, 1.0f, 1.0f);
 
+	/* emission shader data memory used by various functions */
+	ShaderData emission_sd;
+
 	ray.P = sd->P + sd->Ng;
 	ray.D = -sd->Ng;
 	ray.t = FLT_MAX;
@@ -41,7 +44,7 @@ ccl_device void compute_light_pass(KernelGlobals *kg, ShaderData *sd, PathRadian
 	path_radiance_init(&L_sample, kernel_data.film.use_light_pass);
 
 	/* init path state */
-	path_state_init(kg, &state, &rng, sample, NULL);
+	path_state_init(kg, &emission_sd, &state, &rng, sample, NULL);
 
 	/* evaluate surface shader */
 	float rbsdf = path_state_rng_1D(kg, &rng, &state, PRNG_BSDF);
@@ -56,7 +59,7 @@ ccl_device void compute_light_pass(KernelGlobals *kg, ShaderData *sd, PathRadian
 
 		/* sample ambient occlusion */
 		if(pass_filter & BAKE_FILTER_AO) {
-			kernel_path_ao(kg, sd, &L_sample, &state, &rng, throughput);
+			kernel_path_ao(kg, sd, &emission_sd, &L_sample, &state, &rng, throughput);
 		}
 
 		/* sample emission */
@@ -75,6 +78,7 @@ ccl_device void compute_light_pass(KernelGlobals *kg, ShaderData *sd, PathRadian
 			kernel_path_subsurface_init_indirect(&ss_indirect);
 			if(kernel_path_subsurface_scatter(kg,
 			                                  sd,
+			                                  &emission_sd,
 			                                  &L_sample,
 			                                  &state,
 			                                  &rng,
@@ -90,6 +94,7 @@ ccl_device void compute_light_pass(KernelGlobals *kg, ShaderData *sd, PathRadian
 					                                      &L_sample,
 					                                      &throughput);
 					kernel_path_indirect(kg,
+					                     &emission_sd,
 					                     &rng,
 					                     &ray,
 					                     throughput,
@@ -105,14 +110,14 @@ ccl_device void compute_light_pass(KernelGlobals *kg, ShaderData *sd, PathRadian
 
 		/* sample light and BSDF */
 		if(!is_sss_sample && (pass_filter & (BAKE_FILTER_DIRECT | BAKE_FILTER_INDIRECT))) {
-			kernel_path_surface_connect_light(kg, &rng, sd, throughput, &state, &L_sample);
+			kernel_path_surface_connect_light(kg, &rng, sd, &emission_sd, throughput, &state, &L_sample);
 
 			if(kernel_path_surface_bounce(kg, &rng, sd, &throughput, &state, &L_sample, &ray)) {
 #ifdef __LAMP_MIS__
 				state.ray_t = 0.0f;
 #endif
 				/* compute indirect light */
-				kernel_path_indirect(kg, &rng, &ray, throughput, 1, &state, &L_sample);
+				kernel_path_indirect(kg, &emission_sd, &rng, &ray, throughput, 1, &state, &L_sample);
 
 				/* sum and reset indirect light pass variables for the next samples */
 				path_radiance_sum_indirect(&L_sample);
@@ -126,7 +131,7 @@ ccl_device void compute_light_pass(KernelGlobals *kg, ShaderData *sd, PathRadian
 
 		/* sample ambient occlusion */
 		if(pass_filter & BAKE_FILTER_AO) {
-			kernel_branched_path_ao(kg, sd, &L_sample, &state, &rng, throughput);
+			kernel_branched_path_ao(kg, sd, &emission_sd, &L_sample, &state, &rng, throughput);
 		}
 
 		/* sample emission */
@@ -139,7 +144,7 @@ ccl_device void compute_light_pass(KernelGlobals *kg, ShaderData *sd, PathRadian
 		/* sample subsurface scattering */
 		if((pass_filter & BAKE_FILTER_SUBSURFACE) && (sd->flag & SD_BSSRDF)) {
 			/* when mixing BSSRDF and BSDF closures we should skip BSDF lighting if scattering was successful */
-			kernel_branched_path_subsurface_scatter(kg, sd, &L_sample, &state, &rng, &ray, throughput);
+			kernel_branched_path_subsurface_scatter(kg, sd, &emission_sd, &L_sample, &state, &rng, &ray, throughput);
 		}
 #endif
 
@@ -150,13 +155,13 @@ ccl_device void compute_light_pass(KernelGlobals *kg, ShaderData *sd, PathRadian
 			if(kernel_data.integrator.use_direct_light) {
 				int all = kernel_data.integrator.sample_all_lights_direct;
 				kernel_branched_path_surface_connect_light(kg, &rng,
-					sd, &state, throughput, 1.0f, &L_sample, all);
+					sd, &emission_sd, &state, throughput, 1.0f, &L_sample, all);
 			}
 #endif
 
 			/* indirect light */
 			kernel_branched_path_surface_indirect_light(kg, &rng,
-				sd, throughput, 1.0f, &state, &L_sample);
+				sd, &emission_sd, throughput, 1.0f, &state, &L_sample);
 		}
 	}
 #endif
diff --git a/intern/cycles/kernel/kernel_emission.h b/intern/cycles/kernel/kernel_emission.h
index 5cf52f9d176..4de8e0f698a 100644
--- a/intern/cycles/kernel/kernel_emission.h
+++ b/intern/cycles/kernel/kernel_emission.h
@@ -18,6 +18,7 @@ CCL_NAMESPACE_BEGIN
 
 /* Direction Emission */
 ccl_device_noinline float3 direct_emissive_eval(KernelGlobals *kg,
+                                                ShaderData *emission_sd,
                                                 LightSample *ls,
                                                 ccl_addr_space PathState *state,
                                                 float3 I,
@@ -26,12 +27,6 @@ ccl_device_noinline float3 direct_emissive_eval(KernelGlobals *kg,
                                                 float time)
 {
 	/* setup shading at emitter */
-#ifdef __SPLIT_KERNEL__
-	ShaderData *sd = kg->sd_input;
-#else
-	ShaderData sd_object;
-	ShaderData *sd = &sd_object;
-#endif
 	float3 eval;
 
 #ifdef __BACKGROUND_MIS__
@@ -46,28 +41,28 @@ ccl_device_noinline float3 direct_emissive_eval(KernelGlobals *kg,
 		ray.dP = differential3_zero();
 		ray.dD = dI;
 
-		shader_setup_from_background(kg, sd, &ray);
+		shader_setup_from_background(kg, emission_sd, &ray);
 
 		path_state_modify_bounce(state, true);
-		eval = shader_eval_background(kg, sd, state, 0, SHADER_CONTEXT_EMISSION);
+		eval = shader_eval_background(kg, emission_sd, state, 0, SHADER_CONTEXT_EMISSION);
 		path_state_modify_bounce(state, false);
 	}
 	else
 #endif
 	{
-		shader_setup_from_sample(kg, sd, ls->P, ls->Ng, I, ls->shader, ls->object, ls->prim, ls->u, ls->v, t, time);
+		shader_setup_from_sample(kg, emission_sd, ls->P, ls->Ng, I, ls->shader, ls->object, ls->prim, ls->u, ls->v, t, time);
 
-		ls->Ng = ccl_fetch(sd, Ng);
+		ls->Ng = ccl_fetch(emission_sd, Ng);
 
 		/* no path flag, we're evaluating this for all closures. that's weak but
 		 * we'd have to do multiple evaluations otherwise */
 		path_state_modify_bounce(state, true);
-		shader_eval_surface(kg, sd, state, 0.0f, 0, SHADER_CONTEXT_EMISSION);
+		shader_eval_surface(kg, emission_sd, state, 0.0f, 0, SHADER_CONTEXT_EMISSION);
 		path_state_modify_bounce(state, false);
 
