n/b/worley: performance improvements

noise/basis/worley.cpp

@@ -23,92 +23,84 @@ using util::noise::basis::worley;
 
 ///////////////////////////////////////////////////////////////////////////////
 template <typename T>
-util::vector<2,T>
+util::point<2,T>
-generate (util::point<2,T> p, uint64_t seed)
+generate (util::point<2,intmax_t> p, uint64_t seed)
 {
     using util::hash::murmur2::mix;
 
     auto u = mix (seed, mix (uint64_t (p.x), uint64_t (p.y)));
     auto v = mix (u, seed);
 
-    auto r = util::vector<2,T> {
+    auto r = util::point<2,T> {
         (u & 0xffff) / T{0xffff},
         (v & 0xffff) / T{0xffff}
-    } * 2 - 1;
+    };
-
-    CHECK_GE (r, T{-1});
-    CHECK_LE (r, T{ 1});
 
+    CHECK_LIMIT (r, T{0}, T{1});
     return r;
 }
 
 
 ///////////////////////////////////////////////////////////////////////////////
-template <typename T>
+template <typename T, size_t F>
-worley<T>::worley (seed_t _seed):
+worley<T,F>::worley (seed_t _seed):
     seed (_seed)
 { ; }
 
 
 //-----------------------------------------------------------------------------
-template <typename T>
+template <typename T, size_t F>
-worley<T>::worley ():
+worley<T,F>::worley ():
-    seed (time (nullptr))
+    worley (time (nullptr))
 { ; }
 
 
 //-----------------------------------------------------------------------------
-template <typename T>
+template <typename T, size_t F>
 util::range<T>
-worley<T>::bounds (void) const
+worley<T,F>::bounds (void) const
 {
     return { 0.0, 1.5 };
 }
 
 
 //-----------------------------------------------------------------------------
-template <typename T>
+template <typename T, size_t F>
 T
-worley<T>::operator() (util::point<2,T> p) const
+worley<T,F>::operator() (util::point<2,T> p) const
 {
     // extract integer and fractional parts. be careful to always round down
     // (particularly with negatives) and avoid rounding errors.
     auto p_int = p.template cast<intmax_t> ();
     if (p.x < 0) p_int.x -= 1;
     if (p.y < 0) p_int.y -= 1;
-    auto p_rem = abs (p - p_int);
+    auto p_rem = abs (p - p_int).template as<point> ();
 
-    // +---+---+---+
+    // setup an array of distances
-    // | 0 | 1 | 2 |
+    static constexpr size_t RADIUS = 1;
-    // +---+---+---+
+    static constexpr size_t COUNT = pow2 (RADIUS * 2 + 1);
-    // | 3 | 4 | 5 |
+    T distances[COUNT] = { std::numeric_limits<T>::quiet_NaN () };
-    // +---+-------+
-    // | 6 | 7 | 8 |
-    // +---+---+---+
-
-    T distances[9] = { std::numeric_limits<T>::quiet_NaN () };
     T *cursor = distances;
 
-    for (signed y_off = -1; y_off <= 1 ; ++y_off)
+    // record the distances to each candidate point
-        for (signed x_off = -1; x_off <= 1; ++x_off) {
+    for (signed y_off = -signed(RADIUS); y_off <= signed(RADIUS) ; ++y_off) {
-            auto pos = vector<2,T> (T (x_off), T (y_off));
+        for (signed x_off = -signed(RADIUS); x_off <= signed(RADIUS); ++x_off) {
-            auto off = generate<T> (p_int + pos, this->seed);
+            auto off = vector<2,intmax_t> {x_off, y_off};
-            off += T{1};
+            auto pos = generate<T> (p_int + off, this->seed);
-            off /= T{2};
 
-            CHECK (off.x >= 0 && off.x <= 1);
+            CHECK_LIMIT (pos.x, T{0}, T{1});
-            CHECK (off.y >= 0 && off.y <= 1);
+            CHECK_LIMIT (pos.y, T{0}, T{1});
 
-            pos += off;
             *cursor = distance2 (pos + off, p_rem);
             cursor++;
         }
     }
 
-    std::sort (std::begin (distances), std::end (distances));
+    // find the f'th lowest value
-    CHECK_GE (distances[0], 0);
+    static_assert (F < COUNT, "worley order must be less than search radius");
-    CHECK (bounds ().contains (distances[0]));
+    std::partial_sort (distances, distances + F, distances + COUNT);
-    return distances[0];
+    CHECK_GE (distances[F], 0);
+    return distances[F];
 }
 
 

noise/basis/worley.hpp

@@ -23,7 +23,7 @@
 #include "../../range.hpp"
 
 namespace util { namespace noise { namespace basis {
-    template <typename T>
+    template <typename T, size_t F = 0>
     struct worley {
         worley (seed_t);
         worley ();