some more optimizations (elalish#503)

* simple speedup from stl algorithms

* fix build

* std::list allocator optimization

* added documentation

* fix weird formatting issues

* try fixing MacOS build

it seems that Apple Clang 14 does not support pstl

* fix again

* fix macos build

* fix apple

* clean up using macro

* fix apple?

* small fix...

* fix apple (again...)

Author: pca006132

src/collider/src/collider.cpp

@@ -287,7 +287,8 @@ SparseIndices Collider::Collisions(const VecDH<T>& queriesIn) const {
                  {thrust::pair<int*, int*>(nullptr, nullptr), counts.ptrD(),
                   nodeBBox_.ptrD(), internalChildren_.ptrD()}));
   // compute start index for each query and total count
-  exclusive_scan(policy, counts.begin(), counts.end(), counts.begin());
+  exclusive_scan(policy, counts.begin(), counts.end(), counts.begin(), 0,
+                 std::plus<int>());
   SparseIndices queryTri(counts.back());
   // actually recording collisions
   for_each_n(policy, zip(queriesIn.cbegin(), countAt(0)), queriesIn.size(),

src/polygon/src/polygon.cpp

@@ -15,8 +15,14 @@
 #include "polygon.h"
 
 #include <algorithm>
+#if MANIFOLD_PAR == 'T'
+#include <execution>
+#endif
 #include <list>
 #include <map>
+#if !__APPLE__
+#include <memory_resource>
+#endif
 #include <queue>
 #include <set>
 #include <stack>
@@ -303,14 +309,25 @@ class Monotones {
 
  private:
   struct VertAdj;
-  typedef std::list<VertAdj>::iterator VertItr;
   struct EdgePair;
-  typedef std::list<EdgePair>::iterator PairItr;
   enum VertType { Start, WestSide, EastSide, Merge, End, Skip };
+#if __APPLE__
+  typedef std::list<VertAdj>::iterator VertItr;
+  typedef std::list<EdgePair>::iterator PairItr;
 
-  std::list<VertAdj> monotones_;     // sweep-line list of verts
-  std::list<EdgePair> activePairs_;  // west to east list of monotone edge pairs
+  std::list<VertAdj> monotones_;       // sweep-line list of verts
+  std::list<EdgePair> activePairs_;    // west to east monotone edges
   std::list<EdgePair> inactivePairs_;  // completed monotones
+#else
+  typedef std::pmr::list<VertAdj>::iterator VertItr;
+  typedef std::pmr::list<EdgePair>::iterator PairItr;
+
+  std::pmr::monotonic_buffer_resource mbr;
+  std::pmr::polymorphic_allocator<int> pa{&mbr};
+  std::pmr::list<VertAdj> monotones_{pa};       // sweep-line list of verts
+  std::pmr::list<EdgePair> activePairs_{pa};    // west to east monotone edges
+  std::pmr::list<EdgePair> inactivePairs_{pa};  // completed monotones
+#endif
   float precision_;  // a triangle of this height or less is degenerate
 
   /**
@@ -392,7 +409,7 @@ class Monotones {
   class Triangulator {
    public:
     Triangulator(VertItr vert, float precision) : precision_(precision) {
-      reflex_chain_.push(vert);
+      reflex_chain_.push_back(vert);
       other_side_ = vert;
     }
     int NumTriangles() const { return triangles_output_; }
@@ -407,14 +424,14 @@ class Monotones {
      */
     void ProcessVert(const VertItr vi, bool onRight, bool last,
                      std::vector<glm::ivec3> &triangles) {
-      VertItr v_top = reflex_chain_.top();
+      VertItr v_top = reflex_chain_.back();
       if (reflex_chain_.size() < 2) {
-        reflex_chain_.push(vi);
+        reflex_chain_.push_back(vi);
         onRight_ = onRight;
         return;
       }
-      reflex_chain_.pop();
-      VertItr vj = reflex_chain_.top();
+      reflex_chain_.pop_back();
+      VertItr vj = reflex_chain_.back();
       if (onRight_ == onRight && !last) {
         // This only creates enough triangles to ensure the reflex chain is
         // still reflex.
@@ -423,13 +440,13 @@ class Monotones {
         while (ccw == (onRight_ ? 1 : -1) || ccw == 0) {
           AddTriangle(triangles, vi, vj, v_top);
           v_top = vj;
-          reflex_chain_.pop();
+          reflex_chain_.pop_back();
           if (reflex_chain_.empty()) break;
-          vj = reflex_chain_.top();
+          vj = reflex_chain_.back();
           ccw = CCW(vi->pos, vj->pos, v_top->pos, precision_);
         }
-        reflex_chain_.push(v_top);
-        reflex_chain_.push(vi);
+        reflex_chain_.push_back(v_top);
+        reflex_chain_.push_back(vi);
       } else {
         // This branch empties the reflex chain and switches sides. It must be
         // used for the last point, as it will output all the triangles
@@ -438,19 +455,19 @@ class Monotones {
         onRight_ = !onRight_;
         VertItr v_last = v_top;
         while (!reflex_chain_.empty()) {
-          vj = reflex_chain_.top();
+          vj = reflex_chain_.back();
           AddTriangle(triangles, vi, v_last, vj);
           v_last = vj;
-          reflex_chain_.pop();
+          reflex_chain_.pop_back();
         }
-        reflex_chain_.push(v_top);
-        reflex_chain_.push(vi);
+        reflex_chain_.push_back(v_top);
+        reflex_chain_.push_back(vi);
         other_side_ = v_top;
       }
     }
 
