blockmat/seq/blockmat.c

#define PETSCMAT_DLL

/*
   This provides a matrix that consists of Mats
*/

#include "src/mat/matimpl.h"              /*I "petscmat.h" I*/
#include "src/mat/impls/baij/seq/baij.h"    /* use the common AIJ data-structure */
#include "petscksp.h"

#define CHUNKSIZE   15

typedef struct {
  SEQAIJHEADER(Mat);
  SEQBAIJHEADER;
  Mat               *diags;

  Vec               left,right,middle,workb;   /* dummy vectors to perform local parts of product */
} Mat_BlockMat;

#undef __FUNCT__
#define __FUNCT__ "MatRelax_BlockMat_Symmetric"
PetscErrorCode MatRelax_BlockMat_Symmetric(Mat A,Vec bb,PetscReal omega,MatSORType flag,PetscReal fshift,PetscInt its,PetscInt lits,Vec xx)
{
  Mat_BlockMat       *a = (Mat_BlockMat*)A->data;
  PetscScalar        *x;
  const Mat          *v = a->a;
  const PetscScalar  *b;
  PetscErrorCode     ierr;
  PetscInt           n = A->cmap.n,i,mbs = n/A->rmap.bs,j,bs = A->rmap.bs;
  const PetscInt     *idx;
  IS                 row,col;
  MatFactorInfo      info;
  Vec                left = a->left,right = a->right, middle = a->middle;
  Mat                *diag;

  PetscFunctionBegin;
  its = its*lits;
  if (its <= 0) SETERRQ2(PETSC_ERR_ARG_WRONG,"Relaxation requires global its %D and local its %D both positive",its,lits);
  if (flag & SOR_EISENSTAT) SETERRQ(PETSC_ERR_SUP,"No support yet for Eisenstat");
  if (omega != 1.0) SETERRQ(PETSC_ERR_SUP,"No support yet for omega not equal to 1.0");
  if (fshift) SETERRQ(PETSC_ERR_SUP,"No support yet for fshift");
  if ((flag & SOR_BACKWARD_SWEEP || flag & SOR_LOCAL_BACKWARD_SWEEP) && !(flag & SOR_FORWARD_SWEEP || flag & SOR_LOCAL_FORWARD_SWEEP))
    SETERRQ(PETSC_ERR_SUP,"Cannot do backward sweep without forward sweep");

  if (!a->diags) {
    ierr = PetscMalloc(mbs*sizeof(Mat),&a->diags);CHKERRQ(ierr);
    ierr = MatFactorInfoInitialize(&info);CHKERRQ(ierr);
    for (i=0; i<mbs; i++) {
      ierr = MatGetOrdering(a->a[a->diag[i]], MATORDERING_ND,&row,&col);CHKERRQ(ierr);
      ierr = MatCholeskyFactorSymbolic(a->a[a->diag[i]],row,&info,a->diags+i);CHKERRQ(ierr);
      ierr = MatCholeskyFactorNumeric(a->a[a->diag[i]],&info,a->diags+i);CHKERRQ(ierr);
      ierr = ISDestroy(row);CHKERRQ(ierr);
      ierr = ISDestroy(col);CHKERRQ(ierr);
    }
    ierr = VecDuplicate(bb,&a->workb);CHKERRQ(ierr);
  }
  diag    = a->diags;

  ierr = VecSet(xx,0.0);CHKERRQ(ierr);
  ierr = VecGetArray(xx,&x);CHKERRQ(ierr);
  /* copy right hand side because it must be modified during iteration */
  ierr = VecCopy(bb,a->workb);CHKERRQ(ierr);
  ierr = VecGetArray(a->workb,(PetscScalar**)&b);CHKERRQ(ierr);

