pylibsparseir 0.1.0__cp313-cp313-macosx_15_0_arm64.whl → 0.5.2__cp313-cp313-macosx_15_0_arm64.whl

include/eigen3/Eigen/src/SparseLU/SparseLU_column_dfs.h

@@ -0,0 +1,179 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2012 Désiré Nuentsa-Wakam <desire.nuentsa_wakam@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+/* 
+ 
+ * NOTE: This file is the modified version of [s,d,c,z]column_dfs.c file in SuperLU 
+ 
+ * -- SuperLU routine (version 2.0) --
+ * Univ. of California Berkeley, Xerox Palo Alto Research Center,
+ * and Lawrence Berkeley National Lab.
+ * November 15, 1997
+ *
+ * Copyright (c) 1994 by Xerox Corporation.  All rights reserved.
+ *
+ * THIS MATERIAL IS PROVIDED AS IS, WITH ABSOLUTELY NO WARRANTY
+ * EXPRESSED OR IMPLIED.  ANY USE IS AT YOUR OWN RISK.
+ *
+ * Permission is hereby granted to use or copy this program for any
+ * purpose, provided the above notices are retained on all copies.
+ * Permission to modify the code and to distribute modified code is
+ * granted, provided the above notices are retained, and a notice that
+ * the code was modified is included with the above copyright notice.
+ */
+#ifndef SPARSELU_COLUMN_DFS_H
+#define SPARSELU_COLUMN_DFS_H
+
+template <typename Scalar, typename StorageIndex> class SparseLUImpl;
+namespace Eigen {
+
+namespace internal {
+
+template<typename IndexVector, typename ScalarVector>
+struct column_dfs_traits : no_assignment_operator
+{
+  typedef typename ScalarVector::Scalar Scalar;
+  typedef typename IndexVector::Scalar StorageIndex;
+  column_dfs_traits(Index jcol, Index& jsuper, typename SparseLUImpl<Scalar, StorageIndex>::GlobalLU_t& glu, SparseLUImpl<Scalar, StorageIndex>& luImpl)
+   : m_jcol(jcol), m_jsuper_ref(jsuper), m_glu(glu), m_luImpl(luImpl)
+ {}
+  bool update_segrep(Index /*krep*/, Index /*jj*/)
+  {
+    return true;
+  }
+  void mem_expand(IndexVector& lsub, Index& nextl, Index chmark)
+  {
+    if (nextl >= m_glu.nzlmax)
+      m_luImpl.memXpand(lsub, m_glu.nzlmax, nextl, LSUB, m_glu.num_expansions); 
+    if (chmark != (m_jcol-1)) m_jsuper_ref = emptyIdxLU;
+  }
+  enum { ExpandMem = true };
+  
+  Index m_jcol;
+  Index& m_jsuper_ref;
+  typename SparseLUImpl<Scalar, StorageIndex>::GlobalLU_t& m_glu;
+  SparseLUImpl<Scalar, StorageIndex>& m_luImpl;
+};
+
+
+/**
+ * \brief Performs a symbolic factorization on column jcol and decide the supernode boundary
+ * 
+ * A supernode representative is the last column of a supernode.
+ * The nonzeros in U[*,j] are segments that end at supernodes representatives. 
+ * The routine returns a list of the supernodal representatives 
+ * in topological order of the dfs that generates them. 
+ * The location of the first nonzero in each supernodal segment 
+ * (supernodal entry location) is also returned. 
