ngsolve 6.2.2505__cp39-cp39-macosx_10_15_universal2.whl → 6.2.2505.post94.dev0__cp39-cp39-macosx_10_15_universal2.whl

netgen/include/ngblas.hpp

@@ -584,12 +584,31 @@ namespace ngbla
       }
   }
 
-  template <typename TM, typename TVX, typename TVY>
-  extern void TestFunc (TM m, TVX x, TVY y);
-
+  // template <typename TM, typename TVX, typename TVY>
+  // extern void TestFunc (TM m, TVX x, TVY y);
+  
+  
+  template <typename TS, typename T> constexpr bool IsVec = false;
+  template <typename TS, int S> constexpr bool IsVec<TS, Vec<S,TS>> = true;  
+  
   template <bool ADD, bool POS, typename TM, ORDERING ORD, typename TX, typename TY>
   INLINE void NgGEMV (BareSliceMatrix<TM,ORD> a, FlatVector<const TX> x, FlatVector<TY> y)
   {
+    if constexpr (std::is_same<TM,double>() && std::is_same<TX,TY>() && IsVec<Complex,TX>)
+      {
+        FlatMatrix<double> mx(x.Size(), sizeof(TX)/sizeof(double), (double*)(void*)x.Addr(0));
+        FlatMatrix<double> my(y.Size(), sizeof(TX)/sizeof(double), (double*)(void*)y.Addr(0));
+        NgGEMM<ADD,POS> (a.AddSize(y.Size(), x.Size()),make_SliceMatrix(mx), make_SliceMatrix(my));
+        return;
+      }
+    if constexpr (std::is_same<TM,Complex>() && std::is_same<TX,TY>() && IsVec<Complex,TX>)
+      {
+        FlatMatrix<Complex> mx(x.Size(), sizeof(TX)/sizeof(Complex), &const_cast<Complex&>(*(x.Data()->Data())));
+        FlatMatrix<Complex> my(y.Size(), sizeof(TX)/sizeof(Complex), y.Data()->Data());
+        NgGEMM<ADD,POS> (a.AddSize(y.Size(), x.Size()),make_SliceMatrix(mx), make_SliceMatrix(my));
+        return;
+      }
+
     if (!ADD)
       {
         if (!POS)
@@ -635,6 +654,16 @@ namespace ngbla
   extern NGS_DLL_HEADER  
   void NgGEMV (double s, BareSliceMatrix<double,ord> a, SliceVector<double> x, SliceVector<double> y) NETGEN_NOEXCEPT;
   */
+
+
+
+
+
+
+
+  /* *********************** GEMV - SliceVector **************************** */
+
+
   
   template <bool ADD, ORDERING ord>
   extern NGS_DLL_HEADER  
@@ -714,7 +743,30 @@ namespace ngbla
   }
 
 
-
+  
+  
+  template <bool ADD, bool POS, typename TM, ORDERING ORD, typename TX, typename TY>
+  INLINE void NgGEMV (BareSliceMatrix<TM,ORD> a, SliceVector<TX> x, SliceVector<TY> y)
+  {
+    if constexpr (std::is_same<TM,double>() && std::is_same<TX,TY>() && IsVec<Complex,TX>)
+      {
+        constexpr int VS = sizeof(TX)/sizeof(double);
+        SliceMatrix<double> mx(x.Size(), VS, x.Dist()*VS, (double*)(void*)x.Addr(0));
+        SliceMatrix<double> my(y.Size(), VS, y.Dist()*VS, (double*)(void*)y.Addr(0));
+        NgGEMM<ADD,POS> (a.AddSize(y.Size(), x.Size()),make_SliceMatrix(mx), make_SliceMatrix(my));
+        return;
+      }
+    else if constexpr (std::is_same<TM,Complex>() && std::is_same<TX,TY>() && IsVec<Complex,TX>)
+      {
+        constexpr int VS = sizeof(TX)/sizeof(Complex);
+        SliceMatrix<Complex> mx(x.Size(), VS, x.Dist()*VS, &const_cast<Complex&>(*(x.Data()->Data())));
+        SliceMatrix<Complex> my(y.Size(), VS, y.Dist()*VS, y.Data()->Data());
+        NgGEMM<ADD,POS> (a.AddSize(y.Size(), x.Size()),make_SliceMatrix(mx), make_SliceMatrix(my));
+        return;
+      }
+    else
+      NgGEMV<ADD> (POS ? 1.0 : -1.0, a, x, y);
+  }
 
