Cycles: use defensive sampling for picking BSDFs and BSSRDFs.

For the first bounce we now give each BSDF or BSSRDF a minimum sample weight,
which helps reduce noise for a typical case where you have a glossy BSDF with
a small weight due to Fresnel, but not necessarily small contribution relative
to a diffuse or transmission BSDF below.

We can probably find a better heuristic that also enables this on further
bounces, for example when looking through a perfect mirror, but I wasn't able
to find a robust one so far.

3 files changed, 38 insertions, 5 deletions

diff --git a/intern/cycles/kernel/kernel_path.h b/intern/cycles/kernel/kernel_path.h
index 2df21f1cda3..d43d6374c13 100644
--- a/intern/cycles/kernel/kernel_path.h
+++ b/intern/cycles/kernel/kernel_path.h
@@ -435,9 +435,7 @@ ccl_device void kernel_path_indirect(KernelGlobals *kg,
 		                      &isect,
 		                      ray);
 		shader_eval_surface(kg, sd, state, state->flag);
-#ifdef __BRANCHED_PATH__
-		shader_merge_closures(sd);
-#endif  /* __BRANCHED_PATH__ */
+		shader_prepare_closures(sd, state);
 
 		/* Apply shadow catcher, holdout, emission. */
 		if(!kernel_path_shader_apply(kg,
@@ -588,6 +586,7 @@ ccl_device_forceinline void kernel_path_integrate(
 		/* Setup and evaluate shader. */
 		shader_setup_from_ray(kg, &sd, &isect, ray);
 		shader_eval_surface(kg, &sd, state, state->flag);
+		shader_prepare_closures(&sd, state);
 
 		/* Apply shadow catcher, holdout, emission. */
 		if(!kernel_path_shader_apply(kg,
diff --git a/intern/cycles/kernel/kernel_shader.h b/intern/cycles/kernel/kernel_shader.h
index 5ef4475e259..bb3add5d7ca 100644
--- a/intern/cycles/kernel/kernel_shader.h
+++ b/intern/cycles/kernel/kernel_shader.h
@@ -494,6 +494,34 @@ ccl_device_inline void shader_merge_closures(ShaderData *sd)
 }
 #endif
 
+/* Defensive sampling. */
+
+ccl_device_inline void shader_prepare_closures(ShaderData *sd,
+                                               ccl_addr_space PathState *state)
+{
+	/* We can likely also do defensive sampling at deeper bounces, particularly
+	 * for cases like a perfect mirror but possibly also others. This will need
+	 * a good heuristic. */
+	if(state->bounce + state->transparent_bounce == 0 && sd->num_closure > 1) {
+		float sum = 0.0f;
+
+		for(int i = 0; i < sd->num_closure; i++) {
+			ShaderClosure *sc = &sd->closure[i];
+			if(CLOSURE_IS_BSDF_OR_BSSRDF(sc->type)) {
+				sum += sc->sample_weight;
+			}
+		}
+
+		for(int i = 0; i < sd->num_closure; i++) {
+			ShaderClosure *sc = &sd->closure[i];
+			if(CLOSURE_IS_BSDF_OR_BSSRDF(sc->type)) {
+				sc->sample_weight = max(sc->sample_weight, 0.125f * sum);
+			}
+		}
+	}
+}
+
+
 /* BSDF */
 
 ccl_device_inline void _shader_bsdf_multi_eval(KernelGlobals *kg, ShaderData *sd, const float3 omega_in, float *pdf,
diff --git a/intern/cycles/kernel/split/kernel_shader_eval.h b/intern/cycles/kernel/split/kernel_shader_eval.h
index c792c4adb0e..7032461b04a 100644
--- a/intern/cycles/kernel/split/kernel_shader_eval.h
+++ b/intern/cycles/kernel/split/kernel_shader_eval.h
@@ -52,8 +52,14 @@ ccl_device void kernel_shader_eval(KernelGlobals *kg)
 
 		shader_eval_surface(kg, &kernel_split_state.sd[ray_index], state, state->flag);
 #ifdef __BRANCHED_PATH__
-		shader_merge_closures(&kernel_split_state.sd[ray_index]);
-#endif  /* __BRANCHED_PATH__ */
+		if(kernel_data.integrator.branched) {
+			shader_merge_closures(&kernel_split_state.sd[ray_index]);
+		}
+		else
+#endif
+		{
+			shader_prepare_closures(&kernel_split_state.sd[ray_index], state);
+		}
 	}
 }