+ * 
+ * \param m number of rows in the matrix
+ * \param jcol Current column 
+ * \param perm_r Row permutation
+ * \param maxsuper  Maximum number of column allowed in a supernode
+ * \param [in,out] nseg Number of segments in current U[*,j] - new segments appended
+ * \param lsub_col defines the rhs vector to start the dfs
+ * \param [in,out] segrep Segment representatives - new segments appended 
+ * \param repfnz  First nonzero location in each row
+ * \param xprune 
+ * \param marker  marker[i] == jj, if i was visited during dfs of current column jj;
+ * \param parent
+ * \param xplore working array
+ * \param glu global LU data 
+ * \return 0 success
+ *         > 0 number of bytes allocated when run out of space
+ * 
+ */
+template <typename Scalar, typename StorageIndex>
+Index SparseLUImpl<Scalar,StorageIndex>::column_dfs(const Index m, const Index jcol, IndexVector& perm_r, Index maxsuper, Index& nseg,
+                                                    BlockIndexVector lsub_col, IndexVector& segrep, BlockIndexVector repfnz, IndexVector& xprune,
+                                                    IndexVector& marker, IndexVector& parent, IndexVector& xplore, GlobalLU_t& glu)
+{
+  
+  Index jsuper = glu.supno(jcol); 
+  Index nextl = glu.xlsub(jcol); 
+  VectorBlock<IndexVector> marker2(marker, 2*m, m); 
+  
+  
+  column_dfs_traits<IndexVector, ScalarVector> traits(jcol, jsuper, glu, *this);
+  
+  // For each nonzero in A(*,jcol) do dfs 
+  for (Index k = 0; ((k < m) ? lsub_col[k] != emptyIdxLU : false) ; k++)
+  {
+    Index krow = lsub_col(k); 
+    lsub_col(k) = emptyIdxLU; 
+    Index kmark = marker2(krow); 
+    
+    // krow was visited before, go to the next nonz; 
+    if (kmark == jcol) continue;
+    
+    dfs_kernel(StorageIndex(jcol), perm_r, nseg, glu.lsub, segrep, repfnz, xprune, marker2, parent,
+                   xplore, glu, nextl, krow, traits);
+  } // for each nonzero ... 
+  
+  Index fsupc;
+  StorageIndex nsuper = glu.supno(jcol);
+  StorageIndex jcolp1 = StorageIndex(jcol) + 1;
+  Index jcolm1 = jcol - 1;
+  
+  // check to see if j belongs in the same supernode as j-1
+  if ( jcol == 0 )
+  { // Do nothing for column 0 
+    nsuper = glu.supno(0) = 0 ;
+  }
+  else 
+  {
+    fsupc = glu.xsup(nsuper); 
+    StorageIndex jptr = glu.xlsub(jcol); // Not yet compressed
+    StorageIndex jm1ptr = glu.xlsub(jcolm1); 
+    
+    // Use supernodes of type T2 : see SuperLU paper
+    if ( (nextl-jptr != jptr-jm1ptr-1) ) jsuper = emptyIdxLU;
+    
+    // Make sure the number of columns in a supernode doesn't
+    // exceed threshold
+    if ( (jcol - fsupc) >= maxsuper) jsuper = emptyIdxLU; 
+    
+    /* If jcol starts a new supernode, reclaim storage space in
+     * glu.lsub from previous supernode. Note we only store 
+     * the subscript set of the first and last columns of 
+     * a supernode. (first for num values, last for pruning)
+     */
+    if (jsuper == emptyIdxLU)
+    { // starts a new supernode 
+      if ( (fsupc < jcolm1-1) ) 
+      { // >= 3 columns in nsuper
+        StorageIndex ito = glu.xlsub(fsupc+1);
+        glu.xlsub(jcolm1) = ito; 
+        StorageIndex istop = ito + jptr - jm1ptr; 
+        xprune(jcolm1) = istop; // initialize xprune(jcol-1)
+        glu.xlsub(jcol) = istop; 
+        
+        for (StorageIndex ifrom = jm1ptr; ifrom < nextl; ++ifrom, ++ito)
+          glu.lsub(ito) = glu.lsub(ifrom); 
+        nextl = ito;  // = istop + length(jcol)
+      }
+      nsuper++; 
+      glu.supno(jcol) = nsuper; 
+    } // if a new supernode 
+  } // end else:  jcol > 0
+  
+  // Tidy up the pointers before exit
+  glu.xsup(nsuper+1) = jcolp1; 
+  glu.supno(jcolp1) = nsuper; 
+  xprune(jcol) = StorageIndex(nextl);  // Initialize upper bound for pruning
+  glu.xlsub(jcolp1) = StorageIndex(nextl); 
+  
+  return 0; 
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif

include/eigen3/Eigen/src/SparseLU/SparseLU_copy_to_ucol.h

@@ -0,0 +1,107 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2012 Désiré Nuentsa-Wakam <desire.nuentsa_wakam@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+/* 
+ 
+ * NOTE: This file is the modified version of [s,d,c,z]copy_to_ucol.c file in SuperLU 
+ 
+ * -- SuperLU routine (version 2.0) --
+ * Univ. of California Berkeley, Xerox Palo Alto Research Center,
+ * and Lawrence Berkeley National Lab.
