Cycles: option to make background visible through glass transparent.
[blender-staging.git] / intern / cycles / kernel / kernel_path.h
index b74822bc0957faa41e9deb3b7361180ce149db7b..1e98bca66ad5ae6a48ee6c82c3d977d5020d9732 100644 (file)
  * distributed under the License is distributed on an "AS IS" BASIS,
  * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  * See the License for the specific language governing permissions and
- * limitations under the License
+ * limitations under the License.
  */
 
 #ifdef __OSL__
-#include "osl_shader.h"
+#  include "kernel/osl/osl_shader.h"
 #endif
 
-#include "kernel_random.h"
-#include "kernel_projection.h"
-#include "kernel_montecarlo.h"
-#include "kernel_differential.h"
-#include "kernel_camera.h"
+#include "kernel/kernel_random.h"
+#include "kernel/kernel_projection.h"
+#include "kernel/kernel_montecarlo.h"
+#include "kernel/kernel_differential.h"
+#include "kernel/kernel_camera.h"
 
-#include "geom/geom.h"
+#include "kernel/geom/geom.h"
+#include "kernel/bvh/bvh.h"
 
-#include "kernel_accumulate.h"
-#include "kernel_shader.h"
-#include "kernel_light.h"
-#include "kernel_passes.h"
+#include "kernel/kernel_accumulate.h"
+#include "kernel/kernel_shader.h"
+#include "kernel/kernel_light.h"
+#include "kernel/kernel_passes.h"
 
 #ifdef __SUBSURFACE__
-#include "kernel_subsurface.h"
+#  include "kernel/kernel_subsurface.h"
 #endif
 
 #ifdef __VOLUME__
-#include "kernel_volume.h"
+#  include "kernel/kernel_volume.h"
 #endif
 
-#include "kernel_path_state.h"
-#include "kernel_shadow.h"
-#include "kernel_emission.h"
-#include "kernel_path_surface.h"
-#include "kernel_path_volume.h"
+#include "kernel/kernel_path_state.h"
+#include "kernel/kernel_shadow.h"
+#include "kernel/kernel_emission.h"
+#include "kernel/kernel_path_common.h"
+#include "kernel/kernel_path_surface.h"
+#include "kernel/kernel_path_volume.h"
+#include "kernel/kernel_path_subsurface.h"
 
 CCL_NAMESPACE_BEGIN
 
-ccl_device void kernel_path_indirect(KernelGlobals *kg, RNG *rng, Ray ray,
-       float3 throughput, int num_samples, PathState state, PathRadiance *L)
+ccl_device_forceinline bool kernel_path_scene_intersect(
+       KernelGlobals *kg,
+       ccl_addr_space PathState *state,
+       Ray *ray,
+       Intersection *isect,
+       PathRadiance *L)
 {
-       /* path iteration */
-       for(;;) {
-               /* intersect scene */
-               Intersection isect;
-               uint visibility = path_state_ray_visibility(kg, &state);
+       uint visibility = path_state_ray_visibility(kg, state);
+
+       if(path_state_ao_bounce(kg, state)) {
+               visibility = PATH_RAY_SHADOW;
+               ray->t = kernel_data.background.ao_distance;
+       }
+
 #ifdef __HAIR__
-               bool hit = scene_intersect(kg, &ray, visibility, &isect, NULL, 0.0f, 0.0f);
+       float difl = 0.0f, extmax = 0.0f;
+       uint lcg_state = 0;
+
+       if(kernel_data.bvh.have_curves) {
+               if((kernel_data.cam.resolution == 1) && (state->flag & PATH_RAY_CAMERA)) {
+                       float3 pixdiff = ray->dD.dx + ray->dD.dy;
+                       /*pixdiff = pixdiff - dot(pixdiff, ray.D)*ray.D;*/
+                       difl = kernel_data.curve.minimum_width * len(pixdiff) * 0.5f;
+               }
+
+               extmax = kernel_data.curve.maximum_width;
+               lcg_state = lcg_state_init_addrspace(state, 0x51633e2d);
+       }
+
+       bool hit = scene_intersect(kg, *ray, visibility, isect, &lcg_state, difl, extmax);
 #else
-               bool hit = scene_intersect(kg, &ray, visibility, &isect);
-#endif
+       bool hit = scene_intersect(kg, *ray, visibility, isect, NULL, 0.0f, 0.0f);
+#endif  /* __HAIR__ */
+
+#ifdef __KERNEL_DEBUG__
+       if(state->flag & PATH_RAY_CAMERA) {
+               L->debug_data.num_bvh_traversed_nodes += isect->num_traversed_nodes;
+               L->debug_data.num_bvh_traversed_instances += isect->num_traversed_instances;
+               L->debug_data.num_bvh_intersections += isect->num_intersections;
+       }
+       L->debug_data.num_ray_bounces++;
+#endif  /* __KERNEL_DEBUG__ */
 
+       return hit;
+}
+
+ccl_device_forceinline void kernel_path_lamp_emission(
+       KernelGlobals *kg,
+       ccl_addr_space PathState *state,
+       Ray *ray,
+       float3 throughput,
+       ccl_addr_space Intersection *isect,
+       ShaderData *emission_sd,
+       PathRadiance *L)
+{
 #ifdef __LAMP_MIS__
-               if(kernel_data.integrator.use_lamp_mis && !(state.flag & PATH_RAY_CAMERA)) {
-                       /* ray starting from previous non-transparent bounce */
-                       Ray light_ray;
-
-                       light_ray.P = ray.P - state.ray_t*ray.D;
-                       state.ray_t += isect.t;
-                       light_ray.D = ray.D;
-                       light_ray.t = state.ray_t;
-                       light_ray.time = ray.time;
-                       light_ray.dD = ray.dD;
-                       light_ray.dP = ray.dP;
-
-                       /* intersect with lamp */
-                       float3 emission;
-
-                       if(indirect_lamp_emission(kg, &state, &light_ray, &emission))
-                               path_radiance_accum_emission(L, throughput, emission, state.bounce);
-               }
-#endif
+       if(kernel_data.integrator.use_lamp_mis && !(state->flag & PATH_RAY_CAMERA)) {
+               /* ray starting from previous non-transparent bounce */
+               Ray light_ray;
 
-#ifdef __VOLUME__
-               /* volume attenuation, emission, scatter */
-               if(state.volume_stack[0].shader != SHADER_NONE) {
-                       Ray volume_ray = ray;
-                       volume_ray.t = (hit)? isect.t: FLT_MAX;
-
-                       bool heterogeneous = volume_stack_is_heterogeneous(kg, state.volume_stack);
-
-#ifdef __VOLUME_DECOUPLED__
-                       int sampling_method = volume_stack_sampling_method(kg, state.volume_stack);
-                       bool decoupled = kernel_volume_use_decoupled(kg, heterogeneous, false, sampling_method);
-
-                       if(decoupled) {
-                               /* cache steps along volume for repeated sampling */
-                               VolumeSegment volume_segment;
-                               ShaderData volume_sd;
-
-                               shader_setup_from_volume(kg, &volume_sd, &volume_ray, state.bounce, state.transparent_bounce);
-                               kernel_volume_decoupled_record(kg, &state,
-                                       &volume_ray, &volume_sd, &volume_segment, heterogeneous);
-                               
-                               volume_segment.sampling_method = sampling_method;
-
-                               /* emission */
-                               if(volume_segment.closure_flag & SD_EMISSION)
-                                       path_radiance_accum_emission(L, throughput, volume_segment.accum_emission, state.bounce);
-
-                               /* scattering */
-                               VolumeIntegrateResult result = VOLUME_PATH_ATTENUATED;
-
-                               if(volume_segment.closure_flag & SD_SCATTER) {
-                                       bool all = kernel_data.integrator.sample_all_lights_indirect;
-
-                                       /* direct light sampling */
-                                       kernel_branched_path_volume_connect_light(kg, rng, &volume_sd,
-                                               throughput, &state, L, 1.0f, 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
-                                        * this computation, but makes code a bit complex */
-                                       float rphase = path_state_rng_1D_for_decision(kg, rng, &state, PRNG_PHASE);
-                                       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,
-                                               rphase, rscatter, &volume_segment, NULL, true);
-                               }
-
-                               if(result != VOLUME_PATH_SCATTERED)
-                                       throughput *= volume_segment.accum_transmittance;
-
-                               /* free cached steps */
-                               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))
-                                               continue;
-                                       else
-                                               break;
-                               }
-                       }
-                       else
-#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);
-
-#ifdef __VOLUME_SCATTER__
-                               if(result == VOLUME_PATH_SCATTERED) {
-                                       /* direct lighting */
-                                       kernel_path_volume_connect_light(kg, rng, &volume_sd, throughput, &state, L);
-
-                                       /* indirect light bounce */
-                                       if(kernel_path_volume_bounce(kg, rng, &volume_sd, &throughput, &state, L, &ray))
-                                               continue;
-                                       else
-                                               break;
-                               }
-#endif
-                       }
-               }
-#endif
+               light_ray.P = ray->P - state->ray_t*ray->D;
+               state->ray_t += isect->t;
+               light_ray.D = ray->D;
+               light_ray.t = state->ray_t;
+               light_ray.time = ray->time;
+               light_ray.dD = ray->dD;
+               light_ray.dP = ray->dP;
+
+               /* intersect with lamp */
+               float3 emission;
+
+               if(indirect_lamp_emission(kg, emission_sd, state, &light_ray, &emission))
+                       path_radiance_accum_emission(L, state, throughput, emission);
+       }
+#endif  /* __LAMP_MIS__ */
+}
+
+ccl_device_forceinline void kernel_path_background(
+       KernelGlobals *kg,
+       ccl_addr_space PathState *state,
+       ccl_addr_space Ray *ray,
+       float3 throughput,
+       ShaderData *sd,
+       PathRadiance *L)
+{
+       /* eval background shader if nothing hit */
+       if(kernel_data.background.transparent && (state->flag & PATH_RAY_TRANSPARENT_BACKGROUND)) {
+               L->transparent += average(throughput);
+
+#ifdef __PASSES__
+               if(!(kernel_data.film.light_pass_flag & PASSMASK(BACKGROUND)))
+#endif  /* __PASSES__ */
+                       return;
+       }
+
+       /* When using the ao bounces approximation, adjust background
+        * shader intensity with ao factor. */
+       if(path_state_ao_bounce(kg, state)) {
+               throughput *= kernel_data.background.ao_bounces_factor;
+       }
 
