Merge pull request #66 from jacobmerson/feature-scalarMatMult

Create ScalarMatMultiply

Author: jacobmerson

core/lasCSRCore.cc

@@ -40,10 +40,34 @@ namespace las
       }
     }
   };
+  class IdentityCSR : public CSRBuilder
+  {
+  protected:
+    int ndofs;
+  public:
+    IdentityCSR(int ndofs)
+      : CSRBuilder(ndofs,ndofs)
+      , ndofs(ndofs)
+    {
+    }
+    void run()
+    {
+      for(int i=0; i<ndofs;++i)
+      {
+        add(i,i);
+      }
+    }
+  };
   Sparsity * createCSR(apf::Numbering * num, int ndofs)
   {
     CSRFromNumbering bldr(num,ndofs);
     bldr.run();
     return bldr.finalize();
   }
+  Sparsity * createIdentityCSR(int ndofs)
+  {
+    IdentityCSR bldr(ndofs);
+    bldr.run();
+    return bldr.finalize();
+  }
 }

core/lasCSRCore.h

@@ -10,5 +10,6 @@ namespace las
    *  values of the numbering, no mesh partitioning is considered.
    */
   Sparsity * createCSR(apf::Numbering * num, int ndofs);
+  Sparsity * createIdentityCSR(int ndofs);
 }
 #endif

src/las.h

@@ -71,6 +71,10 @@ namespace las
     {
       static_cast<T*>(this)->_set(v,cnt,rws,vls);
     }
+    void set(Vec * v, scalar * vls)
+    {
+      static_cast<T*>(this)->_set(v,vls);
+    }
     void set(Mat * m, int cntr, int * rws, int cntc, int * cls, scalar * vls)
     {
       static_cast<T*>(this)->_set(m,cntr,rws,cntc,cls,vls);
@@ -198,6 +202,8 @@ namespace las
     virtual void solve(Mat * k, Vec * u, Vec * f) = 0;
     virtual ~Solve() {}
   };
+  template <typename T>
+  Solve * getSolve(int id);
   /**
    * Interface for Matrix-Vector multiplication
    * @todo Retrieve backend-specific solvers using
@@ -210,6 +216,8 @@ namespace las
     virtual void exec(Mat * x, Vec * a, Vec * b) = 0;
     virtual ~MatVecMult() {}
   };
+  template <class T>
+  MatVecMult * getMatVecMult();
   /**
    * Interface for Matrix-Matrix multiplication
    * @todo Retrieve backend-specific solvers using
@@ -222,6 +230,56 @@ namespace las
     virtual void exec(Mat * a, Mat * b, Mat ** c) = 0;
     virtual ~MatMatMult() {}
   };
+  template <class T>
+  MatMatMult * getMatMatMult();
+  /**
+   * Interface for Scalar-Matrix multiplication
+   * If c is NULL performs an in place multiplication
+   * @todo Retrieve backend-specific solvers using
+   *       backend id classes to do template
+   *       specialization, as above.
+   */
+  class ScalarMatMult
+  {
+    public:
+    virtual void exec(scalar s, Mat * a, Mat ** c) = 0;
+    virtual ~ScalarMatMult() {}
+  };
+  template <class T>
+  ScalarMatMult * getScalarMatMult();
+  /*
+   * interface for C = alpha_1*A+alpha_2*B
+   */
+  class MatMatAdd 
+  {
+    public:
+      virtual void exec(scalar s1, Mat * a, scalar s2, Mat * b, Mat ** c) = 0;
+      virtual ~MatMatAdd() {}
+  };
+  template <class T>
+  MatMatAdd * getMatMatAdd();
+  /*
+   * interface for vector vector addition
+   */
+  class VecVecAdd
+  {
+    public:
+    virtual void exec(scalar s1, Vec * v1, scalar s2, Vec * v2, Vec *& v3) = 0;
+    virtual ~VecVecAdd() {}
+  };
+  template <class T>
+  VecVecAdd * getVecVecAdd();
+  /*
+   * Interface for scalar-vector multiplication
+   */
+  class ScalarVecMult
+  {
+    public:
+      virtual void exec(scalar s, Vec * x, Vec ** y) = 0;
+      virtual ~ScalarVecMult() {}
+  };
+  template <class T>
+  ScalarVecMult * getScalarVecMult();
   /*
    * Finalize routines which must be called on a matrix when switching from
    * add mode to set mode