  /* need to add code for when initial guess is zero, see MatRelax_SeqAIJ */
  while (its--) {
    if (flag & SOR_FORWARD_SWEEP || flag & SOR_LOCAL_FORWARD_SWEEP){

      for (i=0; i<mbs; i++) {
        n    = a->i[i+1] - a->i[i] - 1;
        idx  = a->j + a->i[i] + 1;
        v    = a->a + a->i[i] + 1;

        ierr = VecSet(left,0.0);CHKERRQ(ierr);
        for (j=0; j<n; j++) {
          ierr = VecPlaceArray(right,x + idx[j]*bs);CHKERRQ(ierr);
          ierr = MatMultAdd(v[j],right,left,left);CHKERRQ(ierr);
          ierr = VecResetArray(right);CHKERRQ(ierr);
        }
        ierr = VecPlaceArray(right,b + i*bs);CHKERRQ(ierr);
        ierr = VecAYPX(left,-1.0,right);CHKERRQ(ierr);
        ierr = VecResetArray(right);CHKERRQ(ierr);

        ierr = VecPlaceArray(right,x + i*bs);CHKERRQ(ierr);
        ierr = MatSolve(diag[i],left,right);CHKERRQ(ierr);

        /* now adjust right hand side, see MatRelax_SeqSBAIJ */
        for (j=0; j<n; j++) {
          ierr = MatMultTranspose(v[j],right,left);CHKERRQ(ierr);
          ierr = VecPlaceArray(middle,b + idx[j]*bs);CHKERRQ(ierr);
          ierr = VecAXPY(middle,-1.0,left);CHKERRQ(ierr);
          ierr = VecResetArray(middle);CHKERRQ(ierr);
        }
        ierr = VecResetArray(right);CHKERRQ(ierr);

      }
    }
    if (flag & SOR_BACKWARD_SWEEP || flag & SOR_LOCAL_BACKWARD_SWEEP){

      for (i=mbs-1; i>=0; i--) {
        n    = a->i[i+1] - a->i[i] - 1;
        idx  = a->j + a->i[i] + 1;
        v    = a->a + a->i[i] + 1;

        ierr = VecSet(left,0.0);CHKERRQ(ierr);
        for (j=0; j<n; j++) {
          ierr = VecPlaceArray(right,x + idx[j]*bs);CHKERRQ(ierr);
          ierr = MatMultAdd(v[j],right,left,left);CHKERRQ(ierr);
          ierr = VecResetArray(right);CHKERRQ(ierr);
        }
        ierr = VecPlaceArray(right,b + i*bs);CHKERRQ(ierr);
        ierr = VecAYPX(left,-1.0,right);CHKERRQ(ierr);
        ierr = VecResetArray(right);CHKERRQ(ierr);

        ierr = VecPlaceArray(right,x + i*bs);CHKERRQ(ierr);
        ierr = MatSolve(diag[i],left,right);CHKERRQ(ierr);
        ierr = VecResetArray(right);CHKERRQ(ierr);

      }
    }
  }
  ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr);
  ierr = VecRestoreArray(a->workb,(PetscScalar**)&b);CHKERRQ(ierr);
  PetscFunctionReturn(0);
}

#undef __FUNCT__
#define __FUNCT__ "MatRelax_BlockMat"
PetscErrorCode MatRelax_BlockMat(Mat A,Vec bb,PetscReal omega,MatSORType flag,PetscReal fshift,PetscInt its,PetscInt lits,Vec xx)
{
  Mat_BlockMat       *a = (Mat_BlockMat*)A->data;
  PetscScalar        *x;
  const Mat          *v = a->a;
  const PetscScalar  *b;
  PetscErrorCode     ierr;
  PetscInt           n = A->cmap.n,i,mbs = n/A->rmap.bs,j,bs = A->rmap.bs;
  const PetscInt     *idx;
  IS                 row,col;
  MatFactorInfo      info;
  Vec                left = a->left,right = a->right;
  Mat                *diag;

