ruby-eigen 0.0.9 → 0.0.10.pre1

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

@@ -0,0 +1,279 @@
+// 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,typename Index>
+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_aligned(A,m);
+  
+  eigen_internal_assert(((lda%PacketSize)==0) && ((ldc%PacketSize)==0) && (i0==internal::first_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 matrux-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);
+        
+        // agressive 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

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

@@ -0,0 +1,127 @@
+// 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 Index>
+void SparseLUImpl<Scalar,Index>::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(n, et, post); // Post order etree
+  IndexVector inv_post(n+1); 
+  Index i;
+  for (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 (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 
+  Index 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 
+  Index 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 = n;
+    for (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 (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

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

@@ -0,0 +1,130 @@
+// 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/.
+
+#ifndef SPARSELU_KERNEL_BMOD_H
+#define SPARSELU_KERNEL_BMOD_H
+
+namespace Eigen {
+namespace internal {
+  
+/**
+ * \brief Performs numeric block updates from a given supernode to a single column
+ * 
+ * \param segsize Size of the segment (and blocks ) to use for updates
+ * \param[in,out] dense Packed values of the original matrix
+ * \param tempv temporary vector to use for updates
+ * \param lusup array containing the supernodes
+ * \param lda Leading dimension in the supernode
+ * \param nrow Number of rows in the rectangular part of the supernode
+ * \param lsub compressed row subscripts of supernodes
+ * \param lptr pointer to the first column of the current supernode in lsub
+ * \param no_zeros Number of nonzeros elements before the diagonal part of the supernode
+ * \return 0 on success
+ */
+template <int SegSizeAtCompileTime> struct LU_kernel_bmod
+{
+  template <typename BlockScalarVector, typename ScalarVector, typename IndexVector, typename Index>
+  static EIGEN_DONT_INLINE void run(const int segsize, BlockScalarVector& dense, ScalarVector& tempv, ScalarVector& lusup, Index& luptr, const Index lda,
+                                    const Index nrow, IndexVector& lsub, const Index lptr, const Index no_zeros);
+};
+
+template <int SegSizeAtCompileTime>
+template <typename BlockScalarVector, typename ScalarVector, typename IndexVector, typename Index>
+EIGEN_DONT_INLINE void LU_kernel_bmod<SegSizeAtCompileTime>::run(const int segsize, BlockScalarVector& dense, ScalarVector& tempv, ScalarVector& lusup, Index& luptr, const Index lda,
+                                                                  const Index nrow, IndexVector& lsub, const Index lptr, const Index no_zeros)
+{
+  typedef typename ScalarVector::Scalar Scalar;
+  // First, copy U[*,j] segment from dense(*) to tempv(*)
+  // The result of triangular solve is in tempv[*]; 
+    // The result of matric-vector update is in dense[*]
+  Index isub = lptr + no_zeros; 
+  int i;
+  Index irow;
+  for (i = 0; i < ((SegSizeAtCompileTime==Dynamic)?segsize:SegSizeAtCompileTime); i++)
+  {
+    irow = lsub(isub); 
+    tempv(i) = dense(irow); 
+    ++isub; 
+  }
+  // Dense triangular solve -- start effective triangle
+  luptr += lda * no_zeros + no_zeros; 
+  // Form Eigen matrix and vector 
+  Map<Matrix<Scalar,SegSizeAtCompileTime,SegSizeAtCompileTime, ColMajor>, 0, OuterStride<> > A( &(lusup.data()[luptr]), segsize, segsize, OuterStride<>(lda) );
+  Map<Matrix<Scalar,SegSizeAtCompileTime,1> > u(tempv.data(), segsize);
+  
+  u = A.template triangularView<UnitLower>().solve(u); 
+  
+  // Dense matrix-vector product y <-- B*x 
+  luptr += segsize;
+  const Index PacketSize = internal::packet_traits<Scalar>::size;
+  Index ldl = internal::first_multiple(nrow, PacketSize);
+  Map<Matrix<Scalar,Dynamic,SegSizeAtCompileTime, ColMajor>, 0, OuterStride<> > B( &(lusup.data()[luptr]), nrow, segsize, OuterStride<>(lda) );
+  Index aligned_offset = internal::first_aligned(tempv.data()+segsize, PacketSize);
+  Index aligned_with_B_offset = (PacketSize-internal::first_aligned(B.data(), PacketSize))%PacketSize;
+  Map<Matrix<Scalar,Dynamic,1>, 0, OuterStride<> > l(tempv.data()+segsize+aligned_offset+aligned_with_B_offset, nrow, OuterStride<>(ldl) );
+  
+  l.setZero();
+  internal::sparselu_gemm<Scalar>(l.rows(), l.cols(), B.cols(), B.data(), B.outerStride(), u.data(), u.outerStride(), l.data(), l.outerStride());
+  
+  // Scatter tempv[] into SPA dense[] as a temporary storage 
+  isub = lptr + no_zeros;
+  for (i = 0; i < ((SegSizeAtCompileTime==Dynamic)?segsize:SegSizeAtCompileTime); i++)
+  {
+    irow = lsub(isub++); 
+    dense(irow) = tempv(i);
+  }
+  
+  // Scatter l into SPA dense[]
+  for (i = 0; i < nrow; i++)
+  {
+    irow = lsub(isub++); 
+    dense(irow) -= l(i);
+  } 
+}
+
+template <> struct LU_kernel_bmod<1>
+{
+  template <typename BlockScalarVector, typename ScalarVector, typename IndexVector, typename Index>
+  static EIGEN_DONT_INLINE void run(const int /*segsize*/, BlockScalarVector& dense, ScalarVector& /*tempv*/, ScalarVector& lusup, Index& luptr,
+                                    const Index lda, const Index nrow, IndexVector& lsub, const Index lptr, const Index no_zeros);
+};
+
+
+template <typename BlockScalarVector, typename ScalarVector, typename IndexVector, typename Index>
+EIGEN_DONT_INLINE void LU_kernel_bmod<1>::run(const int /*segsize*/, BlockScalarVector& dense, ScalarVector& /*tempv*/, ScalarVector& lusup, Index& luptr,
+                                              const Index lda, const Index nrow, IndexVector& lsub, const Index lptr, const Index no_zeros)
+{
+  typedef typename ScalarVector::Scalar Scalar;
+  Scalar f = dense(lsub(lptr + no_zeros));
+  luptr += lda * no_zeros + no_zeros + 1;
+  const Scalar* a(lusup.data() + luptr);
+  const /*typename IndexVector::Scalar*/Index*  irow(lsub.data()+lptr + no_zeros + 1);
+  Index i = 0;
+  for (; i+1 < nrow; i+=2)
+  {
+    Index i0 = *(irow++);
+    Index i1 = *(irow++);
+    Scalar a0 = *(a++);
+    Scalar a1 = *(a++);
+    Scalar d0 = dense.coeff(i0);
+    Scalar d1 = dense.coeff(i1);
+    d0 -= f*a0;
+    d1 -= f*a1;
+    dense.coeffRef(i0) = d0;
+    dense.coeffRef(i1) = d1;
+  }
+  if(i<nrow)
+    dense.coeffRef(*(irow++)) -= f * *(a++);
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
+#endif // SPARSELU_KERNEL_BMOD_H