precompute values in linear, bilinear sampling; make box muller a struct

Author: keptsecret

include/nbl/builtin/hlsl/sampling/bilinear.hlsl

@@ -28,13 +28,13 @@ struct Bilinear
     {
         Bilinear<T> retval;
         retval.bilinearCoeffs = bilinearCoeffs;
+        retval.twiceAreasUnderXCurve = vector2_type(bilinearCoeffs[0] + bilinearCoeffs[1], bilinearCoeffs[2] + bilinearCoeffs[3]);
         return retval;
     }
 
     vector2_type generate(NBL_REF_ARG(scalar_type) rcpPdf, NBL_CONST_REF_ARG(vector2_type) _u)
     {
         vector2_type u = _u;
-        const vector2_type twiceAreasUnderXCurve = vector2_type(bilinearCoeffs[0] + bilinearCoeffs[1], bilinearCoeffs[2] + bilinearCoeffs[3]);
         Linear<scalar_type> lineary = Linear<scalar_type>::create(twiceAreasUnderXCurve);
         u.y = lineary.generate(u.y);
 
@@ -52,7 +52,10 @@ struct Bilinear
         return 4.0 * nbl::hlsl::mix(nbl::hlsl::mix(bilinearCoeffs[0], bilinearCoeffs[1], u.x), nbl::hlsl::mix(bilinearCoeffs[2], bilinearCoeffs[3], u.x), u.y) / (bilinearCoeffs[0] + bilinearCoeffs[1] + bilinearCoeffs[2] + bilinearCoeffs[3]);
     }
 
-    vector4_type bilinearCoeffs;
+    // unit square: x0y0    x1y0
+    //              x0y1    x1y1
+    vector4_type bilinearCoeffs;    // (x0y0, x0y1, x1y0, x1y1)
+    vector2_type twiceAreasUnderXCurve;
 };
 
 }

include/nbl/builtin/hlsl/sampling/box_muller_transform.hlsl

@@ -2,8 +2,8 @@
 // This file is part of the "Nabla Engine".
 // For conditions of distribution and use, see copyright notice in nabla.h
 
-#ifndef _NBL_BUILTIN_HLSL_BOX_MULLER_TRANSFORM_INCLUDED_
-#define _NBL_BUILTIN_HLSL_BOX_MULLER_TRANSFORM_INCLUDED_
+#ifndef _NBL_BUILTIN_HLSL_SAMPLING_BOX_MULLER_TRANSFORM_INCLUDED_
+#define _NBL_BUILTIN_HLSL_SAMPLING_BOX_MULLER_TRANSFORM_INCLUDED_
 
 #include "nbl/builtin/hlsl/math/functions.hlsl"
 #include "nbl/builtin/hlsl/numbers.hlsl"
@@ -12,15 +12,26 @@ namespace nbl
 {
 namespace hlsl
 {
+namespace sampling
+{
 
-template<typename T>
-vector<T,2> boxMullerTransform(vector<T,2> xi, T stddev)
+template<typename T NBL_PRIMARY_REQUIRES(concepts::FloatingPointLikeScalar<T>)
+struct BoxMullerTransform
 {
-    T sinPhi, cosPhi;
-    math::sincos<T>(2.0 * numbers::pi<float> * xi.y - numbers::pi<float>, sinPhi, cosPhi);
-    return vector<T,2>(cosPhi, sinPhi) * nbl::hlsl::sqrt(-2.0 * nbl::hlsl::log(xi.x)) * stddev;
-}
+    using scalar_type = T;
+    using vector2_type = vector<T,2>;
+
+    vector2_type operator()(vector2_type xi)
+    {
+        scalar_type sinPhi, cosPhi;
+        math::sincos<scalar_type>(2.0 * numbers::pi<scalar_type> * xi.y - numbers::pi<scalar_type>, sinPhi, cosPhi);
+        return vector2_type(cosPhi, sinPhi) * nbl::hlsl::sqrt(-2.0 * nbl::hlsl::log(xi.x)) * stddev;
+    }
 
+    T stddev;
+};
+
+}
 }
 }
 

include/nbl/builtin/hlsl/sampling/linear.hlsl

@@ -21,21 +21,26 @@ struct Linear
     using scalar_type = T;
     using vector2_type = vector<T, 2>;
 
-    static Linear<T> create(NBL_CONST_REF_ARG(vector2_type) linearCoeffs)
+    static Linear<T> create(NBL_CONST_REF_ARG(vector2_type) linearCoeffs)   // start and end importance values (start, end)
     {
         Linear<T> retval;
-        retval.linearCoeffs = linearCoeffs;
+        retval.linearCoeffStart = linearCoeffs[0];
+        retval.rcpDiff = 1.0 / (linearCoeffs[0] - linearCoeffs[1]);
+        vector2_type squaredCoeffs = linearCoeffs * linearCoeffs;
+        retval.squaredCoeffStart = squaredCoeffs[0];
+        retval.squaredCoeffDiff = squaredCoeffs[1] - squaredCoeffs[0];
         return retval;
     }
 
     scalar_type generate(scalar_type u)
     {
-        const scalar_type rcpDiff = 1.0 / (linearCoeffs[0] - linearCoeffs[1]);
-        const vector2_type squaredCoeffs = linearCoeffs * linearCoeffs;
-        return nbl::hlsl::abs(rcpDiff) < numeric_limits<scalar_type>::max ? (linearCoeffs[0] - nbl::hlsl::sqrt(nbl::hlsl::mix(squaredCoeffs[0], squaredCoeffs[1], u))) * rcpDiff : u;
+        return hlsl::mix(u, (linearCoeffStart - hlsl::sqrt(squaredCoeffStart + u * squaredCoeffDiff)) * rcpDiff, hlsl::abs(rcpDiff) < numeric_limits<scalar_type>::max);
     }
 
-    vector2_type linearCoeffs;
+    scalar_type linearCoeffStart;  
+    scalar_type rcpDiff;
+    scalar_type squaredCoeffStart;
+    scalar_type squaredCoeffDiff;
 };
 
 }