  PetscFunctionBegin;
  its = its*lits;
  if (its <= 0) SETERRQ2(PETSC_ERR_ARG_WRONG,"Relaxation requires global its %D and local its %D both positive",its,lits);
  if (flag & SOR_EISENSTAT) SETERRQ(PETSC_ERR_SUP,"No support yet for Eisenstat");
  if (omega != 1.0) SETERRQ(PETSC_ERR_SUP,"No support yet for omega not equal to 1.0");
  if (fshift) SETERRQ(PETSC_ERR_SUP,"No support yet for fshift");

  if (!a->diags) {
    ierr = PetscMalloc(mbs*sizeof(Mat),&a->diags);CHKERRQ(ierr);
    ierr = MatFactorInfoInitialize(&info);CHKERRQ(ierr);
    for (i=0; i<mbs; i++) {
      ierr = MatGetOrdering(a->a[a->diag[i]], MATORDERING_ND,&row,&col);CHKERRQ(ierr);
      ierr = MatLUFactorSymbolic(a->a[a->diag[i]],row,col,&info,a->diags+i);CHKERRQ(ierr);
      ierr = MatLUFactorNumeric(a->a[a->diag[i]],&info,a->diags+i);CHKERRQ(ierr);
      ierr = ISDestroy(row);CHKERRQ(ierr);
      ierr = ISDestroy(col);CHKERRQ(ierr);
    }
  }
  diag = a->diags;

  ierr = VecSet(xx,0.0);CHKERRQ(ierr);
  ierr = VecGetArray(xx,&x);CHKERRQ(ierr);
  ierr = VecGetArray(bb,(PetscScalar**)&b);CHKERRQ(ierr);

  /* need to add code for when initial guess is zero, see MatRelax_SeqAIJ */
  while (its--) {
    if (flag & SOR_FORWARD_SWEEP || flag & SOR_LOCAL_FORWARD_SWEEP){

      for (i=0; i<mbs; i++) {
        n    = a->i[i+1] - a->i[i];
        idx  = a->j + a->i[i];
        v    = a->a + a->i[i];

        ierr = VecSet(left,0.0);CHKERRQ(ierr);
        for (j=0; j<n; j++) {
          if (idx[j] != i) {
            ierr = VecPlaceArray(right,x + idx[j]*bs);CHKERRQ(ierr);
            ierr = MatMultAdd(v[j],right,left,left);CHKERRQ(ierr);
            ierr = VecResetArray(right);CHKERRQ(ierr);
          }
        }
        ierr = VecPlaceArray(right,b + i*bs);CHKERRQ(ierr);
        ierr = VecAYPX(left,-1.0,right);CHKERRQ(ierr);
        ierr = VecResetArray(right);CHKERRQ(ierr);

        ierr = VecPlaceArray(right,x + i*bs);CHKERRQ(ierr);
        ierr = MatSolve(diag[i],left,right);CHKERRQ(ierr);
        ierr = VecResetArray(right);CHKERRQ(ierr);
      }
    }
    if (flag & SOR_BACKWARD_SWEEP || flag & SOR_LOCAL_BACKWARD_SWEEP){

      for (i=mbs-1; i>=0; i--) {
        n    = a->i[i+1] - a->i[i];
        idx  = a->j + a->i[i];
        v    = a->a + a->i[i];

        ierr = VecSet(left,0.0);CHKERRQ(ierr);
        for (j=0; j<n; j++) {
          if (idx[j] != i) {
            ierr = VecPlaceArray(right,x + idx[j]*bs);CHKERRQ(ierr);
            ierr = MatMultAdd(v[j],right,left,left);CHKERRQ(ierr);
            ierr = VecResetArray(right);CHKERRQ(ierr);
          }
        }
        ierr = VecPlaceArray(right,b + i*bs);CHKERRQ(ierr);
        ierr = VecAYPX(left,-1.0,right);CHKERRQ(ierr);
        ierr = VecResetArray(right);CHKERRQ(ierr);

        ierr = VecPlaceArray(right,x + i*bs);CHKERRQ(ierr);
        ierr = MatSolve(diag[i],left,right);CHKERRQ(ierr);
        ierr = VecResetArray(right);CHKERRQ(ierr);

