Cycles: Implement denoising option for reducing noise in the rendered image

This commit contains the first part of the new Cycles denoising option,
which filters the resulting image using information gathered during rendering
to get rid of noise while preserving visual features as well as possible.

To use the option, enable it in the render layer options. The default settings
fit a wide range of scenes, but the user can tweak individual settings to
control the tradeoff between a noise-free image, image details, and calculation
time.

Note that the denoiser may still change in the future and that some features
are not implemented yet. The most important missing feature is animation
denoising, which uses information from multiple frames at once to produce a
flicker-free and smoother result. These features will be added in the future.

Finally, thanks to all the people who supported this project:

- Google (through the GSoC) and Theory Studios for sponsoring the development
- The authors of the papers I used for implementing the denoiser (more details
  on them will be included in the technical docs)
- The other Cycles devs for feedback on the code, especially Sergey for
  mentoring the GSoC project and Brecht for the code review!
- And of course the users who helped with testing, reported bugs and things
  that could and/or should work better!

1 files changed, 203 insertions, 0 deletions

diff --git a/intern/cycles/kernel/kernel_passes.h b/intern/cycles/kernel/kernel_passes.h
index ed523696571..8ab4c724829 100644
--- a/intern/cycles/kernel/kernel_passes.h
+++ b/intern/cycles/kernel/kernel_passes.h
@@ -60,6 +60,135 @@ ccl_device_inline void kernel_write_pass_float4(ccl_global float *buffer, int sa
 #endif  /* __SPLIT_KERNEL__ */
 }
 