 		/* evaluate emissive closure */
-		if(ccl_fetch(sd, flag) & SD_EMISSION)
-			eval = shader_emissive_eval(kg, sd);
+		if(ccl_fetch(emission_sd, flag) & SD_EMISSION)
+			eval = shader_emissive_eval(kg, emission_sd);
 		else
 			eval = make_float3(0.0f, 0.0f, 0.0f);
 	}
@@ -79,6 +74,7 @@ ccl_device_noinline float3 direct_emissive_eval(KernelGlobals *kg,
 
 ccl_device_noinline bool direct_emission(KernelGlobals *kg,
                                          ShaderData *sd,
+                                         ShaderData *emission_sd,
                                          LightSample *ls,
                                          ccl_addr_space PathState *state,
                                          Ray *ray,
@@ -94,6 +90,7 @@ ccl_device_noinline bool direct_emission(KernelGlobals *kg,
 	/* evaluate closure */
 
 	float3 light_eval = direct_emissive_eval(kg,
+	                                         emission_sd,
 	                                         ls,
 	                                         state,
 	                                         -ls->D,
@@ -198,6 +195,7 @@ ccl_device_noinline float3 indirect_primitive_emission(KernelGlobals *kg, Shader
 /* Indirect Lamp Emission */
 
 ccl_device_noinline bool indirect_lamp_emission(KernelGlobals *kg,
+                                                ShaderData *emission_sd,
                                                 ccl_addr_space PathState *state,
                                                 Ray *ray,
                                                 float3 *emission)
@@ -225,6 +223,7 @@ ccl_device_noinline bool indirect_lamp_emission(KernelGlobals *kg,
 #endif
 
 		float3 L = direct_emissive_eval(kg,
+		                                emission_sd,
 		                                &ls,
 		                                state,
 		                                -ray->D,
@@ -238,7 +237,7 @@ ccl_device_noinline bool indirect_lamp_emission(KernelGlobals *kg,
 			Ray volume_ray = *ray;
 			volume_ray.t = ls.t;
 			float3 volume_tp = make_float3(1.0f, 1.0f, 1.0f);
-			kernel_volume_shadow(kg, state, &volume_ray, &volume_tp);
+			kernel_volume_shadow(kg, emission_sd, state, &volume_ray, &volume_tp);
 			L *= volume_tp;
 		}
 #endif
@@ -260,6 +259,7 @@ ccl_device_noinline bool indirect_lamp_emission(KernelGlobals *kg,
 /* Indirect Background */
 
 ccl_device_noinline float3 indirect_background(KernelGlobals *kg,
+                                               ShaderData *emission_sd,
                                                ccl_addr_space PathState *state,
                                                ccl_addr_space Ray *ray)
 {
@@ -280,19 +280,14 @@ ccl_device_noinline float3 indirect_background(KernelGlobals *kg,
 	/* evaluate background closure */
 #  ifdef __SPLIT_KERNEL__
 	Ray priv_ray = *ray;
-	shader_setup_from_background(kg, kg->sd_input, &priv_ray);
-
-	path_state_modify_bounce(state, true);
-	float3 L = shader_eval_background(kg, kg->sd_input, state, state->flag, SHADER_CONTEXT_EMISSION);
-	path_state_modify_bounce(state, false);
+	shader_setup_from_background(kg, emission_sd, &priv_ray);
 #  else
-	ShaderData sd;
-	shader_setup_from_background(kg, &sd, ray);
+	shader_setup_from_background(kg, emission_sd, ray);
+#  endif
 
 	path_state_modify_bounce(state, true);
-	float3 L = shader_eval_background(kg, &sd, state, state->flag, SHADER_CONTEXT_EMISSION);
+	float3 L = shader_eval_background(kg, emission_sd, state, state->flag, SHADER_CONTEXT_EMISSION);
 	path_state_modify_bounce(state, false);
-#  endif
 
 #ifdef __BACKGROUND_MIS__
 	/* check if background light exists or if we should skip pdf */
diff --git a/intern/cycles/kernel/kernel_path.h b/intern/cycles/kernel/kernel_path.h
index c136c85df59..5527d8aa861 100644
--- a/intern/cycles/kernel/kernel_path.h
+++ b/intern/cycles/kernel/kernel_path.h
@@ -53,6 +53,7 @@
 CCL_NAMESPACE_BEGIN
 
 ccl_device void kernel_path_indirect(KernelGlobals *kg,
+                                     ShaderData *emission_sd,
                                      RNG *rng,
                                      Ray *ray,
                                      float3 throughput,
@@ -60,6 +61,9 @@ ccl_device void kernel_path_indirect(KernelGlobals *kg,
                                      PathState *state,
                                      PathRadiance *L)
 {
+	/* shader data memory used for both volumes and surfaces, saves stack space */
+	ShaderData sd;
+
 	/* path iteration */
 	for(;;) {
 		/* intersect scene */
@@ -87,7 +91,7 @@ ccl_device void kernel_path_indirect(KernelGlobals *kg,
 
 			/* intersect with lamp */
 			float3 emission;
-			if(indirect_lamp_emission(kg, state, &light_ray, &emission)) {
+			if(indirect_lamp_emission(kg, emission_sd, state, &light_ray, &emission)) {
 				path_radiance_accum_emission(L,
 				                             throughput,
 				                             emission,
@@ -115,15 +119,14 @@ ccl_device void kernel_path_indirect(KernelGlobals *kg,
 			if(decoupled) {
 				/* cache steps along volume for repeated sampling */
 				VolumeSegment volume_segment;
-				ShaderData volume_sd;
 
 				shader_setup_from_volume(kg,
-				                         &volume_sd,
+				                         &sd,
 				                         &volume_ray);
 				kernel_volume_decoupled_record(kg,
 				                               state,
 				                               &volume_ray,
-				                               &volume_sd,
+				                               &sd,
 				                               &volume_segment,
 				                               heterogeneous);
 
@@ -146,7 +149,8 @@ ccl_device void kernel_path_indirect(KernelGlobals *kg,
 					/* direct light sampling */
 					kernel_branched_path_volume_connect_light(kg,
 					                                          rng,
-					                                          &volume_sd,
+					                                          &sd,
+					                                          emission_sd,
 					                                          throughput,
 					                                          state,
 					                                          L,
@@ -163,7 +167,7 @@ ccl_device void kernel_path_indirect(KernelGlobals *kg,
 					result = kernel_volume_decoupled_scatter(kg,
 					                                         state,
 					                                         &volume_ray,
-					                                         &volume_sd,
+					                                         &sd,
 					                                         &throughput,
 					                                         rphase,
 					                                         rscatter,
@@ -178,7 +182,7 @@ ccl_device void kernel_path_indirect(KernelGlobals *kg,
 				if(result == VOLUME_PATH_SCATTERED) {
 					if(kernel_path_volume_bounce(kg,
 					                             rng,
-					                             &volume_sd,
+					                             &sd,
 					                             &throughput,
 					                             state,
 					                             L,
@@ -198,16 +202,16 @@ ccl_device void kernel_path_indirect(KernelGlobals *kg,
 #  endif
 			{
 				/* integrate along volume segment with distance sampling */
-				ShaderData volume_sd;
 				VolumeIntegrateResult result = kernel_volume_integrate(
-					kg, state, &volume_sd, &volume_ray, L, &throughput, rng, heterogeneous);
+					kg, state, &sd, &volume_ray, L, &throughput, rng, heterogeneous);
 
 #  ifdef __VOLUME_SCATTER__
 				if(result == VOLUME_PATH_SCATTERED) {
 					/* direct lighting */
 					kernel_path_volume_connect_light(kg,
 					                                 rng,
-					                                 &volume_sd,
+					                                 &sd,
+					                                 emission_sd,
 					                                 throughput,
 					                                 state,
 					                                 L);
@@ -215,7 +219,7 @@ ccl_device void kernel_path_indirect(KernelGlobals *kg,
 					/* indirect light bounce */
 					if(kernel_path_volume_bounce(kg,
 					                             rng,
-					                             &volume_sd,
+					                             &sd,
 					                             &throughput,
 					                             state,
 					                             L,
@@ -235,7 +239,7 @@ ccl_device void kernel_path_indirect(KernelGlobals *kg,
 		if(!hit) {
 #ifdef __BACKGROUND__
 			/* sample background shader */
-			float3 L_background = indirect_background(kg, state, ray);
+			float3 L_background = indirect_background(kg, emission_sd, state, ray);
 			path_radiance_accum_background(L,
 			                               throughput,
 			                               L_background,
@@ -246,7 +250,6 @@ ccl_device void kernel_path_indirect(KernelGlobals *kg,
 		}
 
 		/* setup shading */
-		ShaderData sd;
 		shader_setup_from_ray(kg,
 		                      &sd,
 		                      &isect,
@@ -328,7 +331,7 @@ ccl_device void kernel_path_indirect(KernelGlobals *kg,
 				light_ray.dP = sd.dP;
 				light_ray.dD = differential3_zero();
 
-				if(!shadow_blocked(kg, state, &light_ray, &ao_shadow)) {
+				if(!shadow_blocked(kg, emission_sd, state, &light_ray, &ao_shadow)) {
 					path_radiance_accum_ao(L,
 					                       throughput,
 					                       ao_alpha,
@@ -378,6 +381,7 @@ ccl_device void kernel_path_indirect(KernelGlobals *kg,
 			kernel_branched_path_surface_connect_light(kg,
 			                                           rng,
 			                                           &sd,
+			                                           emission_sd,
 			                                           state,
 			                                           throughput,
 			                                           1.0f,
@@ -393,6 +397,7 @@ ccl_device void kernel_path_indirect(KernelGlobals *kg,
 
 ccl_device_noinline void kernel_path_ao(KernelGlobals *kg,
                                         ShaderData *sd,
+                                        ShaderData *emission_sd,
                                         PathRadiance *L,
                                         PathState *state,
                                         RNG *rng,
@@ -425,7 +430,7 @@ ccl_device_noinline void kernel_path_ao(KernelGlobals *kg,
 		light_ray.dP = ccl_fetch(sd, dP);
 		light_ray.dD = differential3_zero();
 