+ * November 15, 1997
+ *
+ * Copyright (c) 1994 by Xerox Corporation.  All rights reserved.
+ *
+ * THIS MATERIAL IS PROVIDED AS IS, WITH ABSOLUTELY NO WARRANTY
+ * EXPRESSED OR IMPLIED.  ANY USE IS AT YOUR OWN RISK.
+ *
+ * Permission is hereby granted to use or copy this program for any
+ * purpose, provided the above notices are retained on all copies.
+ * Permission to modify the code and to distribute modified code is
+ * granted, provided the above notices are retained, and a notice that
+ * the code was modified is included with the above copyright notice.
+ */
+#ifndef SPARSELU_COPY_TO_UCOL_H
+#define SPARSELU_COPY_TO_UCOL_H
+
+namespace Eigen {
+namespace internal {
+
+/**
+ * \brief Performs numeric block updates (sup-col) in topological order
+ * 
+ * \param jcol current column to update
+ * \param nseg Number of segments in the U part
+ * \param segrep segment representative ...
+ * \param repfnz First nonzero column in each row  ...
+ * \param perm_r Row permutation 
+ * \param dense Store the full representation of the column
+ * \param glu Global LU data. 
+ * \return 0 - successful return 
+ *         > 0 - number of bytes allocated when run out of space
+ * 
+ */
+template <typename Scalar, typename StorageIndex>
+Index SparseLUImpl<Scalar,StorageIndex>::copy_to_ucol(const Index jcol, const Index nseg, IndexVector& segrep,
+                                                      BlockIndexVector repfnz ,IndexVector& perm_r, BlockScalarVector dense, GlobalLU_t& glu)
+{  
+  Index ksub, krep, ksupno; 
+    
+  Index jsupno = glu.supno(jcol);
+  
+  // For each nonzero supernode segment of U[*,j] in topological order 
+  Index k = nseg - 1, i; 
+  StorageIndex nextu = glu.xusub(jcol); 
+  Index kfnz, isub, segsize; 
+  Index new_next,irow; 
+  Index fsupc, mem; 
+  for (ksub = 0; ksub < nseg; ksub++)
+  {
+    krep = segrep(k); k--; 
+    ksupno = glu.supno(krep); 
+    if (jsupno != ksupno ) // should go into ucol(); 
+    {
+      kfnz = repfnz(krep); 
+      if (kfnz != emptyIdxLU)
+      { // Nonzero U-segment 
+        fsupc = glu.xsup(ksupno); 
+        isub = glu.xlsub(fsupc) + kfnz - fsupc; 
+        segsize = krep - kfnz + 1; 
+        new_next = nextu + segsize; 
+        while (new_next > glu.nzumax) 
+        {
+          mem = memXpand<ScalarVector>(glu.ucol, glu.nzumax, nextu, UCOL, glu.num_expansions); 
+          if (mem) return mem; 
+          mem = memXpand<IndexVector>(glu.usub, glu.nzumax, nextu, USUB, glu.num_expansions); 
+          if (mem) return mem; 
+          
+        }
+        
+        for (i = 0; i < segsize; i++)
+        {
+          irow = glu.lsub(isub); 
+          glu.usub(nextu) = perm_r(irow); // Unlike the L part, the U part is stored in its final order
+          glu.ucol(nextu) = dense(irow); 
+          dense(irow) = Scalar(0.0); 
+          nextu++;
+          isub++;
+        }
+        
+      } // end nonzero U-segment 
+      
+    } // end if jsupno 
+    
+  } // end for each segment
+  glu.xusub(jcol + 1) = nextu; // close U(*,jcol)
+  return 0; 
+}
+
+} // namespace internal
+} // end namespace Eigen
+
+#endif // SPARSELU_COPY_TO_UCOL_H

include/eigen3/Eigen/src/SparseLU/SparseLU_gemm_kernel.h

@@ -0,0 +1,280 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2012 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SPARSELU_GEMM_KERNEL_H
+#define EIGEN_SPARSELU_GEMM_KERNEL_H
+
+namespace Eigen {
+
+namespace internal {
+
+
+/** \internal
+  * A general matrix-matrix product kernel optimized for the SparseLU factorization.