 
 
@@ -880,9 +932,14 @@ namespace ngbla
                          FlatVector<const TB>(prod.View().B().Range(0,w)),
                          FlatVector<T>(self.Spec().Range(0,h)));
       else
+        NgGEMV<ADD,POS> (make_BareSliceMatrix(prod.View().A()),
+                         SliceVector<TB>(prod.View().B().Range(0,w)),
+                         SliceVector<T>(self.Spec().Range(0,h)));
+        /*
         NgGEMV<ADD> (POS ? 1.0 : -1.0, make_BareSliceMatrix(prod.View().A()),
                      SliceVector<TB>(prod.View().B().Range(0,w)),
                      SliceVector<T>(self.Spec().Range(0,h)));
+        */
       return self.Spec();
     }
   };
@@ -1099,8 +1156,60 @@ namespace ngbla
     }
   };
   
+  // typedef void (*pmatmatcRR)(size_t, size_t, BareSliceMatrix<Complex,RowMajor>, BareSliceMatrix<Complex,RowMajor>,BareSliceMatrix<Complex,RowMajor>);
+
+  template <ORDERING OA, ORDERING OB>
+  using pmatmatc =  void (*)(size_t, size_t, BareSliceMatrix<Complex, OA>, BareSliceMatrix<Complex,OB>,BareSliceMatrix<Complex,RowMajor>);
+  template <bool ADD, bool POS, ORDERING OA, ORDERING OB>
+  extern NGS_DLL_HEADER pmatmatc<OA,OB> dispatch_matmatc[9];
+
+  template <bool ADD, bool POS, ORDERING OA, ORDERING OB>
+  extern NGS_DLL_HEADER void NgGEMMBare (size_t ah, size_t aw, size_t bw, BareSliceMatrix<Complex,OA> a, BareSliceMatrix<Complex,OB> b, BareSliceMatrix<Complex,RowMajor> c);
 
 
+  template <bool ADD, bool POS, ORDERING OA, ORDERING OB>
+  void NgGEMM (SliceMatrix<Complex,OA> a, SliceMatrix<Complex,OB> b, SliceMatrix<Complex,RowMajor> c)
+  {
+    size_t ah = a.Height();
+    size_t aw = a.Width();
+    size_t bw = b.Width();
+    if (aw < std::size(dispatch_matmatc<ADD,POS,OA,OB>))
+      {
+        (*dispatch_matmatc<ADD,POS,OA,OB>[aw])(ah, bw, make_BareSliceMatrix(a), make_BareSliceMatrix(b), make_BareSliceMatrix(c));
+        return;
+      }
+
+    NgGEMMBare<ADD,POS>(ah, aw, bw, make_BareSliceMatrix(a), make_BareSliceMatrix(b), make_BareSliceMatrix(c));
+  }
+  
+  template <bool ADD, bool POS, ORDERING OA, ORDERING OB>
+  void NgGEMM (SliceMatrix<Complex,OA> a, SliceMatrix<Complex,OB> b, SliceMatrix<Complex,ColMajor> c)
+  {
+    NgGEMM<ADD,POS> (Trans(b), Trans(a), Trans(c));
+  }
+  
+  template <typename OP, typename T, typename TA, typename TB>
+  class assign_trait<OP, T, MultExpr<TA, TB>,
+                     enable_if_t<IsConvertibleToSliceMatrix<TA,Complex>() &&
+                                 IsConvertibleToSliceMatrix<TB,Complex>() &&
+                                 IsConvertibleToSliceMatrix<T,Complex>(), int>>
+  {
+  public:
+    static inline T & Assign (MatExpr<T> & self, const Expr<MultExpr<TA, TB>> & prod) 
+    {
+      constexpr bool ADD = std::is_same<OP,typename MatExpr<T>::AsAdd>::value || std::is_same<OP,typename MatExpr<T>::AsSub>::value;
+      constexpr bool POS = std::is_same<OP,typename MatExpr<T>::As>::value || std::is_same<OP,typename MatExpr<T>::AsAdd>::value;
+
+      size_t n = CombinedSize(prod.View().A().Height(), self.Spec().Height());
+      size_t m = CombinedSize(prod.View().B().Width(), self.Spec().Width());
+      size_t k = CombinedSize(prod.View().A().Width(), prod.View().B().Height());
+      
+      NgGEMM<ADD,POS> (make_BareSliceMatrix(prod.View().A()).AddSize(n,k).RemoveConst(),
+                       make_BareSliceMatrix(prod.View().B()).AddSize(k,m).RemoveConst(),
+                       make_BareSliceMatrix(self.Spec()).AddSize(n,m));
+      return self.Spec();
+    }
+  };
 
