ruby-eigen 0.0.9 → 0.0.10.pre1

data/ext/eigen/eigen3/Eigen/src/SparseLU/SparseLU_Utils.h

@@ -0,0 +1,80 @@
+// 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/.
+
+
+#ifndef EIGEN_SPARSELU_UTILS_H
+#define EIGEN_SPARSELU_UTILS_H
+
+namespace Eigen {
+namespace internal {
+
+/**
+ * \brief Count Nonzero elements in the factors
+ */
+template <typename Scalar, typename Index>
+void SparseLUImpl<Scalar,Index>::countnz(const Index n, Index& nnzL, Index& nnzU, GlobalLU_t& glu)
+{
+ nnzL = 0; 
+ nnzU = (glu.xusub)(n); 
+ Index nsuper = (glu.supno)(n); 
+ Index jlen; 
+ Index i, j, fsupc;
+ if (n <= 0 ) return; 
+ // For each supernode
+ for (i = 0; i <= nsuper; i++)
+ {
+   fsupc = glu.xsup(i); 
+   jlen = glu.xlsub(fsupc+1) - glu.xlsub(fsupc); 
+   
+   for (j = fsupc; j < glu.xsup(i+1); j++)
+   {
+     nnzL += jlen; 
+     nnzU += j - fsupc + 1; 
+     jlen--; 
+   }
+ }
+}
+
+/**
+ * \brief Fix up the data storage lsub for L-subscripts. 
+ * 
+ * It removes the subscripts sets for structural pruning, 
+ * and applies permutation to the remaining subscripts
+ * 
+ */
+template <typename Scalar, typename Index>
+void SparseLUImpl<Scalar,Index>::fixupL(const Index n, const IndexVector& perm_r, GlobalLU_t& glu)
+{
+  Index fsupc, i, j, k, jstart; 
+  
+  Index nextl = 0; 
+  Index nsuper = (glu.supno)(n); 
+  
+  // For each supernode 
+  for (i = 0; i <= nsuper; i++)
+  {
+    fsupc = glu.xsup(i); 
+    jstart = glu.xlsub(fsupc); 
+    glu.xlsub(fsupc) = nextl; 
+    for (j = jstart; j < glu.xlsub(fsupc + 1); j++)
+    {
+      glu.lsub(nextl) = perm_r(glu.lsub(j)); // Now indexed into P*A
+      nextl++;
+    }
+    for (k = fsupc+1; k < glu.xsup(i+1); k++)
+      glu.xlsub(k) = nextl; // other columns in supernode i
+  }
+  
+  glu.xlsub(n) = nextl; 
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
+#endif // EIGEN_SPARSELU_UTILS_H