      }
    }
  }
  ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr);
  ierr = VecRestoreArray(bb,(PetscScalar**)&b);CHKERRQ(ierr);
  PetscFunctionReturn(0);
}

#undef __FUNCT__
#define __FUNCT__ "MatSetValues_BlockMat"
PetscErrorCode MatSetValues_BlockMat(Mat A,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode is)
{
  Mat_BlockMat   *a = (Mat_BlockMat*)A->data;
  PetscInt       *rp,k,low,high,t,ii,row,nrow,i,col,l,rmax,N,lastcol = -1;
  PetscInt       *imax=a->imax,*ai=a->i,*ailen=a->ilen;
  PetscInt       *aj=a->j,nonew=a->nonew,bs=A->rmap.bs,brow,bcol;
  PetscErrorCode ierr;
  PetscInt       ridx,cidx;
  PetscTruth     roworiented=a->roworiented;
  MatScalar      value;
  Mat            *ap,*aa = a->a;

  PetscFunctionBegin;
  for (k=0; k<m; k++) { /* loop over added rows */
    row  = im[k];
    brow = row/bs;
    if (row < 0) continue;
#if defined(PETSC_USE_DEBUG)
    if (row >= A->rmap.N) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",row,A->rmap.N-1);
#endif
    rp   = aj + ai[brow];
    ap   = aa + ai[brow];
    rmax = imax[brow];
    nrow = ailen[brow];
    low  = 0;
    high = nrow;
    for (l=0; l<n; l++) { /* loop over added columns */
      if (in[l] < 0) continue;
#if defined(PETSC_USE_DEBUG)
      if (in[l] >= A->cmap.n) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %D max %D",in[l],A->cmap.n-1);
#endif
      col = in[l]; bcol = col/bs;
      if (A->symmetric && brow > bcol) continue;
      ridx = row % bs; cidx = col % bs;
      if (roworiented) {
        value = v[l + k*n];
      } else {
        value = v[k + l*m];
      }
      if (col <= lastcol) low = 0; else high = nrow;
      lastcol = col;
      while (high-low > 7) {
        t = (low+high)/2;
        if (rp[t] > bcol) high = t;
        else              low  = t;
      }
      for (i=low; i<high; i++) {
        if (rp[i] > bcol) break;
        if (rp[i] == bcol) {
          goto noinsert1;
        }
      }
      if (nonew == 1) goto noinsert1;
      if (nonew == -1) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero (%D, %D) in the matrix", row, col);
      MatSeqXAIJReallocateAIJ(A,a->mbs,1,nrow,brow,bcol,rmax,aa,ai,aj,rp,ap,imax,nonew,Mat);
      N = nrow++ - 1; high++;
      /* shift up all the later entries in this row */
      for (ii=N; ii>=i; ii--) {
        rp[ii+1] = rp[ii];
        ap[ii+1] = ap[ii];
      }
      if (N>=i) ap[i] = 0;
      rp[i]           = bcol;
      a->nz++;
      noinsert1:;
      if (!*(ap+i)) {
        if (A->symmetric && brow == bcol) {
          /* don't use SBAIJ since want to reorder in sparse factorization */
          ierr = MatCreateSeqAIJ(PETSC_COMM_SELF,bs,bs,0,0,ap+i);CHKERRQ(ierr);
        } else {
          ierr = MatCreateSeqAIJ(PETSC_COMM_SELF,bs,bs,0,0,ap+i);CHKERRQ(ierr);
        }
      }
      ierr = MatSetValues(ap[i],1,&ridx,1,&cidx,&value,is);CHKERRQ(ierr);
      low = i;
    }
    ailen[brow] = nrow;
  }
  A->same_nonzero = PETSC_FALSE;
  PetscFunctionReturn(0);
}

