[path_tracing] cleanup

Author: PanosK92

data/shaders/light_composition.hlsl

@@ -157,16 +157,11 @@ void main_cs(uint3 thread_id : SV_DispatchThreadID)
         // upsample (depth + normal aware) to avoid bleeding across edges
         // also multiply by surface.occlusion to recover contact shadows that
         // restir's spatial reuse and denoiser smear away at small scales
-        // debug mode writes a heatmap into the gi slot, the remodulator is skipped there so
-        // the viridis colors are not tinted by surface albedo
         if (is_restir_pt_enabled())
         {
             float depth_dst_lin = linearize_depth(surface.depth);
-            light_gi = sample_gi_bilateral(surface.uv, depth_dst_lin, surface.normal);
-            if (uint(buffer_frame.restir_pt_debug_mode) == 0u)
-            {
-                light_gi *= restir_albedo_demodulator(surface.albedo);
-            }
+            light_gi  = sample_gi_bilateral(surface.uv, depth_dst_lin, surface.normal);
+            light_gi *= restir_albedo_demodulator(surface.albedo);
             light_gi *= surface.occlusion;
         }
     }

data/shaders/light_image_based.hlsl

@@ -99,23 +99,17 @@ void main_cs(uint3 thread_id : SV_DispatchThreadID)
     float3 diffuse_ibl       = diffuse_skysphere * diffuse_occlusion * diffuse_energy * surface.albedo.rgb;
     float3 specular_ibl      = specular_skysphere * specular_energy * specular_occlusion;
 
-    // ray traced reflections owns specular indirect across the full roughness range when enabled
-    if (is_ray_traced_reflections_enabled())
+    // restir path tracing owns the full brdf at the primary, so it produces both diffuse and
+    // specular indirect, when restir is on we drop both ibl terms to avoid double counting
+    // otherwise rt reflections (when enabled) owns specular and the cubemap diffuse handles ao
+    if (is_restir_pt_enabled())
     {
+        diffuse_ibl  *= 0.0f;
         specular_ibl *= 0.0f;
     }
-
-    // restir path tracing owns diffuse indirect always, and the specular indirect lobe for
-    // moderate to high roughness surfaces (paper-faithful full brdf at the primary), zeroing
-    // specular_ibl in that range avoids double counting when rt reflections is off, near mirror
-    // surfaces still receive specular_ibl as the fallback when rt reflections is unavailable
-    if (is_restir_pt_enabled())
+    else if (is_ray_traced_reflections_enabled())
     {
-        diffuse_ibl *= 0.0f;
-        if (surface.roughness >= 0.2f)
-        {
-            specular_ibl *= 0.0f;
-        }
+        specular_ibl *= 0.0f;
     }
 
     // transparents take full ibl, fresnel inside the split sum already governs the reflection split

data/shaders/restir_pt.hlsl

@@ -24,10 +24,9 @@ CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
 #include "restir_reservoir.hlsl"
 //==============================
 
-// upper bounds on the per pixel ris pool sizes, the actual counts come from cvars
-//   r.restir_pt_initial_candidates  -> brdf stream (cap below)
-//   r.restir_pt_light_candidates    -> nee stream (cap below)
-// caps are sized so the unrolled loop pressure is bounded but the user can tune for quality
+// upper bounds on the per pixel ris pool sizes, the live counts come from get_restir_*
+// helpers in restir_reservoir.hlsl which return paper-faithful hardcoded values (1 brdf
+// candidate, 8 nee candidates), the caps below only bound the unrolled loop size
 static const uint  INITIAL_CANDIDATE_SAMPLES_MAX   = 8;
 static const uint  LIGHT_RIS_CANDIDATE_SAMPLES_MAX = 64;
 static const float MIN_COS_AT_PRIMARY              = 1e-3f;
@@ -594,8 +593,8 @@ void trace_rc_suffix(
 }
 
 // gathers the rc vertex's nee + emission contribution (with rc bsdf baked in at src view dir,
-// view-dep bias bounded by RESTIR_RC_MIN_ROUGHNESS gate on rc reuse) and the suffix radiance
-// past rc (with rc bsdf factored out into rc_outgoing_dir for paper-faithful shift evaluation)
+// view-dep bias bounded by the get_restir_rc_min_roughness floor on rc reuse) and the suffix
+// radiance past rc (with rc bsdf factored out into rc_outgoing_dir for paper-faithful shifts)
 void accumulate_subpath_at_rc(
     PathPayload rc,
     float3 rc_view_dir,
@@ -846,10 +845,18 @@ PathSample trace_path_from_primary(
     L_nee  = max(L_nee,  0.0f);
     L_post = max(L_post, 0.0f);
 