-		if(!shadow_blocked(kg, state, &light_ray, &ao_shadow))
+		if(!shadow_blocked(kg, emission_sd, state, &light_ray, &ao_shadow))
 			path_radiance_accum_ao(L, throughput, ao_alpha, ao_bsdf, ao_shadow, state->bounce);
 	}
 }
@@ -435,6 +440,7 @@ ccl_device_noinline void kernel_path_ao(KernelGlobals *kg,
 ccl_device bool kernel_path_subsurface_scatter(
         KernelGlobals *kg,
         ShaderData *sd,
+        ShaderData *emission_sd,
         PathRadiance *L,
         PathState *state,
         RNG *rng,
@@ -503,7 +509,7 @@ ccl_device bool kernel_path_subsurface_scatter(
 			hit_L->direct_throughput = L->direct_throughput;
 			path_radiance_copy_indirect(hit_L, L);
 
-			kernel_path_surface_connect_light(kg, rng, sd, *hit_tp, state, hit_L);
+			kernel_path_surface_connect_light(kg, rng, sd, emission_sd, *hit_tp, state, hit_L);
 
 			if(kernel_path_surface_bounce(kg,
 			                              rng,
@@ -526,6 +532,7 @@ ccl_device bool kernel_path_subsurface_scatter(
 
 					kernel_volume_stack_update_for_subsurface(
 					    kg,
+					    emission_sd,
 					    &volume_ray,
 					    hit_state->volume_stack);
 				}
@@ -604,8 +611,13 @@ ccl_device_inline float4 kernel_path_integrate(KernelGlobals *kg,
 
 	path_radiance_init(&L, kernel_data.film.use_light_pass);
 
+	/* shader data memory used for both volumes and surfaces, saves stack space */
+	ShaderData sd;
+	/* shader data used by emission, shadows, volume stacks */
+	ShaderData emission_sd;
+
 	PathState state;
-	path_state_init(kg, &state, rng, sample, &ray);
+	path_state_init(kg, &emission_sd, &state, rng, sample, &ray);
 
 #ifdef __KERNEL_DEBUG__
 	DebugData debug_data;
@@ -669,7 +681,7 @@ ccl_device_inline float4 kernel_path_integrate(KernelGlobals *kg,
 			/* intersect with lamp */
 			float3 emission;
 
-			if(indirect_lamp_emission(kg, &state, &light_ray, &emission))
+			if(indirect_lamp_emission(kg, &emission_sd, &state, &light_ray, &emission))
 				path_radiance_accum_emission(&L, throughput, emission, state.bounce);
 		}
 #endif
@@ -689,11 +701,10 @@ ccl_device_inline float4 kernel_path_integrate(KernelGlobals *kg,
 			if(decoupled) {
 				/* cache steps along volume for repeated sampling */
 				VolumeSegment volume_segment;
-				ShaderData volume_sd;
 
-				shader_setup_from_volume(kg, &volume_sd, &volume_ray);
+				shader_setup_from_volume(kg, &sd, &volume_ray);
 				kernel_volume_decoupled_record(kg, &state,
-					&volume_ray, &volume_sd, &volume_segment, heterogeneous);
+					&volume_ray, &sd, &volume_segment, heterogeneous);
 
 				volume_segment.sampling_method = sampling_method;
 
@@ -708,8 +719,9 @@ ccl_device_inline float4 kernel_path_integrate(KernelGlobals *kg,
 					int all = false;
 
 					/* direct light sampling */
-					kernel_branched_path_volume_connect_light(kg, rng, &volume_sd,
-						throughput, &state, &L, all, &volume_ray, &volume_segment);
+					kernel_branched_path_volume_connect_light(kg, rng, &sd,
+						&emission_sd, throughput, &state, &L, all,
+						&volume_ray, &volume_segment);
 
 					/* indirect sample. if we use distance sampling and take just
 					 * one sample for direct and indirect light, we could share
@@ -718,7 +730,7 @@ ccl_device_inline float4 kernel_path_integrate(KernelGlobals *kg,
 					float rscatter = path_state_rng_1D_for_decision(kg, rng, &state, PRNG_SCATTER_DISTANCE);
 
 					result = kernel_volume_decoupled_scatter(kg,
-						&state, &volume_ray, &volume_sd, &throughput,
+						&state, &volume_ray, &sd, &throughput,
 						rphase, rscatter, &volume_segment, NULL, true);
 				}
 
@@ -726,7 +738,7 @@ ccl_device_inline float4 kernel_path_integrate(KernelGlobals *kg,
 				kernel_volume_decoupled_free(kg, &volume_segment);
 
 				if(result == VOLUME_PATH_SCATTERED) {
-					if(kernel_path_volume_bounce(kg, rng, &volume_sd, &throughput, &state, &L, &ray))
+					if(kernel_path_volume_bounce(kg, rng, &sd, &throughput, &state, &L, &ray))
 						continue;
 					else
 						break;
@@ -739,17 +751,16 @@ ccl_device_inline float4 kernel_path_integrate(KernelGlobals *kg,
 #  endif
 			{
 				/* integrate along volume segment with distance sampling */
-				ShaderData volume_sd;
 				VolumeIntegrateResult result = kernel_volume_integrate(
-					kg, &state, &volume_sd, &volume_ray, &L, &throughput, rng, heterogeneous);
+					kg, &state, &sd, &volume_ray, &L, &throughput, rng, heterogeneous);
 
 #  ifdef __VOLUME_SCATTER__
 				if(result == VOLUME_PATH_SCATTERED) {
 					/* direct lighting */
-					kernel_path_volume_connect_light(kg, rng, &volume_sd, throughput, &state, &L);
+					kernel_path_volume_connect_light(kg, rng, &sd, &emission_sd, throughput, &state, &L);
 
 					/* indirect light bounce */
-					if(kernel_path_volume_bounce(kg, rng, &volume_sd, &throughput, &state, &L, &ray))
+					if(kernel_path_volume_bounce(kg, rng, &sd, &throughput, &state, &L, &ray))
 						continue;
 					else
 						break;
@@ -772,7 +783,7 @@ ccl_device_inline float4 kernel_path_integrate(KernelGlobals *kg,
 
 #ifdef __BACKGROUND__
 			/* sample background shader */
-			float3 L_background = indirect_background(kg, &state, &ray);
+			float3 L_background = indirect_background(kg, &emission_sd, &state, &ray);
 			path_radiance_accum_background(&L, throughput, L_background, state.bounce);
 #endif
 
@@ -780,7 +791,6 @@ ccl_device_inline float4 kernel_path_integrate(KernelGlobals *kg,
 		}
 
 		/* setup shading */
-		ShaderData sd;
 		shader_setup_from_ray(kg, &sd, &isect, &ray);
 		float rbsdf = path_state_rng_1D_for_decision(kg, rng, &state, PRNG_BSDF);
 		shader_eval_surface(kg, &sd, &state, rbsdf, state.flag, SHADER_CONTEXT_MAIN);
@@ -848,7 +858,7 @@ ccl_device_inline float4 kernel_path_integrate(KernelGlobals *kg,
 #ifdef __AO__
 		/* ambient occlusion */
 		if(kernel_data.integrator.use_ambient_occlusion || (sd.flag & SD_AO)) {
-			kernel_path_ao(kg, &sd, &L, &state, rng, throughput);
+			kernel_path_ao(kg, &sd, &emission_sd, &L, &state, rng, throughput);
 		}
 #endif
 
@@ -858,6 +868,7 @@ ccl_device_inline float4 kernel_path_integrate(KernelGlobals *kg,
 		if(sd.flag & SD_BSSRDF) {
 			if(kernel_path_subsurface_scatter(kg,
 			                                  &sd,
+			                                  &emission_sd,
 			                                  &L,
 			                                  &state,
 			                                  rng,
@@ -871,7 +882,7 @@ ccl_device_inline float4 kernel_path_integrate(KernelGlobals *kg,
 #endif  /* __SUBSURFACE__ */
 
 		/* direct lighting */
-		kernel_path_surface_connect_light(kg, rng, &sd, throughput, &state, &L);
+		kernel_path_surface_connect_light(kg, rng, &sd, &emission_sd, throughput, &state, &L);
 