#undef __FUNCT__
#define __FUNCT__ "MatLoad_BlockMat"
PetscErrorCode MatLoad_BlockMat(PetscViewer viewer, MatType type,Mat *A)
{
  PetscErrorCode    ierr;
  Mat               tmpA;
  PetscInt          i,j,m,n,bs = 1,ncols,*lens,currentcol,mbs,**ii,*ilens,nextcol,*llens,cnt = 0;
  const PetscInt    *cols;
  const PetscScalar *values;
  PetscTruth        flg,notdone;
  Mat_SeqAIJ        *a;
  Mat_BlockMat      *amat;

  PetscFunctionBegin;
  ierr = MatLoad_SeqAIJ(viewer,MATSEQAIJ,&tmpA);CHKERRQ(ierr);

  ierr = MatGetLocalSize(tmpA,&m,&n);CHKERRQ(ierr);
  ierr = PetscOptionsBegin(PETSC_COMM_SELF,PETSC_NULL,"Options for loading BlockMat matrix 1","Mat");CHKERRQ(ierr);
    ierr = PetscOptionsInt("-matload_block_size","Set the blocksize used to store the matrix","MatLoad",bs,&bs,PETSC_NULL);CHKERRQ(ierr);
    ierr = PetscOptionsName("-matload_symmetric","Store the matrix as symmetric","MatLoad",&flg);CHKERRQ(ierr);
  ierr = PetscOptionsEnd();CHKERRQ(ierr);

  /* Determine number of nonzero blocks for each block row */
  a    = (Mat_SeqAIJ*) tmpA->data;
  mbs  = m/bs;
  ierr = PetscMalloc3(mbs,PetscInt,&lens,bs,PetscInt*,&ii,bs,PetscInt,&ilens);CHKERRQ(ierr);
  ierr = PetscMemzero(lens,mbs*sizeof(PetscInt));CHKERRQ(ierr);

  for (i=0; i<mbs; i++) {
    for (j=0; j<bs; j++) {
      ii[j]         = a->j + a->i[i*bs + j];
      ilens[j]      = a->i[i*bs + j + 1] - a->i[i*bs + j];
    }

    currentcol = -1;
    notdone = PETSC_TRUE;
    while (PETSC_TRUE) {
      notdone = PETSC_FALSE;
      nextcol = 1000000000;
      for (j=0; j<bs; j++) {
        while ((ilens[j] > 0 && ii[j][0]/bs <= currentcol)) {
          ii[j]++;
          ilens[j]--;
        }
        if (ilens[j] > 0) {
          notdone = PETSC_TRUE;
          nextcol = PetscMin(nextcol,ii[j][0]/bs);
        }
      }
      if (!notdone) break;
      if (!flg || (nextcol >= i)) lens[i]++;
      currentcol = nextcol;
    }
  }

  ierr = MatCreateBlockMat(PETSC_COMM_SELF,m,n,bs,0,lens,A);CHKERRQ(ierr);
  if (flg) {
    ierr = MatSetOption(*A,MAT_SYMMETRIC);CHKERRQ(ierr);
  }
  amat = (Mat_BlockMat*)(*A)->data;

  /* preallocate the submatrices */
  ierr = PetscMalloc(bs*sizeof(PetscInt),&llens);CHKERRQ(ierr);
  for (i=0; i<mbs; i++) { /* loops for block rows */
    for (j=0; j<bs; j++) {
      ii[j]         = a->j + a->i[i*bs + j];
      ilens[j]      = a->i[i*bs + j + 1] - a->i[i*bs + j];
    }

    currentcol = 1000000000;
    for (j=0; j<bs; j++) { /* loop over rows in block finding first nonzero block */
      if (ilens[j] > 0) {
	currentcol = PetscMin(currentcol,ii[j][0]/bs);
      }
    }

    notdone = PETSC_TRUE;
    while (PETSC_TRUE) {  /* loops over blocks in block row */