-               if(!hit) {
 #ifdef __BACKGROUND__
-                       /* sample background shader */
-                       float3 L_background = indirect_background(kg, &state, &ray);
-                       path_radiance_accum_background(L, throughput, L_background, state.bounce);
-#endif
+       /* sample background shader */
+       float3 L_background = indirect_background(kg, sd, state, ray);
+       path_radiance_accum_background(L, state, throughput, L_background);
+#endif  /* __BACKGROUND__ */
+}
 
-                       break;
-               }
+#ifndef __SPLIT_KERNEL__
 
-               /* setup shading */
-               ShaderData sd;
-               shader_setup_from_ray(kg, &sd, &isect, &ray, state.bounce, state.transparent_bounce);
-               float rbsdf = path_state_rng_1D_for_decision(kg, rng, &state, PRNG_BSDF);
-               shader_eval_surface(kg, &sd, rbsdf, state.flag, SHADER_CONTEXT_INDIRECT);
-#ifdef __BRANCHED_PATH__
-               shader_merge_closures(&sd);
-#endif
+#ifdef __VOLUME__
+ccl_device_forceinline VolumeIntegrateResult kernel_path_volume(
+       KernelGlobals *kg,
+       ShaderData *sd,
+       PathState *state,
+       Ray *ray,
+       float3 *throughput,
+       ccl_addr_space Intersection *isect,
+       bool hit,
+       ShaderData *emission_sd,
+       PathRadiance *L)
+{
+       /* Sanitize volume stack. */
+       if(!hit) {
+               kernel_volume_clean_stack(kg, state->volume_stack);
+       }
 
-               /* blurring of bsdf after bounces, for rays that have a small likelihood
-                * of following this particular path (diffuse, rough glossy) */
-               if(kernel_data.integrator.filter_glossy != FLT_MAX) {
-                       float blur_pdf = kernel_data.integrator.filter_glossy*state.min_ray_pdf;
+       if(state->volume_stack[0].shader == SHADER_NONE) {
+               return VOLUME_PATH_ATTENUATED;
+       }
 
-                       if(blur_pdf < 1.0f) {
-                               float blur_roughness = sqrtf(1.0f - blur_pdf)*0.5f;
-                               shader_bsdf_blur(kg, &sd, blur_roughness);
-                       }
-               }
+       /* volume attenuation, emission, scatter */
+       Ray volume_ray = *ray;
+       volume_ray.t = (hit)? isect->t: FLT_MAX;
+
+       bool heterogeneous = volume_stack_is_heterogeneous(kg, state->volume_stack);
+
+#  ifdef __VOLUME_DECOUPLED__
+       int sampling_method = volume_stack_sampling_method(kg, state->volume_stack);
+       bool direct = (state->flag & PATH_RAY_CAMERA) != 0;
+       bool decoupled = kernel_volume_use_decoupled(kg, heterogeneous, direct, sampling_method);
+
+       if(decoupled) {
+               /* cache steps along volume for repeated sampling */
+               VolumeSegment volume_segment;
+
+               shader_setup_from_volume(kg, sd, &volume_ray);
+               kernel_volume_decoupled_record(kg, state,
+                       &volume_ray, sd, &volume_segment, heterogeneous);
+
+               volume_segment.sampling_method = sampling_method;
 
-#ifdef __EMISSION__
                /* emission */
-               if(sd.flag & SD_EMISSION) {
-                       float3 emission = indirect_primitive_emission(kg, &sd, isect.t, state.flag, state.ray_pdf);
-                       path_radiance_accum_emission(L, throughput, emission, state.bounce);
-               }
-#endif
+               if(volume_segment.closure_flag & SD_EMISSION)
+                       path_radiance_accum_emission(L, state, *throughput, volume_segment.accum_emission);
 
-               /* path termination. this is a strange place to put the termination, it's
-                * mainly due to the mixed in MIS that we use. gives too many unneeded
-                * shader evaluations, only need emission if we are going to terminate */
-               float probability = path_state_terminate_probability(kg, &state, throughput*num_samples);
+               /* scattering */
+               VolumeIntegrateResult result = VOLUME_PATH_ATTENUATED;
 
-               if(probability == 0.0f) {
-                       break;
-               }
-               else if(probability != 1.0f) {
-                       float terminate = path_state_rng_1D_for_decision(kg, rng, &state, PRNG_TERMINATE);
+               if(volume_segment.closure_flag & SD_SCATTER) {
+                       int all = kernel_data.integrator.sample_all_lights_indirect;
 
-                       if(terminate >= probability)
-                               break;
+                       /* direct light sampling */
+                       kernel_branched_path_volume_connect_light(kg, sd,
+                               emission_sd, *throughput, state, L, all,
+                               &volume_ray, &volume_segment);
 
-                       throughput /= probability;
+                       /* indirect sample. if we use distance sampling and take just
+                        * one sample for direct and indirect light, we could share
+                        * this computation, but makes code a bit complex */
+                       float rphase = path_state_rng_1D(kg, state, PRNG_PHASE_CHANNEL);
+                       float rscatter = path_state_rng_1D(kg, state, PRNG_SCATTER_DISTANCE);
+
+                       result = kernel_volume_decoupled_scatter(kg,
+                               state, &volume_ray, sd, throughput,
+                               rphase, rscatter, &volume_segment, NULL, true);
                }
 
-#ifdef __AO__
-               /* ambient occlusion */
-               if(kernel_data.integrator.use_ambient_occlusion || (sd.flag & SD_AO)) {
-                       float bsdf_u, bsdf_v;
-                       path_state_rng_2D(kg, rng, &state, PRNG_BSDF_U, &bsdf_u, &bsdf_v);
-
-                       float ao_factor = kernel_data.background.ao_factor;
-                       float3 ao_N;
-                       float3 ao_bsdf = shader_bsdf_ao(kg, &sd, ao_factor, &ao_N);
-                       float3 ao_D;
-                       float ao_pdf;
-                       float3 ao_alpha = make_float3(0.0f, 0.0f, 0.0f);
-
-                       sample_cos_hemisphere(ao_N, bsdf_u, bsdf_v, &ao_D, &ao_pdf);
-
-                       if(dot(sd.Ng, ao_D) > 0.0f && ao_pdf != 0.0f) {
-                               Ray light_ray;
-                               float3 ao_shadow;
-
-                               light_ray.P = ray_offset(sd.P, sd.Ng);
-                               light_ray.D = ao_D;
-                               light_ray.t = kernel_data.background.ao_distance;
-#ifdef __OBJECT_MOTION__
-                               light_ray.time = sd.time;
-#endif
-                               light_ray.dP = sd.dP;
-                               light_ray.dD = differential3_zero();
+               /* free cached steps */
+               kernel_volume_decoupled_free(kg, &volume_segment);
 