+  *  - A, B, and C must be column major
+  *  - lda and ldc must be multiples of the respective packet size
+  *  - C must have the same alignment as A
+  */
+template<typename Scalar>
+EIGEN_DONT_INLINE
+void sparselu_gemm(Index m, Index n, Index d, const Scalar* A, Index lda, const Scalar* B, Index ldb, Scalar* C, Index ldc)
+{
+  using namespace Eigen::internal;
+  
+  typedef typename packet_traits<Scalar>::type Packet;
+  enum {
+    NumberOfRegisters = EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS,
+    PacketSize = packet_traits<Scalar>::size,
+    PM = 8,                             // peeling in M
+    RN = 2,                             // register blocking
+    RK = NumberOfRegisters>=16 ? 4 : 2, // register blocking
+    BM = 4096/sizeof(Scalar),           // number of rows of A-C per chunk
+    SM = PM*PacketSize                  // step along M
+  };
+  Index d_end = (d/RK)*RK;    // number of columns of A (rows of B) suitable for full register blocking
+  Index n_end = (n/RN)*RN;    // number of columns of B-C suitable for processing RN columns at once
+  Index i0 = internal::first_default_aligned(A,m);
+  
+  eigen_internal_assert(((lda%PacketSize)==0) && ((ldc%PacketSize)==0) && (i0==internal::first_default_aligned(C,m)));
+  
+  // handle the non aligned rows of A and C without any optimization:
+  for(Index i=0; i<i0; ++i)
+  {
+    for(Index j=0; j<n; ++j)
+    {
+      Scalar c = C[i+j*ldc];
+      for(Index k=0; k<d; ++k)
+        c += B[k+j*ldb] * A[i+k*lda];
+      C[i+j*ldc] = c;
+    }
+  }
+  // process the remaining rows per chunk of BM rows
+  for(Index ib=i0; ib<m; ib+=BM)
+  {
+    Index actual_b = std::min<Index>(BM, m-ib);                 // actual number of rows
+    Index actual_b_end1 = (actual_b/SM)*SM;                   // actual number of rows suitable for peeling
+    Index actual_b_end2 = (actual_b/PacketSize)*PacketSize;   // actual number of rows suitable for vectorization
+    
+    // Let's process two columns of B-C at once
+    for(Index j=0; j<n_end; j+=RN)
+    {
+      const Scalar* Bc0 = B+(j+0)*ldb;
+      const Scalar* Bc1 = B+(j+1)*ldb;
+      
+      for(Index k=0; k<d_end; k+=RK)
+      {
+        
+        // load and expand a RN x RK block of B
+        Packet b00, b10, b20, b30, b01, b11, b21, b31;
+                  { b00 = pset1<Packet>(Bc0[0]); }
+                  { b10 = pset1<Packet>(Bc0[1]); }
+        if(RK==4) { b20 = pset1<Packet>(Bc0[2]); }
+        if(RK==4) { b30 = pset1<Packet>(Bc0[3]); }
+                  { b01 = pset1<Packet>(Bc1[0]); }
+                  { b11 = pset1<Packet>(Bc1[1]); }
+        if(RK==4) { b21 = pset1<Packet>(Bc1[2]); }
+        if(RK==4) { b31 = pset1<Packet>(Bc1[3]); }
+        
+        Packet a0, a1, a2, a3, c0, c1, t0, t1;
+        
+        const Scalar* A0 = A+ib+(k+0)*lda;
+        const Scalar* A1 = A+ib+(k+1)*lda;
+        const Scalar* A2 = A+ib+(k+2)*lda;
+        const Scalar* A3 = A+ib+(k+3)*lda;
+        
+        Scalar* C0 = C+ib+(j+0)*ldc;
+        Scalar* C1 = C+ib+(j+1)*ldc;
+        
+                  a0 = pload<Packet>(A0);
+                  a1 = pload<Packet>(A1);
+        if(RK==4)
+        {
+          a2 = pload<Packet>(A2);
+          a3 = pload<Packet>(A3);
+        }
+        else
+        {
+          // workaround "may be used uninitialized in this function" warning
+          a2 = a3 = a0;
+        }
+        
+#define KMADD(c, a, b, tmp) {tmp = b; tmp = pmul(a,tmp); c = padd(c,tmp);}