      notdone = PETSC_FALSE;
      nextcol = 1000000000;
      ierr = PetscMemzero(llens,bs*sizeof(PetscInt));CHKERRQ(ierr);
      for (j=0; j<bs; j++) { /* loop over rows in block */
        while ((ilens[j] > 0 && ii[j][0]/bs <= currentcol)) { /* loop over columns in row */
          ii[j]++;
          ilens[j]--;
          llens[j]++;
        }
        if (ilens[j] > 0) {
          notdone = PETSC_TRUE;
          nextcol = PetscMin(nextcol,ii[j][0]/bs);
        }
      }
      if (cnt >= amat->maxnz) SETERRQ1(PETSC_ERR_PLIB,"Number of blocks found greater than expected %D",cnt);
      if (!flg || currentcol >= i) {
        amat->j[cnt] = currentcol;
        ierr = MatCreateSeqAIJ(PETSC_COMM_SELF,bs,bs,0,llens,amat->a+cnt++);CHKERRQ(ierr);
      }

      if (!notdone) break;
      currentcol = nextcol;
    }
    amat->ilen[i] = lens[i];
  }
  CHKMEMQ;

  ierr = PetscFree3(lens,ii,ilens);CHKERRQ(ierr);
  ierr = PetscFree(llens);CHKERRQ(ierr);

  /* copy over the matrix, one row at a time */
  for (i=0; i<m; i++) {
    ierr = MatGetRow(tmpA,i,&ncols,&cols,&values);CHKERRQ(ierr);
    ierr = MatSetValues(*A,1,&i,ncols,cols,values,INSERT_VALUES);CHKERRQ(ierr);
    ierr = MatRestoreRow(tmpA,i,&ncols,&cols,&values);CHKERRQ(ierr);
  }
  ierr = MatAssemblyBegin(*A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
  ierr = MatAssemblyEnd(*A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
  PetscFunctionReturn(0);
}

#undef __FUNCT__
#define __FUNCT__ "MatView_BlockMat"
PetscErrorCode MatView_BlockMat(Mat A,PetscViewer viewer)
{
  Mat_SeqAIJ        *a = (Mat_SeqAIJ*)A->data;
  PetscErrorCode    ierr;
  const char        *name;
  PetscViewerFormat format;

  PetscFunctionBegin;
  ierr = PetscObjectGetName((PetscObject)A,&name);CHKERRQ(ierr);
  ierr = PetscViewerGetFormat(viewer,&format);CHKERRQ(ierr);
  if (format == PETSC_VIEWER_ASCII_FACTOR_INFO || format == PETSC_VIEWER_ASCII_INFO) {
    ierr = PetscViewerASCIIPrintf(viewer,"Nonzero block matrices = %D \n",a->nz);CHKERRQ(ierr);
  }
  PetscFunctionReturn(0);
}

#undef __FUNCT__
#define __FUNCT__ "MatDestroy_BlockMat"
PetscErrorCode MatDestroy_BlockMat(Mat mat)
{
  PetscErrorCode ierr;
  Mat_BlockMat   *bmat = (Mat_BlockMat*)mat->data;
  PetscInt       i;

  PetscFunctionBegin;
  if (bmat->right) {
    ierr = VecDestroy(bmat->right);CHKERRQ(ierr);
  }
  if (bmat->left) {
    ierr = VecDestroy(bmat->left);CHKERRQ(ierr);
  }
  if (bmat->middle) {
    ierr = VecDestroy(bmat->middle);CHKERRQ(ierr);
  }
  if (bmat->workb) {
    ierr = VecDestroy(bmat->workb);CHKERRQ(ierr);
  }
  if (bmat->diags) {
    for (i=0; i<mat->rmap.n/mat->rmap.bs; i++) {
      if (bmat->diags[i]) {ierr = MatDestroy(bmat->diags[i]);CHKERRQ(ierr);}
    }
  }
  if (bmat->a) {
    for (i=0; i<bmat->nz; i++) {
      if (bmat->a[i]) {ierr = MatDestroy(bmat->a[i]);CHKERRQ(ierr);}
    }
  }
  ierr = MatSeqXAIJFreeAIJ(mat,(PetscScalar**)&bmat->a,&bmat->j,&bmat->i);CHKERRQ(ierr);
  ierr = PetscFree(bmat);CHKERRQ(ierr);
  PetscFunctionReturn(0);
}