-                               if(!shadow_blocked(kg, &state, &light_ray, &ao_shadow))
-                                       path_radiance_accum_ao(L, throughput, ao_alpha, ao_bsdf, ao_shadow, state.bounce);
-                       }
+               if(result == VOLUME_PATH_SCATTERED) {
+                       if(kernel_path_volume_bounce(kg, sd, throughput, state, &L->state, ray))
+                               return VOLUME_PATH_SCATTERED;
+                       else
+                               return VOLUME_PATH_MISSED;
                }
-#endif
+               else {
+                       *throughput *= volume_segment.accum_transmittance;
+               }
+       }
+       else
+#  endif  /* __VOLUME_DECOUPLED__ */
+       {
+               /* integrate along volume segment with distance sampling */
+               VolumeIntegrateResult result = kernel_volume_integrate(
+                       kg, state, sd, &volume_ray, L, throughput, heterogeneous);
+
+#  ifdef __VOLUME_SCATTER__
+               if(result == VOLUME_PATH_SCATTERED) {
+                       /* direct lighting */
+                       kernel_path_volume_connect_light(kg, sd, emission_sd, *throughput, state, L);
+
+                       /* indirect light bounce */
+                       if(kernel_path_volume_bounce(kg, sd, throughput, state, &L->state, ray))
+                               return VOLUME_PATH_SCATTERED;
+                       else
+                               return VOLUME_PATH_MISSED;
+               }
+#  endif  /* __VOLUME_SCATTER__ */
+       }
 
-#ifdef __SUBSURFACE__
-               /* bssrdf scatter to a different location on the same object, replacing
-                * the closures with a diffuse BSDF */
-               if(sd.flag & SD_BSSRDF) {
-                       float bssrdf_probability;
-                       ShaderClosure *sc = subsurface_scatter_pick_closure(kg, &sd, &bssrdf_probability);
+       return VOLUME_PATH_ATTENUATED;
+}
+#endif  /* __VOLUME__ */
+
+#endif /* __SPLIT_KERNEL__ */
+
+ccl_device_forceinline bool kernel_path_shader_apply(
+       KernelGlobals *kg,
+       ShaderData *sd,
+       ccl_addr_space PathState *state,
+       ccl_addr_space Ray *ray,
+       float3 throughput,
+       ShaderData *emission_sd,
+       PathRadiance *L,
+       ccl_global float *buffer)
+{
+#ifdef __SHADOW_TRICKS__
+       if((sd->object_flag & SD_OBJECT_SHADOW_CATCHER)) {
+               if(state->flag & PATH_RAY_TRANSPARENT_BACKGROUND) {
+                       state->flag |= (PATH_RAY_SHADOW_CATCHER |
+                                                  PATH_RAY_STORE_SHADOW_INFO);
+
+                       float3 bg = make_float3(0.0f, 0.0f, 0.0f);
+                       if(!kernel_data.background.transparent) {
+                               bg = indirect_background(kg, emission_sd, state, ray);
+                       }
+                       path_radiance_accum_shadowcatcher(L, throughput, bg);
+               }
+       }
+       else if(state->flag & PATH_RAY_SHADOW_CATCHER) {
+               /* Only update transparency after shadow catcher bounce. */
+               L->shadow_transparency *=
+                               average(shader_bsdf_transparency(kg, sd));
+       }
+#endif  /* __SHADOW_TRICKS__ */
 
-                       /* modify throughput for picking bssrdf or bsdf */
-                       throughput *= bssrdf_probability;
+       /* holdout */
+#ifdef __HOLDOUT__
+       if(((sd->flag & SD_HOLDOUT) ||
+               (sd->object_flag & SD_OBJECT_HOLDOUT_MASK)) &&
+          (state->flag & PATH_RAY_TRANSPARENT_BACKGROUND))
+       {
+               if(kernel_data.background.transparent) {
+                       float3 holdout_weight;
+                       if(sd->object_flag & SD_OBJECT_HOLDOUT_MASK) {
+                               holdout_weight = make_float3(1.0f, 1.0f, 1.0f);
+                       }
+                       else {
+                               holdout_weight = shader_holdout_eval(kg, sd);
+                       }
+                       /* any throughput is ok, should all be identical here */
+                       L->transparent += average(holdout_weight*throughput);
+               }
 
-                       /* do bssrdf scatter step if we picked a bssrdf closure */
-                       if(sc) {
-                               uint lcg_state = lcg_state_init(rng, &state, 0x68bc21eb);
+               if(sd->object_flag & SD_OBJECT_HOLDOUT_MASK) {
+                       return false;
+               }
+       }
+#endif  /* __HOLDOUT__ */
 
-                               float bssrdf_u, bssrdf_v;
-                               path_state_rng_2D(kg, rng, &state, PRNG_BSDF_U, &bssrdf_u, &bssrdf_v);
-                               subsurface_scatter_step(kg, &sd, state.flag, sc, &lcg_state, bssrdf_u, bssrdf_v, false);
+       /* holdout mask objects do not write data passes */
+       kernel_write_data_passes(kg, buffer, L, sd, state, throughput);
 
-                               state.flag |= PATH_RAY_BSSRDF_ANCESTOR;
-                       }
-               }
-#endif
+       /* blurring of bsdf after bounces, for rays that have a small likelihood
+        * of following this particular path (diffuse, rough glossy) */
+       if(kernel_data.integrator.filter_glossy != FLT_MAX) {
+               float blur_pdf = kernel_data.integrator.filter_glossy*state->min_ray_pdf;
 
-#if defined(__EMISSION__) && defined(__BRANCHED_PATH__)
-               if(kernel_data.integrator.use_direct_light) {
-                       bool all = kernel_data.integrator.sample_all_lights_indirect;
-                       kernel_branched_path_surface_connect_light(kg, rng, &sd, &state, throughput, 1.0f, L, all);
+               if(blur_pdf < 1.0f) {
+                       float blur_roughness = sqrtf(1.0f - blur_pdf)*0.5f;
+                       shader_bsdf_blur(kg, sd, blur_roughness);
                }
-#endif
+       }
 
-               if(!kernel_path_surface_bounce(kg, rng, &sd, &throughput, &state, L, &ray))
-                       break;
+#ifdef __EMISSION__
+       /* emission */
+       if(sd->flag & SD_EMISSION) {
+               float3 emission = indirect_primitive_emission(kg, sd, sd->ray_length, state->flag, state->ray_pdf);
+               path_radiance_accum_emission(L, state, throughput, emission);
        }
+#endif  /* __EMISSION__ */
+
+       return true;
 }
 
-ccl_device void kernel_path_ao(KernelGlobals *kg, ShaderData *sd, PathRadiance *L, PathState *state, RNG *rng, float3 throughput)
+ccl_device_noinline void kernel_path_ao(KernelGlobals *kg,
+                                        ShaderData *sd,
+                                        ShaderData *emission_sd,
+                                        PathRadiance *L,
+                                        ccl_addr_space PathState *state,
+                                        float3 throughput,
+                                        float3 ao_alpha)
 {
        /* todo: solve correlation */
        float bsdf_u, bsdf_v;
 
-       path_state_rng_2D(kg, rng, state, PRNG_BSDF_U, &bsdf_u, &bsdf_v);
+       path_state_rng_2D(kg, state, PRNG_BSDF_U, &bsdf_u, &bsdf_v);
 
        float ao_factor = kernel_data.background.ao_factor;
        float3 ao_N;
        float3 ao_bsdf = shader_bsdf_ao(kg, sd, ao_factor, &ao_N);
        float3 ao_D;
        float ao_pdf;
-       float3 ao_alpha = shader_bsdf_alpha(kg, sd);
 
        sample_cos_hemisphere(ao_N, bsdf_u, bsdf_v, &ao_D, &ao_pdf);
 