+#define WORK(I)  \
+                     c0 = pload<Packet>(C0+i+(I)*PacketSize);    \
+                     c1 = pload<Packet>(C1+i+(I)*PacketSize);    \
+                     KMADD(c0, a0, b00, t0)                      \
+                     KMADD(c1, a0, b01, t1)                      \
+                     a0 = pload<Packet>(A0+i+(I+1)*PacketSize);  \
+                     KMADD(c0, a1, b10, t0)                      \
+                     KMADD(c1, a1, b11, t1)                      \
+                     a1 = pload<Packet>(A1+i+(I+1)*PacketSize);  \
+          if(RK==4){ KMADD(c0, a2, b20, t0)                     }\
+          if(RK==4){ KMADD(c1, a2, b21, t1)                     }\
+          if(RK==4){ a2 = pload<Packet>(A2+i+(I+1)*PacketSize); }\
+          if(RK==4){ KMADD(c0, a3, b30, t0)                     }\
+          if(RK==4){ KMADD(c1, a3, b31, t1)                     }\
+          if(RK==4){ a3 = pload<Packet>(A3+i+(I+1)*PacketSize); }\
+                     pstore(C0+i+(I)*PacketSize, c0);            \
+                     pstore(C1+i+(I)*PacketSize, c1)
+        
+        // process rows of A' - C' with aggressive vectorization and peeling 
+        for(Index i=0; i<actual_b_end1; i+=PacketSize*8)
+        {
+          EIGEN_ASM_COMMENT("SPARSELU_GEMML_KERNEL1");
+                    prefetch((A0+i+(5)*PacketSize));
+                    prefetch((A1+i+(5)*PacketSize));
+          if(RK==4) prefetch((A2+i+(5)*PacketSize));
+          if(RK==4) prefetch((A3+i+(5)*PacketSize));
+
+          WORK(0);
+          WORK(1);
+          WORK(2);
+          WORK(3);
+          WORK(4);
+          WORK(5);
+          WORK(6);
+          WORK(7);
+        }
+        // process the remaining rows with vectorization only
+        for(Index i=actual_b_end1; i<actual_b_end2; i+=PacketSize)
+        {
+          WORK(0);
+        }
+#undef WORK
+        // process the remaining rows without vectorization
+        for(Index i=actual_b_end2; i<actual_b; ++i)
+        {
+          if(RK==4)
+          {
+            C0[i] += A0[i]*Bc0[0]+A1[i]*Bc0[1]+A2[i]*Bc0[2]+A3[i]*Bc0[3];
+            C1[i] += A0[i]*Bc1[0]+A1[i]*Bc1[1]+A2[i]*Bc1[2]+A3[i]*Bc1[3];
+          }
+          else
+          {
+            C0[i] += A0[i]*Bc0[0]+A1[i]*Bc0[1];
+            C1[i] += A0[i]*Bc1[0]+A1[i]*Bc1[1];
+          }
+        }
+        
+        Bc0 += RK;
+        Bc1 += RK;
+      } // peeled loop on k
+    } // peeled loop on the columns j
+    // process the last column (we now perform a matrix-vector product)
+    if((n-n_end)>0)
+    {
+      const Scalar* Bc0 = B+(n-1)*ldb;
+      
+      for(Index k=0; k<d_end; k+=RK)
+      {
+        
+        // load and expand a 1 x RK block of B
+        Packet b00, b10, b20, b30;
+                  b00 = pset1<Packet>(Bc0[0]);
+                  b10 = pset1<Packet>(Bc0[1]);
+        if(RK==4) b20 = pset1<Packet>(Bc0[2]);
+        if(RK==4) b30 = pset1<Packet>(Bc0[3]);
+        
+        Packet a0, a1, a2, a3, c0, t0/*, t1*/;
+        
+        const Scalar* A0 = A+ib+(k+0)*lda;
+        const Scalar* A1 = A+ib+(k+1)*lda;
+        const Scalar* A2 = A+ib+(k+2)*lda;
+        const Scalar* A3 = A+ib+(k+3)*lda;
+        
+        Scalar* C0 = C+ib+(n_end)*ldc;
+        
+                  a0 = pload<Packet>(A0);
+                  a1 = pload<Packet>(A1);
+        if(RK==4)
+        {
+          a2 = pload<Packet>(A2);
