Rewrite SH9 core, more concise

5 files changed, 24 insertions, 35 deletions

diff --git a/3ner.cginc b/3ner.cginc
index cf03caf..0766380 100755
--- a/3ner.cginc
+++ b/3ner.cginc
@@ -40,7 +40,6 @@ v2f vert(appdata v) {
   o.objPos_orig = v.vertex;
 #endif
 
-
   [branch] if (instance_distance_culling()) {
     return (v2f) asfloat(-1);
   }
diff --git a/LUTS/dfg_standard.exr.meta b/LUTS/dfg_standard.exr.meta
index 107cdd2..4f4711f 100755
--- a/LUTS/dfg_standard.exr.meta
+++ b/LUTS/dfg_standard.exr.meta
@@ -22,7 +22,7 @@ TextureImporter:
     normalMapFilter: 0
     flipGreenChannel: 0
   isReadable: 0
-  streamingMipmaps: 0
+  streamingMipmaps: 1
   streamingMipmapsPriority: 0
   vTOnly: 0
   ignoreMipmapLimit: 0
diff --git a/interpolators.cginc b/interpolators.cginc
index 88a4649..a6a24e9 100755
--- a/interpolators.cginc
+++ b/interpolators.cginc
@@ -5,16 +5,17 @@
 
 struct appdata {
   float4 vertex   : POSITION;
-  float2 uv0      : TEXCOORD0;
-  float2 uv1      : TEXCOORD1;
-  float2 uv2      : TEXCOORD2;
-  float2 uv3      : TEXCOORD3;
 #if defined(_RAY_MARCHING_BAKED_ORIGINS)
   float4 color    : COLOR;  // vertex color
 #endif
   float3 normal   : NORMAL;
   float4 tangent  : TANGENT;
 
+  float2 uv0      : TEXCOORD0;
+  float2 uv1      : TEXCOORD1;
+  float2 uv2      : TEXCOORD2;
+  float2 uv3      : TEXCOORD3;
+
   UNITY_VERTEX_INPUT_INSTANCE_ID
 };
 
diff --git a/lighting.cginc b/lighting.cginc
index 2ae7d0d..9a66703 100755
--- a/lighting.cginc
+++ b/lighting.cginc
@@ -139,36 +139,25 @@ float3 yumSH9(float4 n, float3 worldPos, inout LightIndirect light) {
   //   unity_SHB*:     first four of the L2 coefficients
   //   unity_SHC:      last L2 coefficient
 
-  // Parse out coefficients into a simpler but less efficient format.
-  float3 L00  = float3(unity_SHAr.w, unity_SHAg.w, unity_SHAb.w);
-  float3 L1_1 = float3(unity_SHAr.x, unity_SHAg.x, unity_SHAb.x);
-  float3 L10  = float3(unity_SHAr.y, unity_SHAg.y, unity_SHAb.y);
-  float3 L11  = float3(unity_SHAr.z, unity_SHAg.z, unity_SHAb.z);
-  float3 L2_2 = float3(unity_SHBr.x, unity_SHBg.x, unity_SHBb.x);
-  float3 L2_1 = float3(unity_SHBr.y, unity_SHBg.y, unity_SHBb.y);
-  float3 L20  = float3(unity_SHBr.z, unity_SHBg.z, unity_SHBb.z);
-  float3 L21  = float3(unity_SHBr.w, unity_SHBg.w, unity_SHBb.w);
-  float3 L22  = unity_SHC;
-
   // Equation 13 from "An Efficient Representation for Irradiance Environment
-  // Maps" by Ramamoorthi and Hanrahan. Note that the order of some
-  // coefficients is different, and normalization constants have been
-  // premultiplied by Unity.
-  float3 L0 = L00;
-  float3 L1 = L1_1 * n.x + L10 * n.y + L11 * n.z;
-  float3 L2 =
-    L2_2 * n.x * n.y +
-    L2_1 * n.y * n.z +
-    L20  * n.z * n.z +
-    L21 * n.x * n.z +
-    L22 * (n.x * n.x - n.y * n.y);
+  // Maps" by Ramamoorthi and Hanrahan. Normalization constants have been
+  // premultiplied by Unity into the coefficient buffers.
+  //
+  // L0+L1: dot4 per channel (n.w=1 picks up the L0 term from SHA*.w)
+  // L2: four quadratic terms packed into vB via swizzle multiply, plus L22
+  float4 n4 = float4(n.xyz, 1.0);
+  float3 L0 = float3(unity_SHAr.w, unity_SHAg.w, unity_SHAb.w);
+  float3 L01 = float3(dot(unity_SHAr, n4), dot(unity_SHAg, n4), dot(unity_SHAb, n4));
+  float4 vB = n4.xyzz * n4.yzzx;
+  float3 L2 = float3(dot(unity_SHBr, vB), dot(unity_SHBg, vB), dot(unity_SHBb, vB))
+            + unity_SHC * (n.x * n.x - n.y * n.y);
 
   light.L00 = L0;
   light.L01r = unity_SHAr.xyz;
   light.L01g = unity_SHAg.xyz;
   light.L01b = unity_SHAb.xyz;
 
-  return L0 + L1 + L2;
+  return L01 + L2;
 }
 
 float4 getIndirectDiffuse(v2f i, Pbr pbr, inout LightIndirect light) {
diff --git a/pbr.cginc b/pbr.cginc
index a01c185..4389476 100755
--- a/pbr.cginc
+++ b/pbr.cginc
@@ -8,6 +8,12 @@
 #include "texture_utils.cginc"
 #include "impostor.cginc"
 
+// From filament: min roughness s.t. MIN_PERCEPTUAL_ROUGHNESS^4 > 0 in target
+// precision. Although we use fp32, the higher min roughness gives us a broader
+// specular highlight, which is preferable.
+#define MIN_PERCEPTUAL_ROUGHNESS 0.045
+#define MIN_ROUGHNESS            0.002025
+
 struct Pbr {
   float4 albedo;
   float3 normal;
@@ -33,12 +39,6 @@ struct Pbr {
 #endif
 };
 
-// From filament: min roughness s.t. MIN_PERCEPTUAL_ROUGHNESS^4 > 0 in target
-// precision. We use fp32. The smallest non-subnormal is 2^(-126). The 4th
-// root of that is ~3.29 * 10^-10.
-#define MIN_PERCEPTUAL_ROUGHNESS (3.3E-10)
-#define MIN_ROUGHNESS            (1.09E-19)
-
 #if defined(_PARALLAX_HEIGHTMAP)
 float2 parallax_offset(float2 uv, float3 view_dir_world, float3x3 tbn) {
   float3 view_dir_tangent = mul(tbn, view_dir_world);