@@ -314,325 +375,100 @@ ccl_device void kernel_path_ao(KernelGlobals *kg, ShaderData *sd, PathRadiance *
                light_ray.P = ray_offset(sd->P, sd->Ng);
                light_ray.D = ao_D;
                light_ray.t = kernel_data.background.ao_distance;
-#ifdef __OBJECT_MOTION__
                light_ray.time = sd->time;
-#endif
                light_ray.dP = sd->dP;
                light_ray.dD = differential3_zero();
 
-               if(!shadow_blocked(kg, state, &light_ray, &ao_shadow))
-                       path_radiance_accum_ao(L, throughput, ao_alpha, ao_bsdf, ao_shadow, state->bounce);
-       }
-}
-
-ccl_device void kernel_branched_path_ao(KernelGlobals *kg, ShaderData *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;
-       float ao_factor = kernel_data.background.ao_factor;
-       float3 ao_N;
-       float3 ao_bsdf = shader_bsdf_ao(kg, sd, ao_factor, &ao_N);
-       float3 ao_alpha = shader_bsdf_alpha(kg, sd);
-
-       for(int j = 0; j < num_samples; j++) {
-               float bsdf_u, bsdf_v;
-               path_branched_rng_2D(kg, rng, state, j, num_samples, PRNG_BSDF_U, &bsdf_u, &bsdf_v);
-
-               float3 ao_D;
-               float ao_pdf;
-
-               sample_cos_hemisphere(ao_N, bsdf_u, bsdf_v, &ao_D, &ao_pdf);
-
-               if(dot(sd->Ng, ao_D) > 0.0f && ao_pdf != 0.0f) {
-                       Ray light_ray;
-                       float3 ao_shadow;
-
-                       light_ray.P = ray_offset(sd->P, sd->Ng);
-                       light_ray.D = ao_D;
-                       light_ray.t = kernel_data.background.ao_distance;
-#ifdef __OBJECT_MOTION__
-                       light_ray.time = sd->time;
-#endif
-                       light_ray.dP = sd->dP;
-                       light_ray.dD = differential3_zero();
-
-                       if(!shadow_blocked(kg, state, &light_ray, &ao_shadow))
-                               path_radiance_accum_ao(L, throughput*num_samples_inv, ao_alpha, ao_bsdf, ao_shadow, state->bounce);
+               if(!shadow_blocked(kg, sd, emission_sd, state, &light_ray, &ao_shadow)) {
+                       path_radiance_accum_ao(L, state, throughput, ao_alpha, ao_bsdf, ao_shadow);
                }
-       }
-}
-
-#ifdef __SUBSURFACE__
-ccl_device bool kernel_path_subsurface_scatter(KernelGlobals *kg, ShaderData *sd, PathRadiance *L, PathState *state, RNG *rng, Ray *ray, float3 *throughput)
-{
-       float bssrdf_probability;
-       ShaderClosure *sc = subsurface_scatter_pick_closure(kg, sd, &bssrdf_probability);
-
-       /* modify throughput for picking bssrdf or bsdf */
-       *throughput *= bssrdf_probability;
-
-       /* do bssrdf scatter step if we picked a bssrdf closure */
-       if(sc) {
-               uint lcg_state = lcg_state_init(rng, state, 0x68bc21eb);
-
-               ShaderData bssrdf_sd[BSSRDF_MAX_HITS];
-               float bssrdf_u, bssrdf_v;
-               path_state_rng_2D(kg, rng, state, PRNG_BSDF_U, &bssrdf_u, &bssrdf_v);
-               int num_hits = subsurface_scatter_multi_step(kg, sd, bssrdf_sd, state->flag, sc, &lcg_state, bssrdf_u, bssrdf_v, false);
-
-               /* compute lighting with the BSDF closure */
-               for(int hit = 0; hit < num_hits; hit++) {
-                       float3 tp = *throughput;
-                       PathState hit_state = *state;
-                       Ray hit_ray = *ray;
-
-                       hit_state.flag |= PATH_RAY_BSSRDF_ANCESTOR;
-                       hit_state.rng_offset += PRNG_BOUNCE_NUM;
-                       
-                       kernel_path_surface_connect_light(kg, rng, &bssrdf_sd[hit], tp, state, L);
-
-                       if(kernel_path_surface_bounce(kg, rng, &bssrdf_sd[hit], &tp, &hit_state, L, &hit_ray)) {
-#ifdef __LAMP_MIS__
-                               hit_state.ray_t = 0.0f;
-#endif
-
-                               kernel_path_indirect(kg, rng, hit_ray, tp, state->num_samples, hit_state, L);
-
-                               /* for render passes, sum and reset indirect light pass variables
-                                * for the next samples */
-                               path_radiance_sum_indirect(L);
-                               path_radiance_reset_indirect(L);
-                       }
+               else {
+                       path_radiance_accum_total_ao(L, state, throughput, ao_bsdf);
                }
-               return true;
        }
-       return false;
 }
-#endif
 
-ccl_device float4 kernel_path_integrate(KernelGlobals *kg, RNG *rng, int sample, Ray ray, ccl_global float *buffer)
-{
-       /* initialize */
-       PathRadiance L;
-       float3 throughput = make_float3(1.0f, 1.0f, 1.0f);
-       float L_transparent = 0.0f;
+#ifndef __SPLIT_KERNEL__
 
-       path_radiance_init(&L, kernel_data.film.use_light_pass);
-
-       PathState state;
-       path_state_init(kg, &state, rng, sample);
+#if defined(__BRANCHED_PATH__) || defined(__BAKING__)
 
+ccl_device void kernel_path_indirect(KernelGlobals *kg,
+                                     ShaderData *sd,
+                                     ShaderData *emission_sd,
+                                     Ray *ray,
+                                     float3 throughput,
+                                     PathState *state,
+                                     PathRadiance *L)
+{
        /* path iteration */
        for(;;) {
-               /* intersect scene */
+               /* Find intersection with objects in scene. */
                Intersection isect;
-               uint visibility = path_state_ray_visibility(kg, &state);
-
-#ifdef __HAIR__
-               float difl = 0.0f, extmax = 0.0f;
-               uint lcg_state = 0;
-
-               if(kernel_data.bvh.have_curves) {
-                       if((kernel_data.cam.resolution == 1) && (state.flag & PATH_RAY_CAMERA)) {       
-                               float3 pixdiff = ray.dD.dx + ray.dD.dy;
-                               /*pixdiff = pixdiff - dot(pixdiff, ray.D)*ray.D;*/
-                               difl = kernel_data.curve.minimum_width * len(pixdiff) * 0.5f;
-                       }
-
-                       extmax = kernel_data.curve.maximum_width;
-                       lcg_state = lcg_state_init(rng, &state, 0x51633e2d);
-               }
-
-               bool hit = scene_intersect(kg, &ray, visibility, &isect, &lcg_state, difl, extmax);
-#else
-               bool hit = scene_intersect(kg, &ray, visibility, &isect);
-#endif
+               bool hit = kernel_path_scene_intersect(kg, state, ray, &isect, L);
 
-#ifdef __LAMP_MIS__
-               if(kernel_data.integrator.use_lamp_mis && !(state.flag & PATH_RAY_CAMERA)) {
-                       /* ray starting from previous non-transparent bounce */
-                       Ray light_ray;
-
-                       light_ray.P = ray.P - state.ray_t*ray.D;
-                       state.ray_t += isect.t;
-                       light_ray.D = ray.D;
-                       light_ray.t = state.ray_t;
-                       light_ray.time = ray.time;
-                       light_ray.dD = ray.dD;
-                       light_ray.dP = ray.dP;
-
-                       /* intersect with lamp */
-                       float3 emission;
-
-                       if(indirect_lamp_emission(kg, &state, &light_ray, &emission))
-                               path_radiance_accum_emission(&L, throughput, emission, state.bounce);
-               }
-#endif
+               /* Find intersection with lamps and compute emission for MIS. */
+               kernel_path_lamp_emission(kg, state, ray, throughput, &isect, sd, L);
 