 		/* compute direct lighting and next bounce */
 		if(!kernel_path_surface_bounce(kg, rng, &sd, &throughput, &state, &L, &ray))
diff --git a/intern/cycles/kernel/kernel_path_branched.h b/intern/cycles/kernel/kernel_path_branched.h
index 13ae4cf669b..b4dee220aa5 100644
--- a/intern/cycles/kernel/kernel_path_branched.h
+++ b/intern/cycles/kernel/kernel_path_branched.h
@@ -18,7 +18,13 @@ CCL_NAMESPACE_BEGIN
 
 #ifdef __BRANCHED_PATH__
 
-ccl_device void kernel_branched_path_ao(KernelGlobals *kg, ShaderData *sd, PathRadiance *L, PathState *state, RNG *rng, float3 throughput)
+ccl_device void kernel_branched_path_ao(KernelGlobals *kg,
+                                        ShaderData *sd,
+                                        ShaderData *emission_sd,
+                                        PathRadiance *L,
+                                        PathState *state,
+                                        RNG *rng,
+                                        float3 throughput)
 {
 	int num_samples = kernel_data.integrator.ao_samples;
 	float num_samples_inv = 1.0f/num_samples;
@@ -49,7 +55,7 @@ ccl_device void kernel_branched_path_ao(KernelGlobals *kg, ShaderData *sd, PathR
 			light_ray.dP = ccl_fetch(sd, dP);
 			light_ray.dD = differential3_zero();
 
-			if(!shadow_blocked(kg, state, &light_ray, &ao_shadow))
+			if(!shadow_blocked(kg, emission_sd, state, &light_ray, &ao_shadow))
 				path_radiance_accum_ao(L, throughput*num_samples_inv, ao_alpha, ao_bsdf, ao_shadow, state->bounce);
 		}
 	}
@@ -58,8 +64,8 @@ ccl_device void kernel_branched_path_ao(KernelGlobals *kg, ShaderData *sd, PathR
 
 /* bounce off surface and integrate indirect light */
 ccl_device_noinline void kernel_branched_path_surface_indirect_light(KernelGlobals *kg,
-	RNG *rng, ShaderData *sd, float3 throughput, float num_samples_adjust,
-	PathState *state, PathRadiance *L)
+	RNG *rng, ShaderData *sd, ShaderData *emission_sd, float3 throughput,
+	float num_samples_adjust, PathState *state, PathRadiance *L)
 {
 	for(int i = 0; i < ccl_fetch(sd, num_closure); i++) {
 		const ShaderClosure *sc = &ccl_fetch(sd, closure)[i];
@@ -106,6 +112,7 @@ ccl_device_noinline void kernel_branched_path_surface_indirect_light(KernelGloba
 			}
 
 			kernel_path_indirect(kg,
+			                     emission_sd,
 			                     rng,
 			                     &bsdf_ray,
 			                     tp*num_samples_inv,
@@ -124,6 +131,7 @@ ccl_device_noinline void kernel_branched_path_surface_indirect_light(KernelGloba
 #ifdef __SUBSURFACE__
 ccl_device void kernel_branched_path_subsurface_scatter(KernelGlobals *kg,
                                                         ShaderData *sd,
+                                                        ShaderData *emission_sd,
                                                         PathRadiance *L,
                                                         PathState *state,
                                                         RNG *rng,
@@ -186,6 +194,7 @@ ccl_device void kernel_branched_path_subsurface_scatter(KernelGlobals *kg,
 
 					kernel_volume_stack_update_for_subsurface(
 					    kg,
+					    emission_sd,
 					    &volume_ray,
 					    hit_state.volume_stack);
 				}
@@ -199,6 +208,7 @@ ccl_device void kernel_branched_path_subsurface_scatter(KernelGlobals *kg,
 					        kg,
 					        rng,
 					        &bssrdf_sd,
+					        emission_sd,
 					        &hit_state,
 					        throughput,
 					        num_samples_inv,
@@ -212,6 +222,7 @@ ccl_device void kernel_branched_path_subsurface_scatter(KernelGlobals *kg,
 				        kg,
 				        rng,
 				        &bssrdf_sd,
+				        emission_sd,
 				        throughput,
 				        num_samples_inv,
 				        &hit_state,
@@ -231,8 +242,13 @@ ccl_device float4 kernel_branched_path_integrate(KernelGlobals *kg, RNG *rng, in
 
 	path_radiance_init(&L, kernel_data.film.use_light_pass);
 
+	/* shader data memory used for both volumes and surfaces, saves stack space */
+	ShaderData sd;
+	/* shader data used by emission, shadows, volume stacks */
+	ShaderData emission_sd;
+
 	PathState state;
-	path_state_init(kg, &state, rng, sample, &ray);
+	path_state_init(kg, &emission_sd, &state, rng, sample, &ray);
 
 #ifdef __KERNEL_DEBUG__
 	DebugData debug_data;
@@ -287,11 +303,10 @@ ccl_device float4 kernel_branched_path_integrate(KernelGlobals *kg, RNG *rng, in
 
 			/* cache steps along volume for repeated sampling */
 			VolumeSegment volume_segment;
-			ShaderData volume_sd;
 
-			shader_setup_from_volume(kg, &volume_sd, &volume_ray);
+			shader_setup_from_volume(kg, &sd, &volume_ray);
 			kernel_volume_decoupled_record(kg, &state,
-				&volume_ray, &volume_sd, &volume_segment, heterogeneous);
+				&volume_ray, &sd, &volume_segment, heterogeneous);
 
 			/* direct light sampling */
 			if(volume_segment.closure_flag & SD_SCATTER) {
@@ -299,8 +314,9 @@ ccl_device float4 kernel_branched_path_integrate(KernelGlobals *kg, RNG *rng, in
 
 				int all = kernel_data.integrator.sample_all_lights_direct;
 
-				kernel_branched_path_volume_connect_light(kg, rng, &volume_sd,
-					throughput, &state, &L, all, &volume_ray, &volume_segment);
+				kernel_branched_path_volume_connect_light(kg, rng, &sd,
+					&emission_sd, throughput, &state, &L, all,
+					&volume_ray, &volume_segment);
 
 				/* indirect light sampling */
 				int num_samples = kernel_data.integrator.volume_samples;
@@ -326,20 +342,21 @@ ccl_device float4 kernel_branched_path_integrate(KernelGlobals *kg, RNG *rng, in
 					float rscatter = path_state_rng_1D_for_decision(kg, &tmp_rng, &ps, PRNG_SCATTER_DISTANCE);
 
 					VolumeIntegrateResult result = kernel_volume_decoupled_scatter(kg,
-						&ps, &pray, &volume_sd, &tp, rphase, rscatter, &volume_segment, NULL, false);
+						&ps, &pray, &sd, &tp, rphase, rscatter, &volume_segment, NULL, false);
 
 					(void)result;
 					kernel_assert(result == VOLUME_PATH_SCATTERED);
 
 					if(kernel_path_volume_bounce(kg,
 					                             rng,
-					                             &volume_sd,
+					                             &sd,
 					                             &tp,
 					                             &ps,
 					                             &L,
 					                             &pray))
 					{
 						kernel_path_indirect(kg,
+						                     &emission_sd,
 						                     rng,
 						                     &pray,
 						                     tp*num_samples_inv,
@@ -373,30 +390,30 @@ ccl_device float4 kernel_branched_path_integrate(KernelGlobals *kg, RNG *rng, in
 			for(int j = 0; j < num_samples; j++) {
 				PathState ps = state;
 				Ray pray = ray;
-				ShaderData volume_sd;
 				float3 tp = throughput * num_samples_inv;
 
 				/* branch RNG state */
 				path_state_branch(&ps, j, num_samples);
 
 				VolumeIntegrateResult result = kernel_volume_integrate(
-					kg, &ps, &volume_sd, &volume_ray, &L, &tp, rng, heterogeneous);
+					kg, &ps, &sd, &volume_ray, &L, &tp, rng, heterogeneous);
 
 #ifdef __VOLUME_SCATTER__
 				if(result == VOLUME_PATH_SCATTERED) {
 					/* todo: support equiangular, MIS and all light sampling.
 					 * alternatively get decoupled ray marching working on the GPU */
-					kernel_path_volume_connect_light(kg, rng, &volume_sd, tp, &state, &L);
+					kernel_path_volume_connect_light(kg, rng, &sd, &emission_sd, tp, &state, &L);
 
 					if(kernel_path_volume_bounce(kg,
 					                             rng,
-					                             &volume_sd,
+					                             &sd,
 					                             &tp,
 					                             &ps,
 					                             &L,
 					                             &pray))
 					{
 						kernel_path_indirect(kg,
+						                     &emission_sd,
 						                     rng,
 						                     &pray,
 						                     tp,
@@ -414,7 +431,7 @@ ccl_device float4 kernel_branched_path_integrate(KernelGlobals *kg, RNG *rng, in
 			}
 