   
 

netgen/include/recursive_pol.hpp

@@ -1507,9 +1507,24 @@ namespace ngfem
   class NormalizedLegendreFunctions : public RecursivePolynomialNonStatic<NormalizedLegendreFunctions>
   {
     int  m;
+    static Matrix<double,ColMajor> matA, matB;
   public:
     NormalizedLegendreFunctions (size_t _m) : m(_m) { }
     
+
+    static void Calc(int mmax)
+    {
+      if (matA.Height() >= mmax+1) return;
+      matA.SetSize(mmax+1, mmax+1);
+      matB.SetSize(mmax+1, mmax+1);
+      for (int m = 0; m <= mmax; m++)
+        for (int n = m+2; n <= mmax; n++)
+          {
+            matA(n,m) = (2*n-1) / sqrt((n-m+0.0)*(n+m));
+            matB(n,m) = sqrt((n+m-1.0)*(n-m-1.0)) / sqrt((n-m+0.0)*(n+m));
+          }
+    }
+    
     template <typename T, typename S>
     NormalizedLegendreFunctions (size_t mmax, size_t nmax, T x, S && values)
     {
@@ -1524,23 +1539,60 @@ namespace ngfem
       auto y = Trans(values);
       
       y = 0.0;
-      double u = -sqrt((1-x)*(1+x));
+      T u = -sqrt((1-x)*(1+x));
       y(0,0)=1;
-      
-      for (double m = 0; m <= mmax; m++)
+
+      if (mmax+1 > matA.Height() || nmax > matA.Width())
+        {
+          for (double m = 0; m <= mmax; m++)
+            {
+              if (m > 0)
+                y(m,m)=y(m-1,m-1)*u*sqrt((2*m-1.0)/(2*m));
+              if (m < nmax)
+                y(m+1,m)=x*y(m,m)*sqrt(2*m+1.0);
+              for (int n = m+2; n <= nmax; n++)
+                y(n,m)=((2*n-1)*x*y(n-1,m) - 
+                        sqrt((n+m-1.0)*(n-m-1.0))*y(n-2,m))
+                  /sqrt((n-m+0.0)*(n+m));
+            }
+        }
+      else
         {
-          if (m > 0)
-            y(m,m)=y(m-1,m-1)*u*sqrt((2*m-1.0)/(2*m));
-          if (m < nmax)
-            y(m+1,m)=x*y(m,m)*sqrt(2*m+1.0);
-          for (int n = m+2; n <= nmax; n++)
-            y(n,m)=((2*n-1)*x*y(n-1,m) - 
-                    sqrt((n+m-1.0)*(n-m-1.0))*y(n-2,m))
-              /sqrt((n-m+0.0)*(n+m));
+          for (double m = 0; m <= mmax; m++)
+            {
+              if (m > 0)
+                y(m,m)=y(m-1,m-1)*u*sqrt((2*m-1.0)/(2*m));
+              if (m < nmax)
+                {
+                  T valold = y(m,m);
+                  
+                  T val = x*valold*sqrt(2*m+1.0);
+                  y(m+1,m)=val;
+                  
+                  auto coefsA = matA.Col(m);
+                  auto coefsB = matB.Col(m);
+                  
+                  for (int n = m+2; n <= nmax; n++)
+                    {
+                      T valnew = coefsA(n) * x * val - coefsB(n) * valold;
+                      y(n,m) = valnew;
+                      valold = val;
+                      val = valnew;
+                    }
+                }
+              /*
+              for (int n = m+2; n <= nmax; n++)
+                y(n,m) = matA(n,m) * x * y(n-1,m) - matB(n,m) * y(n-2,m);
+              */
+            }
         }
+      
       for (int n = 0; n <= nmax; n++)
+        y.Row(n).Range(0,n+1) *= sqrt(2*n+1.0);
+        /*
         for (int m = 0; m <= n; m++)
           y(n,m)=y(n,m)*sqrt(2*n+1.0);
+        */
     }
     