 #ifdef __VOLUME__
-               /* volume attenuation, emission, scatter */
-               if(state.volume_stack[0].shader != SHADER_NONE) {
-                       Ray volume_ray = ray;
-                       volume_ray.t = (hit)? isect.t: FLT_MAX;
-
-                       bool heterogeneous = volume_stack_is_heterogeneous(kg, state.volume_stack);
-
-#ifdef __VOLUME_DECOUPLED__
-                       int sampling_method = volume_stack_sampling_method(kg, state.volume_stack);
-                       bool decoupled = kernel_volume_use_decoupled(kg, heterogeneous, true, sampling_method);
-
-                       if(decoupled) {
-                               /* cache steps along volume for repeated sampling */
-                               VolumeSegment volume_segment;
-                               ShaderData volume_sd;
-
-                               shader_setup_from_volume(kg, &volume_sd, &volume_ray, state.bounce, state.transparent_bounce);
-                               kernel_volume_decoupled_record(kg, &state,
-                                       &volume_ray, &volume_sd, &volume_segment, heterogeneous);
-
-                               volume_segment.sampling_method = sampling_method;
-
-                               /* emission */
-                               if(volume_segment.closure_flag & SD_EMISSION)
-                                       path_radiance_accum_emission(&L, throughput, volume_segment.accum_emission, state.bounce);
-
-                               /* scattering */
-                               VolumeIntegrateResult result = VOLUME_PATH_ATTENUATED;
-
-                               if(volume_segment.closure_flag & SD_SCATTER) {
-                                       bool all = false;
-
-                                       /* direct light sampling */
-                                       kernel_branched_path_volume_connect_light(kg, rng, &volume_sd,
-                                               throughput, &state, &L, 1.0f, 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
-                                        * this computation, but makes code a bit complex */
-                                       float rphase = path_state_rng_1D_for_decision(kg, rng, &state, PRNG_PHASE);
-                                       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,
-                                               rphase, rscatter, &volume_segment, NULL, true);
-                               }
-
-                               if(result != VOLUME_PATH_SCATTERED)
-                                       throughput *= volume_segment.accum_transmittance;
-
-                               /* free cached steps */
-                               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))
-                                               continue;
-                                       else
-                                               break;
-                               }
-                       }
-                       else 
-#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);
-
-#ifdef __VOLUME_SCATTER__
-                               if(result == VOLUME_PATH_SCATTERED) {
-                                       /* direct lighting */
-                                       kernel_path_volume_connect_light(kg, rng, &volume_sd, throughput, &state, &L);
-
-                                       /* indirect light bounce */
-                                       if(kernel_path_volume_bounce(kg, rng, &volume_sd, &throughput, &state, &L, &ray))
-                                               continue;
-                                       else
-                                               break;
-                               }
-#endif
-                       }
+               /* Volume integration. */
+               VolumeIntegrateResult result = kernel_path_volume(kg,
+                                                                  sd,
+                                                                  state,
+                                                                  ray,
+                                                                  &throughput,
+                                                                  &isect,
+                                                                  hit,
+                                                                  emission_sd,
+                                                                  L);
+
+               if(result == VOLUME_PATH_SCATTERED) {
+                       continue;
                }
-#endif
-
-               if(!hit) {
-                       /* eval background shader if nothing hit */
-                       if(kernel_data.background.transparent && (state.flag & PATH_RAY_CAMERA)) {
-                               L_transparent += average(throughput);
-
-#ifdef __PASSES__
-                               if(!(kernel_data.film.pass_flag & PASS_BACKGROUND))
-#endif
-                                       break;
-                       }
-
-#ifdef __BACKGROUND__
-                       /* sample background shader */
-                       float3 L_background = indirect_background(kg, &state, &ray);
-                       path_radiance_accum_background(&L, throughput, L_background, state.bounce);
-#endif
-
+               else if(result == VOLUME_PATH_MISSED) {
                        break;
                }
+#endif /* __VOLUME__*/
 
-               /* setup shading */
-               ShaderData sd;
-               shader_setup_from_ray(kg, &sd, &isect, &ray, state.bounce, state.transparent_bounce);
-               float rbsdf = path_state_rng_1D_for_decision(kg, rng, &state, PRNG_BSDF);
-               shader_eval_surface(kg, &sd, rbsdf, state.flag, SHADER_CONTEXT_MAIN);
-
-               /* holdout */
-#ifdef __HOLDOUT__
-               if((sd.flag & (SD_HOLDOUT|SD_HOLDOUT_MASK)) && (state.flag & PATH_RAY_CAMERA)) {
-                       if(kernel_data.background.transparent) {
-                               float3 holdout_weight;
-                               
-                               if(sd.flag & SD_HOLDOUT_MASK)
-                                       holdout_weight = make_float3(1.0f, 1.0f, 1.0f);
-                               else
-                                       holdout_weight = shader_holdout_eval(kg, &sd);
-
-                               /* any throughput is ok, should all be identical here */
-                               L_transparent += average(holdout_weight*throughput);
-                       }
-
-                       if(sd.flag & SD_HOLDOUT_MASK)
-                               break;
+               /* Shade background. */
+               if(!hit) {
+                       kernel_path_background(kg, state, ray, throughput, sd, L);
+                       break;
                }
-#endif
-
-               /* holdout mask objects do not write data passes */
-               kernel_write_data_passes(kg, buffer, &L, &sd, sample, &state, throughput);
-
-               /* blurring of bsdf after bounces, for rays that have a small likelihood
-                * of following this particular path (diffuse, rough glossy) */
-               if(kernel_data.integrator.filter_glossy != FLT_MAX) {
-                       float blur_pdf = kernel_data.integrator.filter_glossy*state.min_ray_pdf;
-
-                       if(blur_pdf < 1.0f) {
-                               float blur_roughness = sqrtf(1.0f - blur_pdf)*0.5f;
-                               shader_bsdf_blur(kg, &sd, blur_roughness);
-                       }
+               else if(path_state_ao_bounce(kg, state)) {
+                       break;
                }
 
-#ifdef __EMISSION__
-               /* emission */
-               if(sd.flag & SD_EMISSION) {
-                       /* todo: is isect.t wrong here for transparent surfaces? */
-                       float3 emission = indirect_primitive_emission(kg, &sd, isect.t, state.flag, state.ray_pdf);
-                       path_radiance_accum_emission(&L, throughput, emission, state.bounce);
+               /* Setup and evaluate shader. */
+               shader_setup_from_ray(kg,
+                                     sd,
+                                     &isect,
+                                     ray);
+               shader_eval_surface(kg, sd, state, state->flag, kernel_data.integrator.max_closures);
+               shader_prepare_closures(sd, state);
+
+               /* Apply shadow catcher, holdout, emission. */
+               if(!kernel_path_shader_apply(kg,
+                                            sd,
+                                            state,
+                                            ray,
+                                            throughput,
+                                            emission_sd,
+                                            L,
+                                            NULL))
+               {
+                       break;
                }
-#endif
 
                /* path termination. this is a strange place to put the termination, it's
                 * mainly due to the mixed in MIS that we use. gives too many unneeded
                 * shader evaluations, only need emission if we are going to terminate */
-               float probability = path_state_terminate_probability(kg, &state, throughput);
+               float probability = path_state_continuation_probability(kg, state, throughput);
 
                if(probability == 0.0f) {
                        break;
                }
                else if(probability != 1.0f) {
-                       float terminate = path_state_rng_1D_for_decision(kg, rng, &state, PRNG_TERMINATE);
+                       float terminate = path_state_rng_1D(kg, state, PRNG_TERMINATE);
 
                        if(terminate >= probability)
                                break;
@@ -640,517 +476,255 @@ ccl_device float4 kernel_path_integrate(KernelGlobals *kg, RNG *rng, int sample,
                        throughput /= probability;
                }
 
+               kernel_update_denoising_features(kg, sd, state, L);
+
 #ifdef __AO__
                /* ambient occlusion */
-               if(kernel_data.integrator.use_ambient_occlusion || (sd.flag & SD_AO)) {
-                       kernel_path_ao(kg, &sd, &L, &state, rng, throughput);
+               if(kernel_data.integrator.use_ambient_occlusion || (sd->flag & SD_AO)) {
+                       kernel_path_ao(kg, sd, emission_sd, L, state, throughput, make_float3(0.0f, 0.0f, 0.0f));
                }
-#endif
+#endif  /* __AO__ */
 