-    // no source side firefly clamp here, the lin 2022 estimator is unbiased and the energy
-    // bound is provided by the m cap, the per sample W clamp, and the variance aware denoiser
-    // downstream, scaling the suffix luminance was a downward bias on bright reflections that
-    // also fought the denoiser's variance estimate
+    // soft luminance clamp, the W clamp + m cap bound the unbiased estimator energy but not
+    // single sample variance, when a glossy rc lands on a bright emitter the rc bsdf baked
+    // into rc_L_nee can spike to thousands of nits and a stable spatial neighbor will then
+    // stamp that pulse across many pixels (boiling), the soft scale preserves chromaticity so
+    // the residual downward bias is uniform across channels and the denoiser absorbs the rest
+    {
+        float lum_nee  = luminance(L_nee);
+        float lum_post = luminance(L_post);
+        const float firefly_ceiling = 250.0f;
+        if (lum_nee  > firefly_ceiling) L_nee  *= firefly_ceiling / lum_nee;
+        if (lum_post > firefly_ceiling) L_post *= firefly_ceiling / lum_post;
+    }
 
     s.rc_L_nee        = L_nee;
     s.rc_L_post       = L_post;

data/shaders/restir_pt_debug.hlsl

@@ -295,16 +295,27 @@ void main_cs(uint3 dispatch_id : SV_DispatchThreadID)
         history_weight   = saturate(min(history_weight, 0.992f));
     }
 
+    // small 3x3 mean+variance pre blur (schied 2017 §4.2) so isolated bright fireflies do not
+    // pin the variance estimate to a single pixel, this stabilizes the spatial filter weights
+    // and gives the firefly clamp below a reliable spatial sigma even when history is young
+    float3 mean_color, sigma_color, min_color, max_color;
+    compute_local_statistics(pixel, resolution, mean_color, sigma_color, min_color, max_color);
+    float spatial_sigma_luma = dot(sigma_color, luminance_weights);
+    float spatial_var_3x3    = spatial_sigma_luma * spatial_sigma_luma;
+
     // pre-ema firefly soft clamp, lin 2022 §6.4 / production svgf trick
     // if the current sample's luma is way above the history band, soft-rescale it down before
     // it poisons the moment update, otherwise a single hot pixel can pin the variance estimate
-    // for many frames and force the spatial filter to over-blur the neighborhood, the gate is
-    // applied only once temporal accumulation is established (n_eff > 4) so the bootstrap
-    // window stays unbiased
-    if (history_ok && history_moments.z > 4.0f)
+    // for many frames and force the spatial filter to over-blur the neighborhood, the gate
+    // fires as soon as we have a temporal partner so disocclusion fireflies do not boil for
+    // 4 frames before the clamp kicks in, the band widens via spatial sigma when history is
+    // young so the clamp does not over-bias on freshly accumulating pixels
+    if (history_ok)
     {
         float history_sigma = sqrt(max(history_moments.y - history_moments.x * history_moments.x, 0.0f));
-        float clamp_high    = history_moments.x + 8.0f * max(history_sigma, 0.05f);
+        float band_sigma    = max(history_sigma, spatial_sigma_luma);
+        float band_widen    = lerp(16.0f, 6.0f, saturate(history_moments.z / 4.0f));
+        float clamp_high    = history_moments.x + band_widen * max(band_sigma, 0.05f);
         if (current_luma > clamp_high && current_luma > 1e-3f)
         {
             float scale    = clamp_high / current_luma;
@@ -335,12 +346,6 @@ void main_cs(uint3 dispatch_id : SV_DispatchThreadID)
         variance_estimate = max(temporal_var, spatial_var * boost);
     }
 
-    // small 3x3 mean+variance pre blur (schied 2017 §4.2) so isolated bright fireflies do not
-    // pin the variance estimate to a single pixel, this stabilizes the spatial filter weights
-    float3 mean_color, sigma_color, min_color, max_color;
-    compute_local_statistics(pixel, resolution, mean_color, sigma_color, min_color, max_color);
-    float spatial_sigma_luma = dot(sigma_color, luminance_weights);
-    float spatial_var_3x3    = spatial_sigma_luma * spatial_sigma_luma;
     variance_estimate = max(variance_estimate, spatial_var_3x3 * 0.25f);
 
     // variance gated history clamp keeps chromatic stability even when the temporal accumulator

data/shaders/restir_pt_denoise_temporal.hlsl

@@ -302,9 +302,9 @@ void main_cs(uint3 dispatch_id : SV_DispatchThreadID)
     // lin 2022 §6.4 sample validation: every N frames a fixed subset of pixels re-traces the
     // chosen sample's primary->rc visibility ray, if rc is no longer reachable from the current
     // primary (light moved, geometry changed, occluder appeared) the reservoir is reset so we
-    // do not drag a stale path across many frames, the period is in cb_frame.restir_pt_validation_period
-    // and a value of 0 disables the pass, the pixel hash cycles deterministically with frame
-    // index so the cost is amortized to ~1/N pixels per frame
+    // do not drag a stale path across many frames, the period comes from get_restir_validation_period
+    // (hardcoded to 8 frames) and the pixel hash cycles deterministically with frame index so
+    // the cost is amortized to ~1/N pixels per frame
     uint validation_period = get_restir_validation_period();
     if (validation_period > 0u && combined.M > 0.0f && combined.W > 0.0f)
     {