 			/* todo: avoid this calculation using decoupled ray marching */
-			kernel_volume_shadow(kg, &state, &volume_ray, &throughput);
+			kernel_volume_shadow(kg, &emission_sd, &state, &volume_ray, &throughput);
 #endif
 		}
 #endif
@@ -432,7 +449,7 @@ ccl_device float4 kernel_branched_path_integrate(KernelGlobals *kg, RNG *rng, in
 
 #ifdef __BACKGROUND__
 			/* sample background shader */
-			float3 L_background = indirect_background(kg, &state, &ray);
+			float3 L_background = indirect_background(kg, &emission_sd, &state, &ray);
 			path_radiance_accum_background(&L, throughput, L_background, state.bounce);
 #endif
 
@@ -440,7 +457,6 @@ ccl_device float4 kernel_branched_path_integrate(KernelGlobals *kg, RNG *rng, in
 		}
 
 		/* setup shading */
-		ShaderData sd;
 		shader_setup_from_ray(kg, &sd, &isect, &ray);
 		shader_eval_surface(kg, &sd, &state, 0.0f, state.flag, SHADER_CONTEXT_MAIN);
 		shader_merge_closures(&sd);
@@ -499,14 +515,14 @@ ccl_device float4 kernel_branched_path_integrate(KernelGlobals *kg, RNG *rng, in
 #ifdef __AO__
 		/* ambient occlusion */
 		if(kernel_data.integrator.use_ambient_occlusion || (sd.flag & SD_AO)) {
-			kernel_branched_path_ao(kg, &sd, &L, &state, rng, throughput);
+			kernel_branched_path_ao(kg, &sd, &emission_sd, &L, &state, rng, throughput);
 		}
 #endif
 
 #ifdef __SUBSURFACE__
 		/* bssrdf scatter to a different location on the same object */
 		if(sd.flag & SD_BSSRDF) {
-			kernel_branched_path_subsurface_scatter(kg, &sd, &L, &state,
+			kernel_branched_path_subsurface_scatter(kg, &sd, &emission_sd, &L, &state,
 			                                        rng, &ray, throughput);
 		}
 #endif
@@ -519,13 +535,13 @@ ccl_device float4 kernel_branched_path_integrate(KernelGlobals *kg, RNG *rng, in
 			if(kernel_data.integrator.use_direct_light) {
 				int all = kernel_data.integrator.sample_all_lights_direct;
 				kernel_branched_path_surface_connect_light(kg, rng,
-					&sd, &hit_state, throughput, 1.0f, &L, all);
+					&sd, &emission_sd, &hit_state, throughput, 1.0f, &L, all);
 			}
 #endif
 
 			/* indirect light */
 			kernel_branched_path_surface_indirect_light(kg, rng,
-				&sd, throughput, 1.0f, &hit_state, &L);
+				&sd, &emission_sd, throughput, 1.0f, &hit_state, &L);
 
 			/* continue in case of transparency */
 			throughput *= shader_bsdf_transparency(kg, &sd);
diff --git a/intern/cycles/kernel/kernel_path_state.h b/intern/cycles/kernel/kernel_path_state.h
index ef3765f7d89..e0e35d792ab 100644
--- a/intern/cycles/kernel/kernel_path_state.h
+++ b/intern/cycles/kernel/kernel_path_state.h
@@ -16,7 +16,12 @@
 
 CCL_NAMESPACE_BEGIN
 
-ccl_device_inline void path_state_init(KernelGlobals *kg, ccl_addr_space PathState *state, ccl_addr_space RNG *rng, int sample, ccl_addr_space Ray *ray)
+ccl_device_inline void path_state_init(KernelGlobals *kg,
+                                       ShaderData *stack_sd,
+                                       ccl_addr_space PathState *state,
+                                       ccl_addr_space RNG *rng,
+                                       int sample,
+                                       ccl_addr_space Ray *ray)
 {
 	state->flag = PATH_RAY_CAMERA|PATH_RAY_MIS_SKIP;
 
@@ -41,7 +46,7 @@ ccl_device_inline void path_state_init(KernelGlobals *kg, ccl_addr_space PathSta
 
 	if(kernel_data.integrator.use_volumes) {
 		/* initialize volume stack with volume we are inside of */
-		kernel_volume_stack_init(kg, ray, state->volume_stack);
+		kernel_volume_stack_init(kg, stack_sd, ray, state->volume_stack);
 		/* seed RNG for cases where we can't use stratified samples */
 		state->rng_congruential = lcg_init(*rng + sample*0x51633e2d);
 	}
diff --git a/intern/cycles/kernel/kernel_path_surface.h b/intern/cycles/kernel/kernel_path_surface.h
index 1818c4fd2da..74b1ae0ca32 100644
--- a/intern/cycles/kernel/kernel_path_surface.h
+++ b/intern/cycles/kernel/kernel_path_surface.h
@@ -20,7 +20,8 @@ CCL_NAMESPACE_BEGIN
 
 /* branched path tracing: connect path directly to position on one or more lights and add it to L */
 ccl_device_noinline void kernel_branched_path_surface_connect_light(KernelGlobals *kg, RNG *rng,
-	ShaderData *sd, PathState *state, float3 throughput, float num_samples_adjust, PathRadiance *L, int sample_all_lights)
+	ShaderData *sd, ShaderData *emission_sd, PathState *state, float3 throughput,
+	float num_samples_adjust, PathRadiance *L, int sample_all_lights)
 {
 #ifdef __EMISSION__
 	/* sample illumination from lights to find path contribution */
@@ -55,11 +56,11 @@ ccl_device_noinline void kernel_branched_path_surface_connect_light(KernelGlobal
 				LightSample ls;
 				lamp_light_sample(kg, i, light_u, light_v, ccl_fetch(sd, P), &ls);
 
-				if(direct_emission(kg, sd, &ls, state, &light_ray, &L_light, &is_lamp)) {
+				if(direct_emission(kg, sd, emission_sd, &ls, state, &light_ray, &L_light, &is_lamp)) {
 					/* trace shadow ray */
 					float3 shadow;
 
-					if(!shadow_blocked(kg, state, &light_ray, &shadow)) {
+					if(!shadow_blocked(kg, emission_sd, state, &light_ray, &shadow)) {
 						/* accumulate */
 						path_radiance_accum_light(L, throughput*num_samples_inv, &L_light, shadow, num_samples_inv, state->bounce, is_lamp);
 					}
@@ -87,11 +88,11 @@ ccl_device_noinline void kernel_branched_path_surface_connect_light(KernelGlobal
 				LightSample ls;
 				light_sample(kg, light_t, light_u, light_v, ccl_fetch(sd, time), ccl_fetch(sd, P), state->bounce, &ls);
 
-				if(direct_emission(kg, sd, &ls, state, &light_ray, &L_light, &is_lamp)) {
+				if(direct_emission(kg, sd, emission_sd, &ls, state, &light_ray, &L_light, &is_lamp)) {
 					/* trace shadow ray */
 					float3 shadow;
 
-					if(!shadow_blocked(kg, state, &light_ray, &shadow)) {
+					if(!shadow_blocked(kg, emission_sd, state, &light_ray, &shadow)) {
 						/* accumulate */
 						path_radiance_accum_light(L, throughput*num_samples_inv, &L_light, shadow, num_samples_inv, state->bounce, is_lamp);
 					}
@@ -109,11 +110,11 @@ ccl_device_noinline void kernel_branched_path_surface_connect_light(KernelGlobal
 		light_sample(kg, light_t, light_u, light_v, ccl_fetch(sd, time), ccl_fetch(sd, P), state->bounce, &ls);
 
 		/* sample random light */
-		if(direct_emission(kg, sd, &ls, state, &light_ray, &L_light, &is_lamp)) {
+		if(direct_emission(kg, sd, emission_sd, &ls, state, &light_ray, &L_light, &is_lamp)) {
 			/* trace shadow ray */
 			float3 shadow;
 