src/lasCSR.h

@@ -3,6 +3,8 @@
 #include "lasSys.h"
 #include <vector>
 #include <iostream>
+#include <algorithm>
+#include <cassert>
 namespace las
 {
   class Sparsity;
@@ -35,21 +37,29 @@ namespace las
      *       they are converted to use 1-indexing (in a debug build this generates a warning).
      */
     CSR(int r, int c, int nnz, int * rs, int * cs);
+    CSR(int r, int c, int nnz, std::vector<int> const & rs, std::vector<int> const & cs);
     int getNumRows() const { return nr; }
     int getNumCols() const { return nc; }
     int getNumNonzero() const { return nnz; }
+    // return the index into the values array
+    // if the location is not stored then return -1
+    // note rw and cl start at zero
     int operator()(int rw, int cl) const
     {
-      int result = -1;
-      int fst = rws[rw] - 1;
-      while((fst < rws[rw+1] - 2) && (cls[fst] - 1 < cl))
-        ++fst;
-      // the column is correct at offset and the row isn't empty
-      if(cls[fst] - 1 == cl && rws[rw] - 1 <= rws[rw+1] - 2)
-        result = fst;
-      else
-        result = -1;
-      return result;
+      assert(rw < nr && rw>=0);
+      assert(cl < nc && cl>=0);
+      // the row is empty
+      if(rws[rw+1]-rws[rw] == 0)
+        return -1;
+      // this approach finds the correct index in log(n) time where
+      // n is the number of elements on the row
+      typedef std::vector<int>::const_iterator vit_t;
+      vit_t bgn = cls.begin()+rws[rw]-1;
+      vit_t end = cls.begin()+rws[rw+1]-1;
+      std::pair<vit_t, vit_t> bounds = equal_range(bgn, end, cl+1);
+      if(bounds.first == bounds.second)
+        return -1;
+      return bounds.first-cls.begin();
     }
     int * getRows() { return &rws[0]; }
     int * getCols() { return &cls[0]; }

src/lasCSRBuilder_impl.h

@@ -29,9 +29,11 @@ namespace las
     auto it = std::unique(coords.begin(), coords.end(), unique_comps<int,int>);
     coords.resize(std::distance(coords.begin(), it));
     assert(nnz >= coords.size());
-    if(coords.size() < nnz) {
-      std::cerr<<"Warning: ignored "<<nnz-coords.size() << " duplicate entries\n";
-    }
+    // this is useful for debugging, but doesn't make sense when the builder is needed
+    // for things like the add function...
+    //if(coords.size() < nnz) {
+    //  std::cerr<<"Warning: ignored "<<nnz-coords.size() << " duplicate entries\n";
+    //}
     nnz = coords.size();
     cls.resize(nnz);
     for(std::size_t i=0; i<coords.size(); ++i) {
@@ -61,7 +63,7 @@ namespace las
       result = false;
     if(rw>=rw_cnt || cl>=cl_cnt) 
     {
-      std::cerr<<"Warning: inserting a row outside the matrix bounds. Skipping it.\n";
+      std::cerr<<"Warning: inserting a row/column outside the matrix bounds ("<<rw<<","<<cl<<"). Skipping it.\n";
       result=false;
     }
     return result;

src/lasCSR_impl.h

@@ -23,6 +23,14 @@ namespace las
         cls[cl]++;
     }
   }
+  LAS_INLINE CSR::CSR(int r, int c, int nnz, std::vector<int> const & rs, std::vector<int> const & cs)
+    : nr(r)
+    , nc(c)
+    , nnz(nnz)
+    , rws(rs)
+    , cls(cs)
+  {
+  }
   LAS_INLINE Sparsity * csrFromArray(int rws, int cls, int nnz, int * row_arr, int * col_arr)
   {
     return reinterpret_cast<Sparsity*>(new CSR(rws,cls,nnz,row_arr,col_arr));

src/lasPETSc.h

@@ -17,6 +17,7 @@ namespace las
   Solve * createPetscQNSolve(void * a);
   MatVecMult * createPetscMatVecMult();
   MatMatMult * createPetscMatMatMult();
+  ScalarMatMult * createPetscScalarMatMult();
   template <>
   void finalizeMatrix<petsc>(Mat * mat);
   template <>

src/lasPETSc_impl.h

@@ -482,6 +482,28 @@ namespace las
   {
     return new PetscMatMatMult;
   }
+  class PetscScalarMatMult : public ScalarMatMult
+  {
+    public:
+    virtual void exec(scalar s, Mat * a, Mat ** c)
+    {
+      if (c == nullptr)
+      {
+        PetscErrorCode ierr = ::MatScale(*getPetscMat(a), s);
+      }
+      else
+      {
+        std::cerr << "Out of place matrix scalar multiplication not "
+                     "implemented in petsc"
+                  << std::endl;
+        std::abort();
+      }
+    }
+  };
+  LAS_INLINE ScalarMatMult * createPetscScalarMatMult()
+  {
+    return new PetscScalarMatMult;
+  }
   template <>
   LAS_INLINE void finalizeMatrix<petsc>(Mat * mat)
   {