 #ifdef __SUBSURFACE__
                /* bssrdf scatter to a different location on the same object, replacing
                 * the closures with a diffuse BSDF */
-               if(sd.flag & SD_BSSRDF) {
-                       if(kernel_path_subsurface_scatter(kg, &sd, &L, &state, rng, &ray, &throughput))
-                               break;
-               }
-#endif
-
-               /* direct lighting */
-               kernel_path_surface_connect_light(kg, rng, &sd, throughput, &state, &L);
-
-               /* compute direct lighting and next bounce */
-               if(!kernel_path_surface_bounce(kg, rng, &sd, &throughput, &state, &L, &ray))
-                       break;
-       }
-
-       float3 L_sum = path_radiance_clamp_and_sum(kg, &L);
-
-       kernel_write_light_passes(kg, buffer, &L, sample);
-
-       return make_float4(L_sum.x, L_sum.y, L_sum.z, 1.0f - L_transparent);
-}
-
-#ifdef __BRANCHED_PATH__
-
-/* branched path tracing: 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)
-{
-       for(int i = 0; i< sd->num_closure; i++) {
-               const ShaderClosure *sc = &sd->closure[i];
-
-               if(!CLOSURE_IS_BSDF(sc->type))
-                       continue;
-               /* transparency is not handled here, but in outer loop */
-               if(sc->type == CLOSURE_BSDF_TRANSPARENT_ID)
-                       continue;
-
-               int num_samples;
-
-               if(CLOSURE_IS_BSDF_DIFFUSE(sc->type))
-                       num_samples = kernel_data.integrator.diffuse_samples;
-               else if(CLOSURE_IS_BSDF_BSSRDF(sc->type))
-                       num_samples = 1;
-               else if(CLOSURE_IS_BSDF_GLOSSY(sc->type))
-                       num_samples = kernel_data.integrator.glossy_samples;
-               else
-                       num_samples = kernel_data.integrator.transmission_samples;
-
-               num_samples = ceil_to_int(num_samples_adjust*num_samples);
-
-               float num_samples_inv = num_samples_adjust/num_samples;
-               RNG bsdf_rng = cmj_hash(*rng, i);
-
-               for(int j = 0; j < num_samples; j++) {
-                       PathState ps = *state;
-                       float3 tp = throughput;
-                       Ray bsdf_ray;
-
-                       if(!kernel_branched_path_surface_bounce(kg, &bsdf_rng, sd, sc, j, num_samples, &tp, &ps, L, &bsdf_ray))
-                               continue;
-
-                       kernel_path_indirect(kg, rng, bsdf_ray, tp*num_samples_inv, num_samples, ps, L);
-
-                       /* for render passes, sum and reset indirect light pass variables
-                        * for the next samples */
-                       path_radiance_sum_indirect(L);
-                       path_radiance_reset_indirect(L);
-               }
-       }
-}
-
-#ifdef __SUBSURFACE__
-ccl_device void kernel_branched_path_subsurface_scatter(KernelGlobals *kg, ShaderData *sd, PathRadiance *L, PathState *state, RNG *rng, float3 throughput)
-{
-       for(int i = 0; i< sd->num_closure; i++) {
-               ShaderClosure *sc = &sd->closure[i];
-
-               if(!CLOSURE_IS_BSSRDF(sc->type))
-                       continue;
-
-               /* set up random number generator */
-               uint lcg_state = lcg_state_init(rng, state, 0x68bc21eb);
-               int num_samples = kernel_data.integrator.subsurface_samples;
-               float num_samples_inv = 1.0f/num_samples;
-               RNG bssrdf_rng = cmj_hash(*rng, i);
-
-               state->flag |= PATH_RAY_BSSRDF_ANCESTOR;
-
-               /* do subsurface scatter step with copy of shader data, this will
-                * replace the BSSRDF with a diffuse BSDF closure */
-               for(int j = 0; j < num_samples; j++) {
-                       ShaderData bssrdf_sd[BSSRDF_MAX_HITS];
+               if(sd->flag & SD_BSSRDF) {
                        float bssrdf_u, bssrdf_v;
-                       path_branched_rng_2D(kg, &bssrdf_rng, state, j, num_samples, PRNG_BSDF_U, &bssrdf_u, &bssrdf_v);
-                       int num_hits = subsurface_scatter_multi_step(kg, sd, bssrdf_sd, state->flag, sc, &lcg_state, bssrdf_u, bssrdf_v, true);
-
-                       /* compute lighting with the BSDF closure */
-                       for(int hit = 0; hit < num_hits; hit++) {
-                               PathState hit_state = *state;
-
-                               path_state_branch(&hit_state, j, num_samples);
-
-#if defined(__EMISSION__) && defined(__BRANCHED_PATH__)
-                               /* direct light */
-                               if(kernel_data.integrator.use_direct_light) {
-                                       bool all = kernel_data.integrator.sample_all_lights_direct;
-                                       kernel_branched_path_surface_connect_light(kg, rng,
-                                               &bssrdf_sd[hit], &hit_state, throughput, num_samples_inv, L, all);
-                               }
-#endif
+                       path_state_rng_2D(kg,
+                                         state,
+                                         PRNG_BSDF_U,
+                                         &bssrdf_u, &bssrdf_v);
 
-                               /* indirect light */
-                               kernel_branched_path_surface_indirect_light(kg, rng,
-                                       &bssrdf_sd[hit], throughput, num_samples_inv,
-                                       &hit_state, L);
+                       const ShaderClosure *sc = shader_bssrdf_pick(sd, &throughput, &bssrdf_u);
+
+                       /* do bssrdf scatter step if we picked a bssrdf closure */
+                       if(sc) {
+                               uint lcg_state = lcg_state_init(state, 0x68bc21eb);
+
+                               subsurface_scatter_step(kg,
+                                                       sd,
+                                                       state,
+                                                       state->flag,
+                                                       sc,
+                                                       &lcg_state,
+                                                       bssrdf_u, bssrdf_v,
+                                                       false);
                        }
                }
+#endif  /* __SUBSURFACE__ */
 
-               state->flag &= ~PATH_RAY_BSSRDF_ANCESTOR;
+#if defined(__EMISSION__)
+               if(kernel_data.integrator.use_direct_light) {
+                       int all = (kernel_data.integrator.sample_all_lights_indirect) ||
+                                 (state->flag & PATH_RAY_SHADOW_CATCHER);
+                       kernel_branched_path_surface_connect_light(kg,
+                                                                  sd,
+                                                                  emission_sd,
+                                                                  state,
+                                                                  throughput,
+                                                                  1.0f,
+                                                                  L,
+                                                                  all);
+               }
+#endif  /* defined(__EMISSION__) */
+
+               if(!kernel_path_surface_bounce(kg, sd, &throughput, state, &L->state, ray))
+                       break;
        }
 }
-#endif
 
-ccl_device float4 kernel_branched_path_integrate(KernelGlobals *kg, RNG *rng, int sample, Ray ray, ccl_global float *buffer)
+#endif /* defined(__BRANCHED_PATH__) || defined(__BAKING__) */
+
+ccl_device_forceinline void kernel_path_integrate(
+       KernelGlobals *kg,
+       PathState *state,
+       float3 throughput,
+       Ray *ray,
+       PathRadiance *L,
+       ccl_global float *buffer,
+       ShaderData *emission_sd)
 {
-       /* initialize */
-       PathRadiance L;
-       float3 throughput = make_float3(1.0f, 1.0f, 1.0f);
-       float L_transparent = 0.0f;
+       /* Shader data memory used for both volumes and surfaces, saves stack space. */
+       ShaderData sd;
 
-       path_radiance_init(&L, kernel_data.film.use_light_pass);
+#ifdef __SUBSURFACE__
+       SubsurfaceIndirectRays ss_indirect;
+       kernel_path_subsurface_init_indirect(&ss_indirect);
 
-       PathState state;
-       path_state_init(kg, &state, rng, sample);
+       for(;;) {
+#endif  /* __SUBSURFACE__ */
 
+       /* path iteration */
        for(;;) {
-               /* intersect scene */
+               /* Find intersection with objects in scene. */
                Intersection isect;
-               uint visibility = path_state_ray_visibility(kg, &state);
-
-#ifdef __HAIR__
-               float difl = 0.0f, extmax = 0.0f;
-               uint lcg_state = 0;
-
-               if(kernel_data.bvh.have_curves) {
-                       if((kernel_data.cam.resolution == 1) && (state.flag & PATH_RAY_CAMERA)) {       
-                               float3 pixdiff = ray.dD.dx + ray.dD.dy;
-                               /*pixdiff = pixdiff - dot(pixdiff, ray.D)*ray.D;*/
-                               difl = kernel_data.curve.minimum_width * len(pixdiff) * 0.5f;
-                       }
-
-                       extmax = kernel_data.curve.maximum_width;
-                       lcg_state = lcg_state_init(rng, &state, 0x51633e2d);
-               }
+               bool hit = kernel_path_scene_intersect(kg, state, ray, &isect, L);
 