-			if(!shadow_blocked(kg, state, &light_ray, &shadow)) {
+			if(!shadow_blocked(kg, emission_sd, state, &light_ray, &shadow)) {
 				/* accumulate */
 				path_radiance_accum_light(L, throughput*num_samples_adjust, &L_light, shadow, num_samples_adjust, state->bounce, is_lamp);
 			}
@@ -184,7 +185,8 @@ ccl_device bool kernel_branched_path_surface_bounce(KernelGlobals *kg, RNG *rng,
 #ifndef __SPLIT_KERNEL__
 /* path tracing: connect path directly to position on a light and add it to L */
 ccl_device_inline void kernel_path_surface_connect_light(KernelGlobals *kg, ccl_addr_space RNG *rng,
-	ShaderData *sd, float3 throughput, ccl_addr_space PathState *state, PathRadiance *L)
+	ShaderData *sd, ShaderData *emission_sd, float3 throughput, ccl_addr_space PathState *state,
+	PathRadiance *L)
 {
 #ifdef __EMISSION__
 	if(!(kernel_data.integrator.use_direct_light && (ccl_fetch(sd, flag) & SD_BSDF_HAS_EVAL)))
@@ -206,11 +208,11 @@ ccl_device_inline void kernel_path_surface_connect_light(KernelGlobals *kg, ccl_
 	LightSample ls;
 	light_sample(kg, light_t, light_u, light_v, ccl_fetch(sd, time), ccl_fetch(sd, P), state->bounce, &ls);
 
-	if(direct_emission(kg, sd, &ls, state, &light_ray, &L_light, &is_lamp)) {
+	if(direct_emission(kg, sd, emission_sd, &ls, state, &light_ray, &L_light, &is_lamp)) {
 		/* trace shadow ray */
 		float3 shadow;
 
-		if(!shadow_blocked(kg, state, &light_ray, &shadow)) {
+		if(!shadow_blocked(kg, emission_sd, state, &light_ray, &shadow)) {
 			/* accumulate */
 			path_radiance_accum_light(L, throughput, &L_light, shadow, 1.0f, state->bounce, is_lamp);
 		}
diff --git a/intern/cycles/kernel/kernel_path_volume.h b/intern/cycles/kernel/kernel_path_volume.h
index 9eb8b240b88..e45522a4641 100644
--- a/intern/cycles/kernel/kernel_path_volume.h
+++ b/intern/cycles/kernel/kernel_path_volume.h
@@ -19,7 +19,7 @@ CCL_NAMESPACE_BEGIN
 #ifdef __VOLUME_SCATTER__
 
 ccl_device void kernel_path_volume_connect_light(KernelGlobals *kg, RNG *rng,
-	ShaderData *sd, float3 throughput, PathState *state, PathRadiance *L)
+	ShaderData *sd, ShaderData *emission_sd, float3 throughput, PathState *state, PathRadiance *L)
 {
 #ifdef __EMISSION__
 	if(!kernel_data.integrator.use_direct_light)
@@ -44,11 +44,11 @@ ccl_device void kernel_path_volume_connect_light(KernelGlobals *kg, RNG *rng,
 	if(ls.pdf == 0.0f)
 		return;
 	
-	if(direct_emission(kg, sd, &ls, state, &light_ray, &L_light, &is_lamp)) {
+	if(direct_emission(kg, sd, emission_sd, &ls, state, &light_ray, &L_light, &is_lamp)) {
 		/* trace shadow ray */
 		float3 shadow;
 
-		if(!shadow_blocked(kg, state, &light_ray, &shadow)) {
+		if(!shadow_blocked(kg, emission_sd, state, &light_ray, &shadow)) {
 			/* accumulate */
 			path_radiance_accum_light(L, throughput, &L_light, shadow, 1.0f, state->bounce, is_lamp);
 		}
@@ -106,7 +106,7 @@ bool kernel_path_volume_bounce(KernelGlobals *kg, RNG *rng,
 }
 
 ccl_device void kernel_branched_path_volume_connect_light(KernelGlobals *kg, RNG *rng,
-	ShaderData *sd, float3 throughput, PathState *state, PathRadiance *L,
+	ShaderData *sd, ShaderData *emission_sd, float3 throughput, PathState *state, PathRadiance *L,
 	bool sample_all_lights, Ray *ray, const VolumeSegment *segment)
 {
 #ifdef __EMISSION__
@@ -160,11 +160,11 @@ ccl_device void kernel_branched_path_volume_connect_light(KernelGlobals *kg, RNG
 				if(ls.pdf == 0.0f)
 					continue;
 
-				if(direct_emission(kg, sd, &ls, state, &light_ray, &L_light, &is_lamp)) {
+				if(direct_emission(kg, sd, emission_sd, &ls, state, &light_ray, &L_light, &is_lamp)) {
 					/* trace shadow ray */
 					float3 shadow;
 
-					if(!shadow_blocked(kg, state, &light_ray, &shadow)) {
+					if(!shadow_blocked(kg, emission_sd, state, &light_ray, &shadow)) {
 						/* accumulate */
 						path_radiance_accum_light(L, tp*num_samples_inv, &L_light, shadow, num_samples_inv, state->bounce, is_lamp);
 					}
@@ -211,11 +211,11 @@ ccl_device void kernel_branched_path_volume_connect_light(KernelGlobals *kg, RNG
 				if(ls.pdf == 0.0f)
 					continue;
 
-				if(direct_emission(kg, sd, &ls, state, &light_ray, &L_light, &is_lamp)) {
+				if(direct_emission(kg, sd, emission_sd, &ls, state, &light_ray, &L_light, &is_lamp)) {
 					/* trace shadow ray */
 					float3 shadow;
 
-					if(!shadow_blocked(kg, state, &light_ray, &shadow)) {
+					if(!shadow_blocked(kg, emission_sd, state, &light_ray, &shadow)) {
 						/* accumulate */
 						path_radiance_accum_light(L, tp*num_samples_inv, &L_light, shadow, num_samples_inv, state->bounce, is_lamp);
 					}
@@ -251,11 +251,11 @@ ccl_device void kernel_branched_path_volume_connect_light(KernelGlobals *kg, RNG
 			return;
 
 		/* sample random light */
-		if(direct_emission(kg, sd, &ls, state, &light_ray, &L_light, &is_lamp)) {
+		if(direct_emission(kg, sd, emission_sd, &ls, state, &light_ray, &L_light, &is_lamp)) {
 			/* trace shadow ray */
 			float3 shadow;
 
-			if(!shadow_blocked(kg, state, &light_ray, &shadow)) {
+			if(!shadow_blocked(kg, emission_sd, state, &light_ray, &shadow)) {
 				/* accumulate */
 				path_radiance_accum_light(L, tp, &L_light, shadow, 1.0f, state->bounce, is_lamp);
 			}
diff --git a/intern/cycles/kernel/kernel_shadow.h b/intern/cycles/kernel/kernel_shadow.h
index 504ac2e40bc..c8f6503cf58 100644
--- a/intern/cycles/kernel/kernel_shadow.h
+++ b/intern/cycles/kernel/kernel_shadow.h
@@ -41,7 +41,7 @@ CCL_NAMESPACE_BEGIN
 
 #define STACK_MAX_HITS 64
 
-ccl_device_inline bool shadow_blocked(KernelGlobals *kg, PathState *state, Ray *ray, float3 *shadow)
+ccl_device_inline bool shadow_blocked(KernelGlobals *kg, ShaderData *shadow_sd, PathState *state, Ray *ray, float3 *shadow)
 {
 	*shadow = make_float3(1.0f, 1.0f, 1.0f);
 
@@ -107,21 +107,20 @@ ccl_device_inline bool shadow_blocked(KernelGlobals *kg, PathState *state, Ray *
 				if(ps.volume_stack[0].shader != SHADER_NONE) {
 					Ray segment_ray = *ray;
 					segment_ray.t = isect->t;
-					kernel_volume_shadow(kg, &ps, &segment_ray, &throughput);
+					kernel_volume_shadow(kg, shadow_sd, &ps, &segment_ray, &throughput);
 				}
 #endif
 
 				/* setup shader data at surface */
-				ShaderData sd;
-				shader_setup_from_ray(kg, &sd, isect, ray);
+				shader_setup_from_ray(kg, shadow_sd, isect, ray);
 
 				/* attenuation from transparent surface */
-				if(!(sd.flag & SD_HAS_ONLY_VOLUME)) {
+				if(!(shadow_sd->flag & SD_HAS_ONLY_VOLUME)) {
 					path_state_modify_bounce(state, true);
-					shader_eval_surface(kg, &sd, state, 0.0f, PATH_RAY_SHADOW, SHADER_CONTEXT_SHADOW);
+					shader_eval_surface(kg, shadow_sd, state, 0.0f, PATH_RAY_SHADOW, SHADER_CONTEXT_SHADOW);
 					path_state_modify_bounce(state, false);
 
-					throughput *= shader_bsdf_transparency(kg, &sd);
+					throughput *= shader_bsdf_transparency(kg, shadow_sd);
 				}
 
 				/* stop if all light is blocked */
@@ -133,13 +132,13 @@ ccl_device_inline bool shadow_blocked(KernelGlobals *kg, PathState *state, Ray *
 				}
 