    private:
-    std::stack<VertItr> reflex_chain_;
+    std::vector<VertItr> reflex_chain_;
     VertItr other_side_;  // The end vertex across from the reflex chain
     bool onRight_;        // The side the reflex chain is on
     int triangles_output_ = 0;
@@ -784,7 +801,11 @@ class Monotones {
         starts.push_back(v);
       }
     }
+#if MANIFOLD_PAR == 'T' && !(__APPLE__)
+    std::sort(std::execution::par_unseq, starts.begin(), starts.end(), cmp);
+#else
     std::sort(starts.begin(), starts.end(), cmp);
+#endif
 
     std::vector<VertItr> skipped;
     VertItr insertAt = monotones_.begin();

src/utilities/include/par.h

@@ -30,6 +30,9 @@
 #define MANIFOLD_PAR_NS omp
 #elif MANIFOLD_PAR == 'T'
 #include <thrust/system/tbb/execution_policy.h>
+
+#include <algorithm>
+#include <execution>
 #define MANIFOLD_PAR_NS tbb
 #else
 #define MANIFOLD_PAR_NS cpp
@@ -136,43 +139,105 @@ inline ExecutionPolicy autoPolicy(int size) {
     thrust::NAME(thrust::cpp::par, args...);                 \
   }
 
+#if MANIFOLD_PAR == 'T' && !(__APPLE__)
+// sometimes stl variant is faster
+#define STL_DYNAMIC_BACKEND(NAME, RET)                        \
+  template <typename Ret = RET, typename... Args>             \
+  Ret NAME(ExecutionPolicy policy, Args... args) {            \
+    switch (policy) {                                         \
+      case ExecutionPolicy::ParUnseq:                         \
+      case ExecutionPolicy::Par:                              \
+        return std::NAME(std::execution::par_unseq, args...); \
+      case ExecutionPolicy::Seq:                              \
+        break;                                                \
+    }                                                         \
+    return std::NAME(args...);                                \
+  }
+#define STL_DYNAMIC_BACKEND_VOID(NAME)                 \
+  template <typename... Args>                          \
+  void NAME(ExecutionPolicy policy, Args... args) {    \
+    switch (policy) {                                  \
+      case ExecutionPolicy::ParUnseq:                  \
+      case ExecutionPolicy::Par:                       \
+        std::NAME(std::execution::par_unseq, args...); \
+        break;                                         \
+      case ExecutionPolicy::Seq:                       \
+        std::NAME(args...);                            \
+        break;                                         \
+    }                                                  \
+  }
+
+template <typename... Args>
+void exclusive_scan(ExecutionPolicy policy, Args... args) {
+  // https://github.com/llvm/llvm-project/issues/59810
+  std::exclusive_scan(args...);
+}
+template <typename DerivedPolicy, typename InputIterator1,
+          typename InputIterator2, typename OutputIterator, typename Predicate>
+OutputIterator copy_if(ExecutionPolicy policy, InputIterator1 first,
+                       InputIterator1 last, InputIterator2 stencil,
+                       OutputIterator result, Predicate pred) {
+  if (policy == ExecutionPolicy::Seq)
+    return thrust::copy_if(thrust::cpp::par, first, last, stencil, result,
+                           pred);
+  else
+    // note: this is not a typo, see
+    // https://github.com/NVIDIA/thrust/issues/1977
+    return thrust::copy_if(first, last, stencil, result, pred);
+}
+template <typename DerivedPolicy, typename InputIterator1,
+          typename OutputIterator, typename Predicate>
+OutputIterator copy_if(ExecutionPolicy policy, InputIterator1 first,
+                       InputIterator1 last, OutputIterator result,
+                       Predicate pred) {
+  if (policy == ExecutionPolicy::Seq)
+    return std::copy_if(first, last, result, pred);
+  else
+    return std::copy_if(std::execution::par_unseq, first, last, result, pred);
+}
+#else
+#define STL_DYNAMIC_BACKEND(NAME, RET) THRUST_DYNAMIC_BACKEND(NAME, RET)
+#define STL_DYNAMIC_BACKEND_VOID(NAME) THRUST_DYNAMIC_BACKEND_VOID(NAME)
+
+THRUST_DYNAMIC_BACKEND_VOID(exclusive_scan)
+THRUST_DYNAMIC_BACKEND(copy_if, void)
+#endif
+
 THRUST_DYNAMIC_BACKEND_HOST_VOID(for_each)
 THRUST_DYNAMIC_BACKEND_HOST_VOID(for_each_n)
 