+#ifdef __DENOISING_FEATURES__
+ccl_device_inline void kernel_write_pass_float_variance(ccl_global float *buffer, int sample, float value)
+{
+	kernel_write_pass_float(buffer, sample, value);
+
+	/* The online one-pass variance update that's used for the megakernel can't easily be implemented
+	 * with atomics, so for the split kernel the E[x^2] - 1/N * (E[x])^2 fallback is used. */
+#  ifdef __SPLIT_KERNEL__
+	kernel_write_pass_float(buffer+1, sample, value*value);
+#  else
+	if(sample == 0) {
+		kernel_write_pass_float(buffer+1, sample, 0.0f);
+	}
+	else {
+		float new_mean = buffer[0] * (1.0f / (sample + 1));
+		float old_mean = (buffer[0] - value) * (1.0f / sample);
+		kernel_write_pass_float(buffer+1, sample, (value - new_mean) * (value - old_mean));
+	}
+#  endif
+}
+
+#  if defined(__SPLIT_KERNEL__)
+#    define kernel_write_pass_float3_unaligned kernel_write_pass_float3
+#  else
+ccl_device_inline void kernel_write_pass_float3_unaligned(ccl_global float *buffer, int sample, float3 value)
+{
+	buffer[0] = (sample == 0)? value.x: buffer[0] + value.x;
+	buffer[1] = (sample == 0)? value.y: buffer[1] + value.y;
+	buffer[2] = (sample == 0)? value.z: buffer[2] + value.z;
+}
+#  endif
+
+ccl_device_inline void kernel_write_pass_float3_variance(ccl_global float *buffer, int sample, float3 value)
+{
+	kernel_write_pass_float3_unaligned(buffer, sample, value);
+#  ifdef __SPLIT_KERNEL__
+	kernel_write_pass_float3_unaligned(buffer+3, sample, value*value);
+#  else
+	if(sample == 0) {
+		kernel_write_pass_float3_unaligned(buffer+3, sample, make_float3(0.0f, 0.0f, 0.0f));
+	}
+	else {
+		float3 sum = make_float3(buffer[0], buffer[1], buffer[2]);
+		float3 new_mean = sum * (1.0f / (sample + 1));
+		float3 old_mean = (sum - value) * (1.0f / sample);
+		kernel_write_pass_float3_unaligned(buffer+3, sample, (value - new_mean) * (value - old_mean));
+	}
+#  endif
+}
+
+ccl_device_inline void kernel_write_denoising_shadow(KernelGlobals *kg, ccl_global float *buffer,
+	int sample, float path_total, float path_total_shaded)
+{
+	if(kernel_data.film.pass_denoising_data == 0)
+		return;
+
+	buffer += (sample & 1)? DENOISING_PASS_SHADOW_B : DENOISING_PASS_SHADOW_A;
+
+	path_total = ensure_finite(path_total);
+	path_total_shaded = ensure_finite(path_total_shaded);
+
+	kernel_write_pass_float(buffer, sample/2, path_total);
+	kernel_write_pass_float(buffer+1, sample/2, path_total_shaded);
+
+	float value = path_total_shaded / max(path_total, 1e-7f);
+#  ifdef __SPLIT_KERNEL__
+	kernel_write_pass_float(buffer+2, sample/2, value*value);
+#  else
+	if(sample < 2) {
+		kernel_write_pass_float(buffer+2, sample/2, 0.0f);
+	}
+	else {
+		float old_value = (buffer[1] - path_total_shaded) / max(buffer[0] - path_total, 1e-7f);
+		float new_value = buffer[1] / max(buffer[0], 1e-7f);
+		kernel_write_pass_float(buffer+2, sample, (value - new_value) * (value - old_value));
+	}
+#  endif
+}
+#endif /* __DENOISING_FEATURES__ */
+
+ccl_device_inline void kernel_update_denoising_features(KernelGlobals *kg,
+                                                        ShaderData *sd,
+                                                        ccl_global PathState *state,
+                                                        PathRadiance *L)
+{
+#ifdef __DENOISING_FEATURES__
+	if(state->denoising_feature_weight == 0.0f) {
+		return;
+	}
+
+	L->denoising_depth += ensure_finite(state->denoising_feature_weight * sd->ray_length);
+
+	float3 normal = make_float3(0.0f, 0.0f, 0.0f);
+	float3 albedo = make_float3(0.0f, 0.0f, 0.0f);
+	float sum_weight = 0.0f, sum_nonspecular_weight = 0.0f;
+
+	for(int i = 0; i < sd->num_closure; i++) {
+		ShaderClosure *sc = &sd->closure[i];
+
+		if(!CLOSURE_IS_BSDF_OR_BSSRDF(sc->type))
+			continue;
+
+		/* All closures contribute to the normal feature, but only diffuse-like ones to the albedo. */
+		normal += sc->N * sc->sample_weight;
+		sum_weight += sc->sample_weight;
+		if(!bsdf_is_specular_like(sc)) {
+			albedo += sc->weight;
+			sum_nonspecular_weight += sc->sample_weight;
+		}
+	}
+
+	/* Wait for next bounce if 75% or more sample weight belongs to specular-like closures. */
+	if((sum_weight == 0.0f) || (sum_nonspecular_weight*4.0f > sum_weight)) {
+		if(sum_weight != 0.0f) {
+			normal /= sum_weight;
+		}
+		L->denoising_normal += ensure_finite3(state->denoising_feature_weight * normal);
+		L->denoising_albedo += ensure_finite3(state->denoising_feature_weight * albedo);
+
+		state->denoising_feature_weight = 0.0f;
+	}
+#else
+	(void) kg;
+	(void) sd;
+	(void) state;
+	(void) L;
+#endif  /* __DENOISING_FEATURES__ */
+}
+
 ccl_device_inline void kernel_write_data_passes(KernelGlobals *kg, ccl_global float *buffer, PathRadiance *L,
 	ShaderData *sd, int sample, ccl_addr_space PathState *state, float3 throughput)
 {
@@ -199,5 +328,79 @@ ccl_device_inline void kernel_write_light_passes(KernelGlobals *kg, ccl_global f
 #endif
 }
 