 				/* move ray forward */
-				ray->P = sd.P;
+				ray->P = shadow_sd->P;
 				if(ray->t != FLT_MAX)
 					ray->D = normalize_len(Pend - ray->P, &ray->t);
 
 #ifdef __VOLUME__
 				/* exit/enter volume */
-				kernel_volume_stack_enter_exit(kg, &sd, ps.volume_stack);
+				kernel_volume_stack_enter_exit(kg, shadow_sd, ps.volume_stack);
 #endif
 
 				bounce++;
@@ -148,7 +147,7 @@ ccl_device_inline bool shadow_blocked(KernelGlobals *kg, PathState *state, Ray *
 #ifdef __VOLUME__
 			/* attenuation for last line segment towards light */
 			if(ps.volume_stack[0].shader != SHADER_NONE)
-				kernel_volume_shadow(kg, &ps, ray, &throughput);
+				kernel_volume_shadow(kg, shadow_sd, &ps, ray, &throughput);
 #endif
 
 			*shadow = throughput;
@@ -164,7 +163,7 @@ ccl_device_inline bool shadow_blocked(KernelGlobals *kg, PathState *state, Ray *
 #ifdef __VOLUME__
 	if(!blocked && state->volume_stack[0].shader != SHADER_NONE) {
 		/* apply attenuation from current volume shader */
-		kernel_volume_shadow(kg, state, ray, shadow);
+		kernel_volume_shadow(kg, shadow_sd, state, ray, shadow);
 	}
 #endif
 
@@ -184,6 +183,7 @@ ccl_device_inline bool shadow_blocked(KernelGlobals *kg, PathState *state, Ray *
  * one extra ray cast for the cases were we do want transparency. */
 
 ccl_device_noinline bool shadow_blocked(KernelGlobals *kg,
+                                        ShaderData *shadow_sd,
                                         ccl_addr_space PathState *state,
                                         ccl_addr_space Ray *ray_input,
                                         float3 *shadow)
@@ -228,7 +228,7 @@ ccl_device_noinline bool shadow_blocked(KernelGlobals *kg,
 #ifdef __VOLUME__
 					/* attenuation for last line segment towards light */
 					if(ps.volume_stack[0].shader != SHADER_NONE)
-						kernel_volume_shadow(kg, &ps, ray, &throughput);
+						kernel_volume_shadow(kg, shadow_sd, &ps, ray, &throughput);
 #endif
 
 					*shadow *= throughput;
@@ -244,39 +244,33 @@ ccl_device_noinline bool shadow_blocked(KernelGlobals *kg,
 				if(ps.volume_stack[0].shader != SHADER_NONE) {
 					Ray segment_ray = *ray;
 					segment_ray.t = isect->t;
-					kernel_volume_shadow(kg, &ps, &segment_ray, &throughput);
+					kernel_volume_shadow(kg, shadow_sd, &ps, &segment_ray, &throughput);
 				}
 #endif
 
 				/* setup shader data at surface */
-#ifdef __SPLIT_KERNEL__
-				ShaderData *sd = kg->sd_input;
-#else
-				ShaderData sd_object;
-				ShaderData *sd = &sd_object;
-#endif
-				shader_setup_from_ray(kg, sd, isect, ray);
+				shader_setup_from_ray(kg, shadow_sd, isect, ray);
 
 				/* attenuation from transparent surface */
-				if(!(ccl_fetch(sd, flag) & SD_HAS_ONLY_VOLUME)) {
+				if(!(ccl_fetch(shadow_sd, flag) & SD_HAS_ONLY_VOLUME)) {
 					path_state_modify_bounce(state, true);
-					shader_eval_surface(kg, sd, state, 0.0f, PATH_RAY_SHADOW, SHADER_CONTEXT_SHADOW);
+					shader_eval_surface(kg, shadow_sd, state, 0.0f, PATH_RAY_SHADOW, SHADER_CONTEXT_SHADOW);
 					path_state_modify_bounce(state, false);
 
-					throughput *= shader_bsdf_transparency(kg, sd);
+					throughput *= shader_bsdf_transparency(kg, shadow_sd);
 				}
 
 				if(is_zero(throughput))
 					return true;
 
 				/* move ray forward */
-				ray->P = ray_offset(ccl_fetch(sd, P), -ccl_fetch(sd, Ng));
+				ray->P = ray_offset(ccl_fetch(shadow_sd, P), -ccl_fetch(shadow_sd, Ng));
 				if(ray->t != FLT_MAX)
 					ray->D = normalize_len(Pend - ray->P, &ray->t);
 
 #ifdef __VOLUME__
 				/* exit/enter volume */
-				kernel_volume_stack_enter_exit(kg, sd, ps.volume_stack);
+				kernel_volume_stack_enter_exit(kg, shadow_sd, ps.volume_stack);
 #endif
 
 				bounce++;
@@ -286,7 +280,7 @@ ccl_device_noinline bool shadow_blocked(KernelGlobals *kg,
 #ifdef __VOLUME__
 	else if(!blocked && state->volume_stack[0].shader != SHADER_NONE) {
 		/* apply attenuation from current volume shader */
-		kernel_volume_shadow(kg, state, ray, shadow);
+		kernel_volume_shadow(kg, shadow_sd, state, ray, shadow);
 	}
 #endif
 #endif
diff --git a/intern/cycles/kernel/kernel_volume.h b/intern/cycles/kernel/kernel_volume.h
index 30a978f6c9e..e1ea60f372e 100644
--- a/intern/cycles/kernel/kernel_volume.h
+++ b/intern/cycles/kernel/kernel_volume.h
@@ -219,15 +219,14 @@ ccl_device void kernel_volume_shadow_heterogeneous(KernelGlobals *kg, PathState
 
 /* get the volume attenuation over line segment defined by ray, with the
  * assumption that there are no surfaces blocking light between the endpoints */
-ccl_device_noinline void kernel_volume_shadow(KernelGlobals *kg, PathState *state, Ray *ray, float3 *throughput)
+ccl_device_noinline void kernel_volume_shadow(KernelGlobals *kg, ShaderData *shadow_sd, PathState *state, Ray *ray, float3 *throughput)
 {
-	ShaderData sd;
-	shader_setup_from_volume(kg, &sd, ray);
+	shader_setup_from_volume(kg, shadow_sd, ray);
 
 	if(volume_stack_is_heterogeneous(kg, state->volume_stack))
-		kernel_volume_shadow_heterogeneous(kg, state, ray, &sd, throughput);
+		kernel_volume_shadow_heterogeneous(kg, state, ray, shadow_sd, throughput);
 	else
-		kernel_volume_shadow_homogeneous(kg, state, ray, &sd, throughput);
+		kernel_volume_shadow_homogeneous(kg, state, ray, shadow_sd, throughput);
 }
 
 /* Equi-angular sampling as in:
@@ -1000,6 +999,7 @@ ccl_device bool kernel_volume_use_decoupled(KernelGlobals *kg, bool heterogeneou
  * is inside of. */
 
 ccl_device void kernel_volume_stack_init(KernelGlobals *kg,
+                                         ShaderData *stack_sd,
                                          Ray *ray,
                                          VolumeStack *stack)
 {
@@ -1040,28 +1040,27 @@ ccl_device void kernel_volume_stack_init(KernelGlobals *kg,
 		qsort(hits, num_hits, sizeof(Intersection), intersections_compare);
 
 		for(uint hit = 0; hit < num_hits; ++hit, ++isect) {
-			ShaderData sd;
-			shader_setup_from_ray(kg, &sd, isect, &volume_ray);
-			if(sd.flag & SD_BACKFACING) {
+			shader_setup_from_ray(kg, stack_sd, isect, &volume_ray);
+			if(stack_sd->flag & SD_BACKFACING) {
 				bool need_add = true;
 				for(int i = 0; i < enclosed_index && need_add; ++i) {
 					/* If ray exited the volume and never entered to that volume
 					 * it means that camera is inside such a volume.
 					 */
-					if(enclosed_volumes[i] == sd.object) {
+					if(enclosed_volumes[i] == stack_sd->object) {
 						need_add = false;
 					}
 				}
 				for(int i = 0; i < stack_index && need_add; ++i) {
 					/* Don't add intersections twice. */
-					if(stack[i].object == sd.object) {
+					if(stack[i].object == stack_sd->object) {
 						need_add = false;
 						break;
 					}
 				}
 				if(need_add) {
-					stack[stack_index].object = sd.object;
-					stack[stack_index].shader = sd.shader;
+					stack[stack_index].object = stack_sd->object;
+					stack[stack_index].shader = stack_sd->shader;
 					++stack_index;
 				}
 			}
@@ -1069,7 +1068,7 @@ ccl_device void kernel_volume_stack_init(KernelGlobals *kg,
 				/* If ray from camera enters the volume, this volume shouldn't
 				 * be added to the stack on exit.
 				 */
-				enclosed_volumes[enclosed_index++] = sd.object;
+				enclosed_volumes[enclosed_index++] = stack_sd->object;
 			}
 		}
 	}
@@ -1086,9 +1085,8 @@ ccl_device void kernel_volume_stack_init(KernelGlobals *kg,
 			break;
 		}
 