 THRUST_DYNAMIC_BACKEND_VOID(gather)
 THRUST_DYNAMIC_BACKEND_VOID(scatter)
 THRUST_DYNAMIC_BACKEND_VOID(for_each)
 THRUST_DYNAMIC_BACKEND_VOID(for_each_n)
-THRUST_DYNAMIC_BACKEND_VOID(sort)
-THRUST_DYNAMIC_BACKEND_VOID(stable_sort)
-THRUST_DYNAMIC_BACKEND_VOID(fill)
 THRUST_DYNAMIC_BACKEND_VOID(sequence)
 THRUST_DYNAMIC_BACKEND_VOID(sort_by_key)
 THRUST_DYNAMIC_BACKEND_VOID(stable_sort_by_key)
-THRUST_DYNAMIC_BACKEND_VOID(copy)
 THRUST_DYNAMIC_BACKEND_VOID(transform)
-THRUST_DYNAMIC_BACKEND_VOID(inclusive_scan)
-THRUST_DYNAMIC_BACKEND_VOID(exclusive_scan)
 THRUST_DYNAMIC_BACKEND_VOID(uninitialized_fill)
 THRUST_DYNAMIC_BACKEND_VOID(uninitialized_copy)
+THRUST_DYNAMIC_BACKEND_VOID(stable_sort)
+THRUST_DYNAMIC_BACKEND_VOID(fill)
+THRUST_DYNAMIC_BACKEND_VOID(copy)
+THRUST_DYNAMIC_BACKEND_VOID(inclusive_scan)
 THRUST_DYNAMIC_BACKEND_VOID(copy_n)
+STL_DYNAMIC_BACKEND_VOID(sort)
 
-THRUST_DYNAMIC_BACKEND(all_of, bool)
-THRUST_DYNAMIC_BACKEND(is_sorted, bool)
-THRUST_DYNAMIC_BACKEND(reduce, void)
-THRUST_DYNAMIC_BACKEND(count_if, int)
-THRUST_DYNAMIC_BACKEND(binary_search, bool)
 // void implies that the user have to specify the return type in the template
 // argument, as we are unable to deduce it
+THRUST_DYNAMIC_BACKEND(transform_reduce, void)
+THRUST_DYNAMIC_BACKEND(gather_if, void)
+THRUST_DYNAMIC_BACKEND(reduce_by_key, void)
+THRUST_DYNAMIC_BACKEND(lower_bound, void)
 THRUST_DYNAMIC_BACKEND(remove, void)
-THRUST_DYNAMIC_BACKEND(copy_if, void)
-THRUST_DYNAMIC_BACKEND(remove_if, void)
-THRUST_DYNAMIC_BACKEND(unique, void)
 THRUST_DYNAMIC_BACKEND(find, void)
 THRUST_DYNAMIC_BACKEND(find_if, void)
-THRUST_DYNAMIC_BACKEND(reduce_by_key, void)
-THRUST_DYNAMIC_BACKEND(transform_reduce, void)
-THRUST_DYNAMIC_BACKEND(lower_bound, void)
-THRUST_DYNAMIC_BACKEND(gather_if, void)
+THRUST_DYNAMIC_BACKEND(all_of, bool)
+THRUST_DYNAMIC_BACKEND(is_sorted, bool)
+THRUST_DYNAMIC_BACKEND(reduce, void)
+THRUST_DYNAMIC_BACKEND(count_if, int)
+THRUST_DYNAMIC_BACKEND(binary_search, bool)
+STL_DYNAMIC_BACKEND(remove_if, void)
+STL_DYNAMIC_BACKEND(unique, void)
 
 }  // namespace manifold