+          a3 = pload<Packet>(A3);
+        }
+        else
+        {
+          // workaround "may be used uninitialized in this function" warning
+          a2 = a3 = a0;
+        }
+        
+#define WORK(I) \
+                   c0 = pload<Packet>(C0+i+(I)*PacketSize);     \
+                   KMADD(c0, a0, b00, t0)                       \
+                   a0 = pload<Packet>(A0+i+(I+1)*PacketSize);   \
+                   KMADD(c0, a1, b10, t0)                       \
+                   a1 = pload<Packet>(A1+i+(I+1)*PacketSize);   \
+        if(RK==4){ KMADD(c0, a2, b20, t0)                      }\
+        if(RK==4){ a2 = pload<Packet>(A2+i+(I+1)*PacketSize);  }\
+        if(RK==4){ KMADD(c0, a3, b30, t0)                      }\
+        if(RK==4){ a3 = pload<Packet>(A3+i+(I+1)*PacketSize);  }\
+                   pstore(C0+i+(I)*PacketSize, c0);
+        
+        // aggressive vectorization and peeling
+        for(Index i=0; i<actual_b_end1; i+=PacketSize*8)
+        {
+          EIGEN_ASM_COMMENT("SPARSELU_GEMML_KERNEL2");
+          WORK(0);
+          WORK(1);
+          WORK(2);
+          WORK(3);
+          WORK(4);
+          WORK(5);
+          WORK(6);
+          WORK(7);
+        }
+        // vectorization only
+        for(Index i=actual_b_end1; i<actual_b_end2; i+=PacketSize)
+        {
+          WORK(0);
+        }
+        // remaining scalars
+        for(Index i=actual_b_end2; i<actual_b; ++i)
+        {
+          if(RK==4) 
+            C0[i] += A0[i]*Bc0[0]+A1[i]*Bc0[1]+A2[i]*Bc0[2]+A3[i]*Bc0[3];
+          else
+            C0[i] += A0[i]*Bc0[0]+A1[i]*Bc0[1];
+        }
+        
+        Bc0 += RK;
+#undef WORK
+      }
+    }
+    
+    // process the last columns of A, corresponding to the last rows of B
+    Index rd = d-d_end;
+    if(rd>0)
+    {
+      for(Index j=0; j<n; ++j)
+      {
+        enum {
+          Alignment = PacketSize>1 ? Aligned : 0
+        };
+        typedef Map<Matrix<Scalar,Dynamic,1>, Alignment > MapVector;
+        typedef Map<const Matrix<Scalar,Dynamic,1>, Alignment > ConstMapVector;
+        if(rd==1)       MapVector(C+j*ldc+ib,actual_b) += B[0+d_end+j*ldb] * ConstMapVector(A+(d_end+0)*lda+ib, actual_b);
+        
+        else if(rd==2)  MapVector(C+j*ldc+ib,actual_b) += B[0+d_end+j*ldb] * ConstMapVector(A+(d_end+0)*lda+ib, actual_b)
+                                                        + B[1+d_end+j*ldb] * ConstMapVector(A+(d_end+1)*lda+ib, actual_b);
+        
+        else            MapVector(C+j*ldc+ib,actual_b) += B[0+d_end+j*ldb] * ConstMapVector(A+(d_end+0)*lda+ib, actual_b)
+                                                        + B[1+d_end+j*ldb] * ConstMapVector(A+(d_end+1)*lda+ib, actual_b)
+                                                        + B[2+d_end+j*ldb] * ConstMapVector(A+(d_end+2)*lda+ib, actual_b);
+      }
+    }
+  
+  } // blocking on the rows of A and C
+}
+#undef KMADD
+
+} // namespace internal
+
+} // namespace Eigen
+
+#endif // EIGEN_SPARSELU_GEMM_KERNEL_H

include/eigen3/Eigen/src/SparseLU/SparseLU_heap_relax_snode.h

@@ -0,0 +1,126 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2012 Désiré Nuentsa-Wakam <desire.nuentsa_wakam@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+/* This file is a modified version of heap_relax_snode.c file in SuperLU
+ * -- SuperLU routine (version 3.0) --
+ * Univ. of California Berkeley, Xerox Palo Alto Research Center,
+ * and Lawrence Berkeley National Lab.