-               bool hit = scene_intersect(kg, &ray, visibility, &isect, &lcg_state, difl, extmax);
-#else
-               bool hit = scene_intersect(kg, &ray, visibility, &isect);
-#endif
+               /* Find intersection with lamps and compute emission for MIS. */
+               kernel_path_lamp_emission(kg, state, ray, throughput, &isect, &sd, L);
 
 #ifdef __VOLUME__
-               /* volume attenuation, emission, scatter */
-               if(state.volume_stack[0].shader != SHADER_NONE) {
-                       Ray volume_ray = ray;
-                       volume_ray.t = (hit)? isect.t: FLT_MAX;
-                       
-                       bool heterogeneous = volume_stack_is_heterogeneous(kg, state.volume_stack);
-
-#ifdef __VOLUME_DECOUPLED__
-                       /* decoupled ray marching only supported on CPU */
-
-                       /* cache steps along volume for repeated sampling */
-                       VolumeSegment volume_segment;
-                       ShaderData volume_sd;
-
-                       shader_setup_from_volume(kg, &volume_sd, &volume_ray, state.bounce, state.transparent_bounce);
-                       kernel_volume_decoupled_record(kg, &state,
-                               &volume_ray, &volume_sd, &volume_segment, heterogeneous);
-
-                       /* direct light sampling */
-                       if(volume_segment.closure_flag & SD_SCATTER) {
-                               volume_segment.sampling_method = volume_stack_sampling_method(kg, state.volume_stack);
-
-                               bool all = kernel_data.integrator.sample_all_lights_direct;
-
-                               kernel_branched_path_volume_connect_light(kg, rng, &volume_sd,
-                                       throughput, &state, &L, 1.0f, all, &volume_ray, &volume_segment);
-
-                               /* indirect light sampling */
-                               int num_samples = kernel_data.integrator.volume_samples;
-                               float num_samples_inv = 1.0f/num_samples;
-
-                               for(int j = 0; j < num_samples; j++) {
-                                       /* workaround to fix correlation bug in T38710, can find better solution
-                                        * in random number generator later, for now this is done here to not impact
-                                        * performance of rendering without volumes */
-                                       RNG tmp_rng = cmj_hash(*rng, state.rng_offset);
-
-                                       PathState ps = state;
-                                       Ray pray = ray;
-                                       float3 tp = throughput;
-
-                                       /* branch RNG state */
-                                       path_state_branch(&ps, j, num_samples);
-
-                                       /* scatter sample. if we use distance sampling and take just one
-                                        * sample for direct and indirect light, we could share this
-                                        * computation, but makes code a bit complex */
-                                       float rphase = path_state_rng_1D_for_decision(kg, &tmp_rng, &ps, PRNG_PHASE);
-                                       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);
-                                               
-                                       (void)result;
-                                       kernel_assert(result == VOLUME_PATH_SCATTERED);
-
-                                       if(kernel_path_volume_bounce(kg, rng, &volume_sd, &tp, &ps, &L, &pray)) {
-                                               kernel_path_indirect(kg, rng, pray, tp*num_samples_inv, num_samples, ps, &L);
-
-                                               /* for render passes, sum and reset indirect light pass variables
-                                                * for the next samples */
-                                               path_radiance_sum_indirect(&L);
-                                               path_radiance_reset_indirect(&L);
-                                       }
-                               }
-                       }
-
-                       /* emission and transmittance */
-                       if(volume_segment.closure_flag & SD_EMISSION)
-                               path_radiance_accum_emission(&L, throughput, volume_segment.accum_emission, state.bounce);
-                       throughput *= volume_segment.accum_transmittance;
-
-                       /* free cached steps */
-                       kernel_volume_decoupled_free(kg, &volume_segment);
-#else
-                       /* GPU: no decoupled ray marching, scatter probalistically */
-                       int num_samples = kernel_data.integrator.volume_samples;
-                       float num_samples_inv = 1.0f/num_samples;
-
-                       /* todo: we should cache the shader evaluations from stepping
-                        * through the volume, for now we redo them multiple times */
-
-                       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);
-                               
-#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);
-
-                                       if(kernel_path_volume_bounce(kg, rng, &volume_sd, &tp, &ps, &L, &pray)) {
-                                               kernel_path_indirect(kg, rng, pray, tp, num_samples, ps, &L);
-
-                                               /* for render passes, sum and reset indirect light pass variables
-                                                * for the next samples */
-                                               path_radiance_sum_indirect(&L);
-                                               path_radiance_reset_indirect(&L);
-                                       }
-                               }
-#endif
-                       }
-
-                       /* todo: avoid this calculation using decoupled ray marching */
-                       kernel_volume_shadow(kg, &state, &volume_ray, &throughput);
-#endif
+               /* Volume integration. */
+               VolumeIntegrateResult result = kernel_path_volume(kg,
+                                                                  &sd,
+                                                                  state,
+                                                                  ray,
+                                                                  &throughput,
+                                                                  &isect,
+                                                                  hit,
+                                                                  emission_sd,
+                                                                  L);
+
+               if(result == VOLUME_PATH_SCATTERED) {
+                       continue;
                }
-#endif
+               else if(result == VOLUME_PATH_MISSED) {
+                       break;
+               }
+#endif /* __VOLUME__*/
 
+               /* Shade background. */
                if(!hit) {
-                       /* eval background shader if nothing hit */
-                       if(kernel_data.background.transparent) {
-                               L_transparent += average(throughput);
-
-#ifdef __PASSES__
-                               if(!(kernel_data.film.pass_flag & PASS_BACKGROUND))
-#endif
-                                       break;
-                       }
-
-#ifdef __BACKGROUND__
-                       /* sample background shader */
-                       float3 L_background = indirect_background(kg, &state, &ray);
-                       path_radiance_accum_background(&L, throughput, L_background, state.bounce);
-#endif
-
+                       kernel_path_background(kg, state, ray, throughput, &sd, L);
                        break;
                }
-
-               /* setup shading */
-               ShaderData sd;
-               shader_setup_from_ray(kg, &sd, &isect, &ray, state.bounce, state.transparent_bounce);
-               shader_eval_surface(kg, &sd, 0.0f, state.flag, SHADER_CONTEXT_MAIN);
-               shader_merge_closures(&sd);
-
-               /* holdout */
-#ifdef __HOLDOUT__
-               if(sd.flag & (SD_HOLDOUT|SD_HOLDOUT_MASK)) {
-                       if(kernel_data.background.transparent) {
-                               float3 holdout_weight;
-                               
-                               if(sd.flag & SD_HOLDOUT_MASK)
-                                       holdout_weight = make_float3(1.0f, 1.0f, 1.0f);
-                               else
-                                       holdout_weight = shader_holdout_eval(kg, &sd);
-
-                               /* any throughput is ok, should all be identical here */
-                               L_transparent += average(holdout_weight*throughput);
-                       }
-
-                       if(sd.flag & SD_HOLDOUT_MASK)
-                               break;
+               else if(path_state_ao_bounce(kg, state)) {
+                       break;
                }
-#endif
-
-               /* holdout mask objects do not write data passes */
-               kernel_write_data_passes(kg, buffer, &L, &sd, sample, &state, throughput);
 
-#ifdef __EMISSION__
-               /* emission */
-               if(sd.flag & SD_EMISSION) {
-                       float3 emission = indirect_primitive_emission(kg, &sd, isect.t, state.flag, state.ray_pdf);
-                       path_radiance_accum_emission(&L, throughput, emission, state.bounce);
+               /* Setup and evaluate shader. */
+               shader_setup_from_ray(kg, &sd, &isect, ray);
+               shader_eval_surface(kg, &sd, state, state->flag, kernel_data.integrator.max_closures);
+               shader_prepare_closures(&sd, state);
+
+               /* Apply shadow catcher, holdout, emission. */
+               if(!kernel_path_shader_apply(kg,
+                                            &sd,
+                                            state,
+                                            ray,
+                                            throughput,
+                                            emission_sd,
+                                            L,
+                                            buffer))
+               {
+                       break;
                }
-#endif
 
