Fix T91064: Cycles low poly meshes having black edges when shade smoothed

Fixes:{T91064}

Caused by {rBcd118c5581f482afc8554ff88b5b6f3b552b1682}

- Applies `ensure_valid_reflection()` to the normal input on all BSDFs for CPU and GPU.
- This doesn't affect hair.
- Removes `ensure_valid_reflection()` from the output of Bump Map and Normal Map nodes for CPU/GPU as it is not needed.
- The fix doesn't touch OSL.

Reviewed By: brecht, leesonw

Maniphest Tasks: T91064

Differential Revision: https://developer.blender.org/D12403

intern/cycles/kernel/closure/bsdf_principled_diffuse.h

@@ -35,21 +35,25 @@ static_assert(sizeof(ShaderClosure) >= sizeof(PrincipledDiffuseBsdf),
               "PrincipledDiffuseBsdf is too large!");
 
 ccl_device float3 calculate_principled_diffuse_brdf(
-    const PrincipledDiffuseBsdf *bsdf, float3 N, float3 V, float3 L, float3 H, float *pdf)
+    const PrincipledDiffuseBsdf *bsdf, float3 N, float3 V, float3 L, float *pdf)
 {
   float NdotL = dot(N, L);
-  float NdotV = dot(N, V);
 
-  if (NdotL <= 0 || NdotV <= 0) {
-    *pdf = 0.0f;
+  if (NdotL <= 0) {
     return make_float3(0.0f, 0.0f, 0.0f);
   }
 
-  float LdotH = dot(L, H);
+  float NdotV = dot(N, V);
+
+  /* H = normalize(L + V);  // Bissector of an angle between L and V
+   * LH2 = 2 * dot(L, H)^2 = 2cos(x)^2 = cos(2x) + 1 = dot(L, V) + 1,
+   * half-angle x between L and V is at most 90 deg
+   */
+  float LH2 = dot(L, V) + 1;
 
   float FL = schlick_fresnel(NdotL), FV = schlick_fresnel(NdotV);
-  const float Fd90 = 0.5f + 2.0f * LdotH * LdotH * bsdf->roughness;
-  float Fd = (1.0f * (1.0f - FL) + Fd90 * FL) * (1.0f * (1.0f - FV) + Fd90 * FV);
+  const float Fd90 = 0.5f + LH2 * bsdf->roughness;
+  float Fd = (1.0f - FL + Fd90 * FL) * (1.0f - FV + Fd90 * FV);
 
   float value = M_1_PI_F * NdotL * Fd;
 
@@ -72,11 +76,10 @@ ccl_device float3 bsdf_principled_diffuse_eval_reflect(const ShaderClosure *sc,
   float3 N = bsdf->N;
   float3 V = I;         // outgoing
   float3 L = omega_in;  // incoming
-  float3 H = normalize(L + V);
 
   if (dot(N, omega_in) > 0.0f) {
     *pdf = fmaxf(dot(N, omega_in), 0.0f) * M_1_PI_F;
-    return calculate_principled_diffuse_brdf(bsdf, N, V, L, H, pdf);
+    return calculate_principled_diffuse_brdf(bsdf, N, V, L, pdf);
   }
   else {
     *pdf = 0.0f;
@@ -112,9 +115,7 @@ ccl_device int bsdf_principled_diffuse_sample(const ShaderClosure *sc,
   sample_cos_hemisphere(N, randu, randv, omega_in, pdf);
 
   if (dot(Ng, *omega_in) > 0) {
-    float3 H = normalize(I + *omega_in);
-
-    *eval = calculate_principled_diffuse_brdf(bsdf, N, I, *omega_in, H, pdf);
+    *eval = calculate_principled_diffuse_brdf(bsdf, N, I, *omega_in, pdf);
 
 #ifdef __RAY_DIFFERENTIALS__
     // TODO: find a better approximation for the diffuse bounce

intern/cycles/kernel/svm/svm_closure.h

@@ -85,6 +85,9 @@ ccl_device_noinline int svm_node_closure_bsdf(
   }
 
   float3 N = stack_valid(data_node.x) ? stack_load_float3(stack, data_node.x) : sd->N;
+  if (!(sd->type & PRIMITIVE_ALL_CURVE)) {
+    N = ensure_valid_reflection(sd->Ng, sd->I, N);
+  }
 
   float param1 = (stack_valid(param1_offset)) ? stack_load_float(stack, param1_offset) :
                                                 __uint_as_float(node.z);
@@ -166,6 +169,9 @@ ccl_device_noinline int svm_node_closure_bsdf(
       float3 clearcoat_normal = stack_valid(data_cn_ssr.x) ?
                                     stack_load_float3(stack, data_cn_ssr.x) :
                                     sd->N;
+      if (!(sd->type & PRIMITIVE_ALL_CURVE)) {
+        clearcoat_normal = ensure_valid_reflection(sd->Ng, sd->I, clearcoat_normal);
+      }
       float3 subsurface_radius = stack_valid(data_cn_ssr.y) ?
                                      stack_load_float3(stack, data_cn_ssr.y) :
                                      make_float3(1.0f, 1.0f, 1.0f);

intern/cycles/kernel/svm/svm_tex_coord.h

@@ -347,8 +347,6 @@ ccl_device_noinline void svm_node_normal_map(const KernelGlobals *kg,
     N = safe_normalize(sd->N + (N - sd->N) * strength);
   }
 
-  N = ensure_valid_reflection(sd->Ng, sd->I, N);
-
   if (is_zero(N)) {
     N = sd->N;
   }