data/ext/eigen/eigen3/Eigen/src/SparseLU/SparseLU_column_bmod.h

@@ -0,0 +1,180 @@
+// 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>
+// 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/.
+
+/* 
+ 
+ * NOTE: This file is the modified version of xcolumn_bmod.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_COLUMN_BMOD_H
+#define SPARSELU_COLUMN_BMOD_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 dense Store the full representation of the column
+ * \param tempv working array 
+ * \param segrep segment representative ...
+ * \param repfnz ??? First nonzero column in each row ???  ...
+ * \param fpanelc First column in the current panel
+ * \param glu Global LU data. 
+ * \return 0 - successful return 
+ *         > 0 - number of bytes allocated when run out of space
+ * 
+ */
+template <typename Scalar, typename Index>
+Index SparseLUImpl<Scalar,Index>::column_bmod(const Index jcol, const Index nseg, BlockScalarVector dense, ScalarVector& tempv, BlockIndexVector segrep, BlockIndexVector repfnz, Index fpanelc, GlobalLU_t& glu)
+{
+  Index  jsupno, k, ksub, krep, ksupno; 
+  Index lptr, nrow, isub, irow, nextlu, new_next, ufirst; 
+  Index fsupc, nsupc, nsupr, luptr, kfnz, no_zeros; 
+  /* krep = representative of current k-th supernode
+    * fsupc =  first supernodal column
+    * nsupc = number of columns in a supernode
+    * nsupr = number of rows in a supernode
+    * luptr = location of supernodal LU-block in storage
+    * kfnz = first nonz in the k-th supernodal segment
+    * no_zeros = no lf leading zeros in a supernodal U-segment
+    */
+  
+  jsupno = glu.supno(jcol);
+  // For each nonzero supernode segment of U[*,j] in topological order 
+  k = nseg - 1; 
+  Index d_fsupc; // distance between the first column of the current panel and the 
+               // first column of the current snode
+  Index fst_col; // First column within small LU update
+  Index segsize; 
+  for (ksub = 0; ksub < nseg; ksub++)
+  {
+    krep = segrep(k); k--; 
+    ksupno = glu.supno(krep); 
+    if (jsupno != ksupno )
+    {
+      // outside the rectangular supernode 
+      fsupc = glu.xsup(ksupno); 
+      fst_col = (std::max)(fsupc, fpanelc); 
+      
+      // Distance from the current supernode to the current panel; 
+      // d_fsupc = 0 if fsupc > fpanelc
+      d_fsupc = fst_col - fsupc; 
+      
+      luptr = glu.xlusup(fst_col) + d_fsupc; 
+      lptr = glu.xlsub(fsupc) + d_fsupc; 
+      
+      kfnz = repfnz(krep); 
+      kfnz = (std::max)(kfnz, fpanelc); 
+      
+      segsize = krep - kfnz + 1; 
+      nsupc = krep - fst_col + 1; 
+      nsupr = glu.xlsub(fsupc+1) - glu.xlsub(fsupc); 
+      nrow = nsupr - d_fsupc - nsupc;
+      Index lda = glu.xlusup(fst_col+1) - glu.xlusup(fst_col);
+      
+      
+      // Perform a triangular solver and block update, 
+      // then scatter the result of sup-col update to dense
+      no_zeros = kfnz - fst_col; 
+      if(segsize==1)
+        LU_kernel_bmod<1>::run(segsize, dense, tempv, glu.lusup, luptr, lda, nrow, glu.lsub, lptr, no_zeros);
+      else
+        LU_kernel_bmod<Dynamic>::run(segsize, dense, tempv, glu.lusup, luptr, lda, nrow, glu.lsub, lptr, no_zeros);
+    } // end if jsupno 
+  } // end for each segment
+  
+  // Process the supernodal portion of  L\U[*,j]
+  nextlu = glu.xlusup(jcol); 
+  fsupc = glu.xsup(jsupno);
+  
+  // copy the SPA dense into L\U[*,j]
+  Index mem; 
+  new_next = nextlu + glu.xlsub(fsupc + 1) - glu.xlsub(fsupc); 
+  Index offset = internal::first_multiple<Index>(new_next, internal::packet_traits<Scalar>::size) - new_next;
+  if(offset)
+    new_next += offset;
+  while (new_next > glu.nzlumax )
+  {
+    mem = memXpand<ScalarVector>(glu.lusup, glu.nzlumax, nextlu, LUSUP, glu.num_expansions);  
+    if (mem) return mem; 
+  }
+  
+  for (isub = glu.xlsub(fsupc); isub < glu.xlsub(fsupc+1); isub++)
+  {
+    irow = glu.lsub(isub);
+    glu.lusup(nextlu) = dense(irow);
+    dense(irow) = Scalar(0.0); 
+    ++nextlu; 
+  }
+  
+  if(offset)
+  {
+    glu.lusup.segment(nextlu,offset).setZero();
+    nextlu += offset;
+  }
+  glu.xlusup(jcol + 1) = nextlu;  // close L\U(*,jcol); 
+  
+  /* For more updates within the panel (also within the current supernode),
+   * should start from the first column of the panel, or the first column
+   * of the supernode, whichever is bigger. There are two cases:
+   *  1) fsupc < fpanelc, then fst_col <-- fpanelc
+   *  2) fsupc >= fpanelc, then fst_col <-- fsupc
+   */
+  fst_col = (std::max)(fsupc, fpanelc); 
+  
+  if (fst_col  < jcol)
+  {
+    // Distance between the current supernode and the current panel
+    // d_fsupc = 0 if fsupc >= fpanelc
+    d_fsupc = fst_col - fsupc; 
+    
+    lptr = glu.xlsub(fsupc) + d_fsupc; 
+    luptr = glu.xlusup(fst_col) + d_fsupc; 
+    nsupr = glu.xlsub(fsupc+1) - glu.xlsub(fsupc); // leading dimension
+    nsupc = jcol - fst_col; // excluding jcol 
+    nrow = nsupr - d_fsupc - nsupc; 
+    
+    // points to the beginning of jcol in snode L\U(jsupno) 
+    ufirst = glu.xlusup(jcol) + d_fsupc; 
+    Index lda = glu.xlusup(jcol+1) - glu.xlusup(jcol);
+    MappedMatrixBlock A( &(glu.lusup.data()[luptr]), nsupc, nsupc, OuterStride<>(lda) );
+    VectorBlock<ScalarVector> u(glu.lusup, ufirst, nsupc); 
+    u = A.template triangularView<UnitLower>().solve(u); 
+    
+    new (&A) MappedMatrixBlock ( &(glu.lusup.data()[luptr+nsupc]), nrow, nsupc, OuterStride<>(lda) );
+    VectorBlock<ScalarVector> l(glu.lusup, ufirst+nsupc, nrow); 
+    l.noalias() -= A * u;
+    
+  } // End if fst_col
+  return 0; 
+}
+
+} // end namespace internal
+} // end namespace Eigen
+
+#endif // SPARSELU_COLUMN_BMOD_H

data/ext/eigen/eigen3/Eigen/src/SparseLU/SparseLU_column_dfs.h

@@ -0,0 +1,177 @@
+// 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 Index> 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 Index;
+  column_dfs_traits(Index jcol, Index& jsuper, typename SparseLUImpl<Scalar, Index>::GlobalLU_t& glu, SparseLUImpl<Scalar, Index>& 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, Index>::GlobalLU_t& m_glu;
+  SparseLUImpl<Scalar, Index>& 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 Index>
+Index SparseLUImpl<Scalar,Index>::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(jcol, perm_r, nseg, glu.lsub, segrep, repfnz, xprune, marker2, parent,
+                   xplore, glu, nextl, krow, traits);
+  } // for each nonzero ... 
+  
+  Index fsupc, jptr, jm1ptr, ito, ifrom, istop;
+  Index nsuper = glu.supno(jcol);
+  Index jcolp1 = 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); 
+    jptr = glu.xlsub(jcol); // Not yet compressed
+    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
+        ito = glu.xlsub(fsupc+1);
+        glu.xlsub(jcolm1) = ito; 
+        istop = ito + jptr - jm1ptr; 
+        xprune(jcolm1) = istop; // intialize xprune(jcol-1)
+        glu.xlsub(jcol) = istop; 
+        
+        for (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) = nextl;  // Intialize upper bound for pruning
+  glu.xlsub(jcolp1) = nextl; 
+  
+  return 0; 
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif

data/ext/eigen/eigen3/Eigen/src/SparseLU/SparseLU_copy_to_ucol.h

@@ -0,0 +1,106 @@
+// 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 Index>
+Index SparseLUImpl<Scalar,Index>::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; 
+  Index 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