+ * October 15, 2003
+ *
+ * Copyright (c) 1994 by Xerox Corporation.  All rights reserved.
+ *
+ * THIS MATERIAL IS PROVIDED AS IS, WITH ABSOLUTELY NO WARRANTY
+ * EXPRESSED OR IMPLIED.  ANY USE IS AT YOUR OWN RISK.
+ *
+ * Permission is hereby granted to use or copy this program for any
+ * purpose, provided the above notices are retained on all copies.
+ * Permission to modify the code and to distribute modified code is
+ * granted, provided the above notices are retained, and a notice that
+ * the code was modified is included with the above copyright notice.
+ */
+
+#ifndef SPARSELU_HEAP_RELAX_SNODE_H
+#define SPARSELU_HEAP_RELAX_SNODE_H
+
+namespace Eigen {
+namespace internal {
+
+/** 
+ * \brief Identify the initial relaxed supernodes
+ * 
+ * This routine applied to a symmetric elimination tree. 
+ * It assumes that the matrix has been reordered according to the postorder of the etree
+ * \param n The number of columns
+ * \param et elimination tree 
+ * \param relax_columns Maximum number of columns allowed in a relaxed snode 
+ * \param descendants Number of descendants of each node in the etree
+ * \param relax_end last column in a supernode
+ */
+template <typename Scalar, typename StorageIndex>
+void SparseLUImpl<Scalar,StorageIndex>::heap_relax_snode (const Index n, IndexVector& et, const Index relax_columns, IndexVector& descendants, IndexVector& relax_end)
+{
+  
+  // The etree may not be postordered, but its heap ordered  
+  IndexVector post;
+  internal::treePostorder(StorageIndex(n), et, post); // Post order etree
+  IndexVector inv_post(n+1); 
+  for (StorageIndex i = 0; i < n+1; ++i) inv_post(post(i)) = i; // inv_post = post.inverse()???
+  
+  // Renumber etree in postorder 
+  IndexVector iwork(n);
+  IndexVector et_save(n+1);
+  for (Index i = 0; i < n; ++i)
+  {
+    iwork(post(i)) = post(et(i));
+  }
+  et_save = et; // Save the original etree
+  et = iwork; 
+  
+  // compute the number of descendants of each node in the etree
+  relax_end.setConstant(emptyIdxLU);
+  Index j, parent; 
+  descendants.setZero();
+  for (j = 0; j < n; j++) 
+  {
+    parent = et(j);
+    if (parent != n) // not the dummy root
+      descendants(parent) += descendants(j) + 1;
+  }
+  // Identify the relaxed supernodes by postorder traversal of the etree
+  Index snode_start; // beginning of a snode 
+  StorageIndex k;
+  Index nsuper_et_post = 0; // Number of relaxed snodes in postordered etree 
+  Index nsuper_et = 0; // Number of relaxed snodes in the original etree 
+  StorageIndex l; 
+  for (j = 0; j < n; )
+  {
+    parent = et(j);
+    snode_start = j; 
+    while ( parent != n && descendants(parent) < relax_columns ) 
+    {
+      j = parent; 
+      parent = et(j);
+    }
+    // Found a supernode in postordered etree, j is the last column 
+    ++nsuper_et_post;
+    k = StorageIndex(n);
+    for (Index i = snode_start; i <= j; ++i)
+      k = (std::min)(k, inv_post(i));
+    l = inv_post(j);
+    if ( (l - k) == (j - snode_start) )  // Same number of columns in the snode
+    {
+      // This is also a supernode in the original etree
+      relax_end(k) = l; // Record last column 
+      ++nsuper_et; 
+    }
+    else 
+    {
+      for (Index i = snode_start; i <= j; ++i) 
+      {
+        l = inv_post(i);
+        if (descendants(i) == 0) 
+        {
+          relax_end(l) = l;
+          ++nsuper_et;
+        }
+      }
+    }
+    j++;
+    // Search for a new leaf
+    while (descendants(j) != 0 && j < n) j++;
+  } // End postorder traversal of the etree
+  
+  // Recover the original etree
+  et = et_save; 
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
+#endif // SPARSELU_HEAP_RELAX_SNODE_H