-		ShaderData sd;
-		shader_setup_from_ray(kg, &sd, &isect, &volume_ray);
-		if(sd.flag & SD_BACKFACING) {
+		shader_setup_from_ray(kg, stack_sd, &isect, &volume_ray);
+		if(stack_sd->flag & SD_BACKFACING) {
 			/* If ray exited the volume and never entered to that volume
 			 * it means that camera is inside such a volume.
 			 */
@@ -1097,20 +1095,20 @@ ccl_device void kernel_volume_stack_init(KernelGlobals *kg,
 				/* If ray exited the volume and never entered to that volume
 				 * it means that camera is inside such a volume.
 				 */
-				if(enclosed_volumes[i] == sd.object) {
+				if(enclosed_volumes[i] == stack_sd->object) {
 					need_add = false;
 				}
 			}
 			for(int i = 0; i < stack_index && need_add; ++i) {
 				/* Don't add intersections twice. */
-				if(stack[i].object == sd.object) {
+				if(stack[i].object == stack_sd->object) {
 					need_add = false;
 					break;
 				}
 			}
 			if(need_add) {
-				stack[stack_index].object = sd.object;
-				stack[stack_index].shader = sd.shader;
+				stack[stack_index].object = stack_sd->object;
+				stack[stack_index].shader = stack_sd->shader;
 				++stack_index;
 			}
 		}
@@ -1118,11 +1116,11 @@ ccl_device void kernel_volume_stack_init(KernelGlobals *kg,
 			/* If ray from camera enters the volume, this volume shouldn't
 			 * be added to the stack on exit.
 			 */
-			enclosed_volumes[enclosed_index++] = sd.object;
+			enclosed_volumes[enclosed_index++] = stack_sd->object;
 		}
 
 		/* Move ray forward. */
-		volume_ray.P = ray_offset(sd.P, -sd.Ng);
+		volume_ray.P = ray_offset(stack_sd->P, -stack_sd->Ng);
 		++step;
 	}
 #endif
@@ -1190,6 +1188,7 @@ ccl_device void kernel_volume_stack_enter_exit(KernelGlobals *kg, ShaderData *sd
 
 #ifdef __SUBSURFACE__
 ccl_device void kernel_volume_stack_update_for_subsurface(KernelGlobals *kg,
+                                                          ShaderData *stack_sd,
                                                           Ray *ray,
                                                           VolumeStack *stack)
 {
@@ -1210,9 +1209,8 @@ ccl_device void kernel_volume_stack_update_for_subsurface(KernelGlobals *kg,
 		qsort(hits, num_hits, sizeof(Intersection), intersections_compare);
 
 		for(uint hit = 0; hit < num_hits; ++hit, ++isect) {
-			ShaderData sd;
-			shader_setup_from_ray(kg, &sd, isect, &volume_ray);
-			kernel_volume_stack_enter_exit(kg, &sd, stack);
+			shader_setup_from_ray(kg, stack_sd, isect, &volume_ray);
+			kernel_volume_stack_enter_exit(kg, stack_sd, stack);
 		}
 	}
 #  else
@@ -1224,13 +1222,12 @@ ccl_device void kernel_volume_stack_update_for_subsurface(KernelGlobals *kg,
 	                             &isect,
 	                             PATH_RAY_ALL_VISIBILITY))
 	{
-		ShaderData sd;
-		shader_setup_from_ray(kg, &sd, &isect, &volume_ray);
-		kernel_volume_stack_enter_exit(kg, &sd, stack);
+		shader_setup_from_ray(kg, stack_sd, &isect, &volume_ray);
+		kernel_volume_stack_enter_exit(kg, stack_sd, stack);
 
 		/* Move ray forward. */
-		volume_ray.P = ray_offset(sd.P, -sd.Ng);
-		volume_ray.t -= sd.ray_length;
+		volume_ray.P = ray_offset(stack_sd->P, -stack_sd->Ng);
+		volume_ray.t -= stack_sd->ray_length;
 		++step;
 	}
 #  endif
diff --git a/intern/cycles/kernel/split/kernel_background_buffer_update.h b/intern/cycles/kernel/split/kernel_background_buffer_update.h
index 3d12a3dd993..f42d0a985bb 100644
--- a/intern/cycles/kernel/split/kernel_background_buffer_update.h
+++ b/intern/cycles/kernel/split/kernel_background_buffer_update.h
@@ -157,7 +157,7 @@ ccl_device char kernel_background_buffer_update(
 		if(IS_STATE(ray_state, ray_index, RAY_HIT_BACKGROUND)) {
 #ifdef __BACKGROUND__
 			/* sample background shader */
-			float3 L_background = indirect_background(kg, state, ray);
+			float3 L_background = indirect_background(kg, kg->sd_input, state, ray);
 			path_radiance_accum_background(L, (*throughput), L_background, state->bounce);
 #endif
 			ASSIGN_RAY_STATE(ray_state, ray_index, RAY_UPDATE_BUFFER);
@@ -226,7 +226,7 @@ ccl_device char kernel_background_buffer_update(
 				*throughput = make_float3(1.0f, 1.0f, 1.0f);
 				*L_transparent = 0.0f;
 				path_radiance_init(L, kernel_data.film.use_light_pass);
-				path_state_init(kg, state, rng, sample, ray);
+				path_state_init(kg, kg->sd_input, state, rng, sample, ray);
 #ifdef __KERNEL_DEBUG__
 				debug_data_init(debug_data);
 #endif
diff --git a/intern/cycles/kernel/split/kernel_data_init.h b/intern/cycles/kernel/split/kernel_data_init.h
index 9891391a3a3..e3dbc43757e 100644
--- a/intern/cycles/kernel/split/kernel_data_init.h
+++ b/intern/cycles/kernel/split/kernel_data_init.h
@@ -207,6 +207,7 @@ ccl_device void kernel_data_init(
 			L_transparent_coop[ray_index] = 0.0f;
 			path_radiance_init(&PathRadiance_coop[ray_index], kernel_data.film.use_light_pass);
 			path_state_init(kg,
+			                kg->sd_input,
 			                &PathState_coop[ray_index],
 			                &rng_coop[ray_index],
 			                my_sample,
diff --git a/intern/cycles/kernel/split/kernel_direct_lighting.h b/intern/cycles/kernel/split/kernel_direct_lighting.h
index c7a2aa6426c..ebe91097496 100644
--- a/intern/cycles/kernel/split/kernel_direct_lighting.h
+++ b/intern/cycles/kernel/split/kernel_direct_lighting.h
@@ -88,7 +88,7 @@ ccl_device char kernel_direct_lighting(
 
 			BsdfEval L_light;
 			bool is_lamp;
-			if(direct_emission(kg, sd, &ls, state, &light_ray, &L_light, &is_lamp)) {
+			if(direct_emission(kg, sd, kg->sd_input, &ls, state, &light_ray, &L_light, &is_lamp)) {
 				/* Write intermediate data to global memory to access from
 				 * the next kernel.
 				 */
diff --git a/intern/cycles/kernel/split/kernel_lamp_emission.h b/intern/cycles/kernel/split/kernel_lamp_emission.h
index dc3b4b34d4e..3bd0e361078 100644
--- a/intern/cycles/kernel/split/kernel_lamp_emission.h
+++ b/intern/cycles/kernel/split/kernel_lamp_emission.h
@@ -74,7 +74,7 @@ ccl_device void kernel_lamp_emission(
 			/* intersect with lamp */
 			float3 emission;
 
-			if(indirect_lamp_emission(kg, state, &light_ray, &emission)) {
+			if(indirect_lamp_emission(kg, kg->sd_input, state, &light_ray, &emission)) {
 				path_radiance_accum_emission(L, throughput, emission, state->bounce);
 			}
 		}
diff --git a/intern/cycles/kernel/split/kernel_shadow_blocked.h b/intern/cycles/kernel/split/kernel_shadow_blocked.h
index 0c989861eef..6153af47f96 100644
--- a/intern/cycles/kernel/split/kernel_shadow_blocked.h
+++ b/intern/cycles/kernel/split/kernel_shadow_blocked.h
@@ -71,6 +71,7 @@ ccl_device void kernel_shadow_blocked(
 
 		float3 shadow;
 		update_path_radiance = !(shadow_blocked(kg,
+		                                        kg->sd_input,
 		                                        state,
 		                                        light_ray_global,
 		                                        &shadow));