     template <class S>

netgen/include/simd_complex.hpp

@@ -56,6 +56,17 @@ namespace ngcore
     auto & imag() { return im; }
 
 
+    auto Lo() const
+    {
+      if constexpr (N == 2) return Complex(re.Lo(), im.Lo());
+      else return SIMD<Complex,N/2> (re.Lo(), im.Lo());
+    }
+    auto Hi() const
+    {
+      if constexpr (N == 2) return Complex(re.Hi(), im.Hi());
+      else return SIMD<Complex,N/2> (re.Hi(), im.Hi());
+    }
+    
     // Numbers in SIMD structure are not necessarily in same order as in memory
     // for instance:
     // [x0,y0,x1,y1,x2,y2,x3,y3] -> [x0,x2,x1,x3,y0,y2,y1,y3]
@@ -266,6 +277,15 @@ namespace ngcore
     return SIMD<Complex, N> (IfPos (a.real(), b.real(), c.real()),
                              IfPos (a.real(), b.imag(), c.imag()));
   }
+
+
+  template <int N>
+  INLINE SIMD<Complex, N> If (SIMD<mask64, N> a, SIMD<Complex, N> b, SIMD<Complex, N> c)
+  {
+    return SIMD<Complex, N> (If (a, b.real(), c.real()),
+                             If (a, b.imag(), c.imag()));
+  }
+
 }
 
 

netgen/include/sparsematrix_dyn.hpp

@@ -50,9 +50,9 @@ namespace ngla
     virtual void MultAdd (double s, const BaseVector & x, BaseVector & y) const override;
     
     AutoVector CreateRowVector() const override
-    { throw make_unique<VVector<TSCAL>> (VWidth()); }
+    { return make_unique<VVector<TSCAL>> (VWidth()); }
     AutoVector CreateColVector() const override
-    { throw make_unique<VVector<TSCAL>> (VHeight()); }
+    { return make_unique<VVector<TSCAL>> (VHeight()); }
     
     virtual tuple<int,int> EntrySizes() const override { return { bh, bw }; }
   };

netgen/include/sparsematrix_impl.hpp

@@ -475,6 +475,31 @@ namespace ngla
           }
       }
     return this->CreateFromCOO(indi, indj, val, this->Height(), this->Width());
+
+    /*
+      // needs parallelization and testing
+    Array<int> cnt(this->Height());
+    cnt = 0;
+    size_t cnt0 = 0;
+    for (auto i : Range(this->Height()))
+      for (auto j : Range(firsti[i], firsti[i+1]))
+        if (ngbla::L2Norm2(data[j]) > tol*tol)
+          cnt[i]++;
+        else
+          cnt0++;
+
+    cout << "zero-els = " << cnt0 << endl;
+
+
+    auto matrix = make_shared<SparseMatrix<TM>> (cnt, this->Width());
+
+    for (auto i : Range(this->Height()))
+      for (auto j : Range(firsti[i], firsti[i+1]))
+        if (ngbla::L2Norm2(data[j]) > tol*tol)
+          (*matrix)(i, colnr[i]) =  data[j];
+
+    return matrix;
+    */
   }
 