-               /* transparency termination */
-               if(state.flag & PATH_RAY_TRANSPARENT) {
-                       /* path termination. this is a strange place to put the termination, it's
-                        * mainly due to the mixed in MIS that we use. gives too many unneeded
-                        * shader evaluations, only need emission if we are going to terminate */
-                       float probability = path_state_terminate_probability(kg, &state, throughput);
+               /* path termination. this is a strange place to put the termination, it's
+                * mainly due to the mixed in MIS that we use. gives too many unneeded
+                * shader evaluations, only need emission if we are going to terminate */
+               float probability = path_state_continuation_probability(kg, state, throughput);
 
-                       if(probability == 0.0f) {
+               if(probability == 0.0f) {
+                       break;
+               }
+               else if(probability != 1.0f) {
+                       float terminate = path_state_rng_1D(kg, state, PRNG_TERMINATE);
+                       if(terminate >= probability)
                                break;
-                       }
-                       else if(probability != 1.0f) {
-                               float terminate = path_state_rng_1D_for_decision(kg, rng, &state, PRNG_TERMINATE);
-
-                               if(terminate >= probability)
-                                       break;
 
-                               throughput /= probability;
-                       }
+                       throughput /= probability;
                }
 
+               kernel_update_denoising_features(kg, &sd, state, L);
+
 #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_path_ao(kg, &sd, emission_sd, L, state, throughput, shader_bsdf_alpha(kg, &sd));
                }
-#endif
+#endif  /* __AO__ */
 
 #ifdef __SUBSURFACE__
-               /* bssrdf scatter to a different location on the same object */
+               /* bssrdf scatter to a different location on the same object, replacing
+                * the closures with a diffuse BSDF */
                if(sd.flag & SD_BSSRDF) {
-                       kernel_branched_path_subsurface_scatter(kg, &sd, &L, &state, rng, throughput);
-               }
-#endif
-
-               if(!(sd.flag & SD_HAS_ONLY_VOLUME)) {
-                       PathState hit_state = state;
-
-#ifdef __EMISSION__
-                       /* direct light */
-                       if(kernel_data.integrator.use_direct_light) {
-                               bool all = kernel_data.integrator.sample_all_lights_direct;
-                               kernel_branched_path_surface_connect_light(kg, rng,
-                                       &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);
-
-                       /* continue in case of transparency */
-                       throughput *= shader_bsdf_transparency(kg, &sd);
-
-                       if(is_zero(throughput))
+                       if(kernel_path_subsurface_scatter(kg,
+                                                         &sd,
+                                                         emission_sd,
+                                                         L,
+                                                         state,
+                                                         ray,
+                                                         &throughput,
+                                                         &ss_indirect))
+                       {
                                break;
+                       }
                }
+#endif  /* __SUBSURFACE__ */
 
-               path_state_next(kg, &state, LABEL_TRANSPARENT);
-               ray.P = ray_offset(sd.P, -sd.Ng);
-               ray.t -= sd.ray_length; /* clipping works through transparent */
+               /* direct lighting */
+               kernel_path_surface_connect_light(kg, &sd, emission_sd, throughput, state, L);
 
-#ifdef __VOLUME__
-               /* enter/exit volume */
-               kernel_volume_stack_enter_exit(kg, &sd, state.volume_stack);
-#endif
+               /* compute direct lighting and next bounce */
+               if(!kernel_path_surface_bounce(kg, &sd, &throughput, state, &L->state, ray))
+                       break;
        }
 
-       float3 L_sum = path_radiance_clamp_and_sum(kg, &L);
-
-       kernel_write_light_passes(kg, buffer, &L, sample);
-
-       return make_float4(L_sum.x, L_sum.y, L_sum.z, 1.0f - L_transparent);
-}
-
-#endif
-
-ccl_device_inline void kernel_path_trace_setup(KernelGlobals *kg, ccl_global uint *rng_state, int sample, int x, int y, RNG *rng, Ray *ray)
-{
-       float filter_u;
-       float filter_v;
-
-       int num_samples = kernel_data.integrator.aa_samples;
-
-       path_rng_init(kg, rng_state, sample, num_samples, rng, x, y, &filter_u, &filter_v);
-
-       /* sample camera ray */
-
-       float lens_u = 0.0f, lens_v = 0.0f;
-
-       if(kernel_data.cam.aperturesize > 0.0f)
-               path_rng_2D(kg, rng, sample, num_samples, PRNG_LENS_U, &lens_u, &lens_v);
-
-       float time = 0.0f;
-
-#ifdef __CAMERA_MOTION__
-       if(kernel_data.cam.shuttertime != -1.0f)
-               time = path_rng_1D(kg, rng, sample, num_samples, PRNG_TIME);
-#endif
-
-       camera_sample(kg, x, y, filter_u, filter_v, lens_u, lens_v, time, ray);
+#ifdef __SUBSURFACE__
+               /* Trace indirect subsurface rays by restarting the loop. this uses less
+                * stack memory than invoking kernel_path_indirect.
+                */
+               if(ss_indirect.num_rays) {
+                       kernel_path_subsurface_setup_indirect(kg,
+                                                             &ss_indirect,
+                                                             state,
+                                                             ray,
+                                                             L,
+                                                             &throughput);
+               }
+               else {
+                       break;
+               }
+       }
+#endif  /* __SUBSURFACE__ */
 }
 
 ccl_device void kernel_path_trace(KernelGlobals *kg,
-       ccl_global float *buffer, ccl_global uint *rng_state,
+       ccl_global float *buffer,
        int sample, int x, int y, int offset, int stride)
 {
        /* buffer offset */
        int index = offset + x + y*stride;
        int pass_stride = kernel_data.film.pass_stride;
 
-       rng_state += index;
        buffer += index*pass_stride;
 
-       /* initialize random numbers and ray */
-       RNG rng;
+       /* Initialize random numbers and sample ray. */
+       uint rng_hash;
        Ray ray;
 
-       kernel_path_trace_setup(kg, rng_state, sample, x, y, &rng, &ray);
+       kernel_path_trace_setup(kg, sample, x, y, &rng_hash, &ray);
 
-       /* integrate */
-       float4 L;
-
-       if(ray.t != 0.0f)
-               L = kernel_path_integrate(kg, &rng, sample, ray, buffer);
-       else
-               L = make_float4(0.0f, 0.0f, 0.0f, 0.0f);
-
-       /* accumulate result in output buffer */
-       kernel_write_pass_float4(buffer, sample, L);
-
-       path_rng_end(kg, rng_state, rng);
-}
-
-#ifdef __BRANCHED_PATH__
-ccl_device void kernel_branched_path_trace(KernelGlobals *kg,
-       ccl_global float *buffer, ccl_global uint *rng_state,
-       int sample, int x, int y, int offset, int stride)
-{
-       /* buffer offset */
-       int index = offset + x + y*stride;
-       int pass_stride = kernel_data.film.pass_stride;
-
-       rng_state += index;
-       buffer += index*pass_stride;
-
-       /* initialize random numbers and ray */
-       RNG rng;
-       Ray ray;
-
-       kernel_path_trace_setup(kg, rng_state, sample, x, y, &rng, &ray);
+       if(ray.t == 0.0f) {
+               return;
+       }
 
-       /* integrate */
-       float4 L;
+       /* Initialize state. */
+       float3 throughput = make_float3(1.0f, 1.0f, 1.0f);
 
-       if(ray.t != 0.0f)
-               L = kernel_branched_path_integrate(kg, &rng, sample, ray, buffer);
-       else
-               L = make_float4(0.0f, 0.0f, 0.0f, 0.0f);
+       PathRadiance L;
+       path_radiance_init(&L, kernel_data.film.use_light_pass);
 
-       /* accumulate result in output buffer */
-       kernel_write_pass_float4(buffer, sample, L);
+       ShaderDataTinyStorage emission_sd_storage;
+       ShaderData *emission_sd = AS_SHADER_DATA(&emission_sd_storage);
 
-       path_rng_end(kg, rng_state, rng);
+       PathState state;
+       path_state_init(kg, emission_sd, &state, rng_hash, sample, &ray);
+
+       /* Integrate. */
+       kernel_path_integrate(kg,
+                             &state,
+                             throughput,
+                             &ray,
+                             &L,
+                             buffer,
+                             emission_sd);
+
+       kernel_write_result(kg, buffer, sample, &L);
 }
-#endif
+
+#endif  /* __SPLIT_KERNEL__ */
 
 CCL_NAMESPACE_END