+ccl_device_inline void kernel_write_result(KernelGlobals *kg, ccl_global float *buffer,
+	int sample, PathRadiance *L, float alpha, bool is_shadow_catcher)
+{
+	if(L) {
+		float3 L_sum;
+#ifdef __SHADOW_TRICKS__
+		if(is_shadow_catcher) {
+			L_sum = path_radiance_sum_shadowcatcher(kg, L, &alpha);
+		}
+		else
+#endif  /* __SHADOW_TRICKS__ */
+		{
+			L_sum = path_radiance_clamp_and_sum(kg, L);
+		}
+
+		kernel_write_pass_float4(buffer, sample, make_float4(L_sum.x, L_sum.y, L_sum.z, alpha));
+
+		kernel_write_light_passes(kg, buffer, L, sample);
+
+#ifdef __DENOISING_FEATURES__
+		if(kernel_data.film.pass_denoising_data) {
+#  ifdef __SHADOW_TRICKS__
+			kernel_write_denoising_shadow(kg, buffer + kernel_data.film.pass_denoising_data, sample, average(L->path_total), average(L->path_total_shaded));
+#  else
+			kernel_write_denoising_shadow(kg, buffer + kernel_data.film.pass_denoising_data, sample, 0.0f, 0.0f);
+#  endif
+			if(kernel_data.film.pass_denoising_clean) {
+				float3 noisy, clean;
+				path_radiance_split_denoising(kg, L, &noisy, &clean);
+				kernel_write_pass_float3_variance(buffer + kernel_data.film.pass_denoising_data + DENOISING_PASS_COLOR,
+				                                  sample, noisy);
+				kernel_write_pass_float3_unaligned(buffer + kernel_data.film.pass_denoising_clean,
+				                                   sample, clean);
+			}
+			else {
+				kernel_write_pass_float3_variance(buffer + kernel_data.film.pass_denoising_data + DENOISING_PASS_COLOR,
+				                                  sample, L_sum);
+			}
+
+			kernel_write_pass_float3_variance(buffer + kernel_data.film.pass_denoising_data + DENOISING_PASS_NORMAL,
+			                                  sample, L->denoising_normal);
+			kernel_write_pass_float3_variance(buffer + kernel_data.film.pass_denoising_data + DENOISING_PASS_ALBEDO,
+			                                  sample, L->denoising_albedo);
+			kernel_write_pass_float_variance(buffer + kernel_data.film.pass_denoising_data + DENOISING_PASS_DEPTH,
+			                                 sample, L->denoising_depth);
+		}
+#endif  /* __DENOISING_FEATURES__ */
+	}
+	else {
+		kernel_write_pass_float4(buffer, sample, make_float4(0.0f, 0.0f, 0.0f, 0.0f));
+
+#ifdef __DENOISING_FEATURES__
+		if(kernel_data.film.pass_denoising_data) {
+			kernel_write_denoising_shadow(kg, buffer + kernel_data.film.pass_denoising_data, sample, 0.0f, 0.0f);
+
+			kernel_write_pass_float3_variance(buffer + kernel_data.film.pass_denoising_data + DENOISING_PASS_COLOR,
+			                                  sample, make_float3(0.0f, 0.0f, 0.0f));
+
+			kernel_write_pass_float3_variance(buffer + kernel_data.film.pass_denoising_data + DENOISING_PASS_NORMAL,
+			                                  sample, make_float3(0.0f, 0.0f, 0.0f));
+			kernel_write_pass_float3_variance(buffer + kernel_data.film.pass_denoising_data + DENOISING_PASS_ALBEDO,
+			                                  sample, make_float3(0.0f, 0.0f, 0.0f));
+			kernel_write_pass_float_variance(buffer + kernel_data.film.pass_denoising_data + DENOISING_PASS_DEPTH,
+			                                 sample, 0.0f);
+
+			if(kernel_data.film.pass_denoising_clean) {
+				kernel_write_pass_float3_unaligned(buffer + kernel_data.film.pass_denoising_clean,
+				                                   sample, make_float3(0.0f, 0.0f, 0.0f));
+			}
+		}
+#endif  /* __DENOISING_FEATURES__ */
+	}
+}
+
 CCL_NAMESPACE_END