   template <class TM>

netgen/include/vector.hpp

@@ -701,7 +701,8 @@ namespace ngbla
     */
     
     Vec (const Vec &) = default;
-    auto & HTData() const { return data; }                                    
+    auto & HTData() { return data; }
+    const auto & HTData() const { return data; }                                        
     template <typename T2>
     Vec (const Vec<S,T2> & v2) : data(v2.HTData()) { ; }
 
@@ -1196,10 +1197,22 @@ namespace ngbla
     for (int i = 0; i < DIM; i++)
       AtomicAdd (x(i), y(i));
   }
-  
 
+
+  template <typename T, typename TS, typename TDIST, typename TB>
+  inline void AtomicAdd (VectorView<T,TS,TDIST> v, const Expr<TB> & v2)
+  {
+    auto viewv2 = v2.View();
+    auto combsize = CombinedSize(v.Size(), v2.Height());
+    for (size_t i = 0; i < combsize; i++)
+      AtomicAdd (v(i), viewv2(i));
+  }
+
+  
 }
 
+
+
 namespace ngstd
 {
   template <typename ARCHIVE, int S, typename T>

ngsolve/cmake/NGSolveConfig.cmake

@@ -1,4 +1,4 @@
-set(PACKAGE_VERSION "6.2.2505")
+set(PACKAGE_VERSION "6.2.2505-94-geb7c3f7d9")
 find_package(Netgen CONFIG REQUIRED HINTS
   ${CMAKE_CURRENT_LIST_DIR}
   ${CMAKE_CURRENT_LIST_DIR}/..

ngsolve/config/config.py

@@ -29,15 +29,15 @@ NGSOLVE_INSTALL_DIR_INCLUDE  = "netgen/include"
 NGSOLVE_INSTALL_DIR_CMAKE    = "ngsolve/cmake"
 NGSOLVE_INSTALL_DIR_RES      = "share"
 
-NGSOLVE_VERSION = "6.2.2505"
-NGSOLVE_VERSION_GIT = "v6.2.2505-0-g8fbbbc389"
-NGSOLVE_VERSION_PYTHON = "6.2.2505"
+NGSOLVE_VERSION = "6.2.2505-94-geb7c3f7d9"
+NGSOLVE_VERSION_GIT = "v6.2.2505-94-geb7c3f7d9"
+NGSOLVE_VERSION_PYTHON = "6.2.2505.post94.dev0"
 
 NGSOLVE_VERSION_MAJOR = "6"
 NGSOLVE_VERSION_MINOR = "2"
-NGSOLVE_VERSION_TWEAK = "0"
+NGSOLVE_VERSION_TWEAK = "94"
 NGSOLVE_VERSION_PATCH = "2505"
-NGSOLVE_VERSION_HASH = "g8fbbbc389"
+NGSOLVE_VERSION_HASH = "geb7c3f7d9"
 
 CMAKE_CXX_COMPILER           = "/Library/Developer/CommandLineTools/usr/bin/c++"
 CMAKE_CUDA_COMPILER          = ""

{ngsolve-6.2.2505.dist-info → ngsolve-6.2.2505.post94.dev0.dist-info}/METADATA

@@ -1,13 +1,13 @@
 Metadata-Version: 2.1
 Name: ngsolve
-Version: 6.2.2505
+Version: 6.2.2505.post94.dev0
 Summary: NGSolve
 Home-page: UNKNOWN
 Author: The NGSolve team
 License: LGPL2.1
 Platform: UNKNOWN
 License-File: LICENSE
-Requires-Dist: netgen-mesher (==6.2.2505)
+Requires-Dist: netgen-mesher (==6.2.2505.post45.dev0)
 
 UNKNOWN