#undef __FUNCT__
#define __FUNCT__ "MatMult_BlockMat"
PetscErrorCode MatMult_BlockMat(Mat A,Vec x,Vec y)
{
  Mat_BlockMat   *bmat = (Mat_BlockMat*)A->data;
  PetscErrorCode ierr;
  PetscScalar    *xx,*yy;
  PetscInt       *aj,i,*ii,jrow,m = A->rmap.n/A->rmap.bs,bs = A->rmap.bs,n,j;
  Mat            *aa;

  PetscFunctionBegin;
  CHKMEMQ;
  /*
     Standard CSR multiply except each entry is a Mat
  */
  ierr = VecGetArray(x,&xx);CHKERRQ(ierr);

  ierr = VecSet(y,0.0);CHKERRQ(ierr);
  ierr = VecGetArray(y,&yy);CHKERRQ(ierr);
  aj  = bmat->j;
  aa  = bmat->a;
  ii  = bmat->i;
  for (i=0; i<m; i++) {
    jrow = ii[i];
    ierr = VecPlaceArray(bmat->left,yy + bs*i);CHKERRQ(ierr);
    n    = ii[i+1] - jrow;
    for (j=0; j<n; j++) {
      ierr = VecPlaceArray(bmat->right,xx + bs*aj[jrow]);CHKERRQ(ierr);
      ierr = MatMultAdd(aa[jrow],bmat->right,bmat->left,bmat->left);CHKERRQ(ierr);
      ierr = VecResetArray(bmat->right);CHKERRQ(ierr);
      jrow++;
    }
    ierr = VecResetArray(bmat->left);CHKERRQ(ierr);
  }
  ierr = VecRestoreArray(x,&xx);CHKERRQ(ierr);
  ierr = VecRestoreArray(y,&yy);CHKERRQ(ierr);
  CHKMEMQ;
  PetscFunctionReturn(0);
}

#undef __FUNCT__
#define __FUNCT__ "MatMult_BlockMat_Symmetric"
PetscErrorCode MatMult_BlockMat_Symmetric(Mat A,Vec x,Vec y)
{
  Mat_BlockMat   *bmat = (Mat_BlockMat*)A->data;
  PetscErrorCode ierr;
  PetscScalar    *xx,*yy;
  PetscInt       *aj,i,*ii,jrow,m = A->rmap.n/A->rmap.bs,bs = A->rmap.bs,n,j;
  Mat            *aa;

  PetscFunctionBegin;
  CHKMEMQ;
  /*
     Standard CSR multiply except each entry is a Mat
  */
  ierr = VecGetArray(x,&xx);CHKERRQ(ierr);

  ierr = VecSet(y,0.0);CHKERRQ(ierr);
  ierr = VecGetArray(y,&yy);CHKERRQ(ierr);
  aj  = bmat->j;
  aa  = bmat->a;
  ii  = bmat->i;
  for (i=0; i<m; i++) {
    jrow = ii[i];
    n    = ii[i+1] - jrow;
    ierr = VecPlaceArray(bmat->left,yy + bs*i);CHKERRQ(ierr);
    ierr = VecPlaceArray(bmat->middle,xx + bs*i);CHKERRQ(ierr);
    /* if we ALWAYS required a diagonal entry then could remove this if test */
    if (aj[jrow] == i) {
      ierr = VecPlaceArray(bmat->right,xx + bs*aj[jrow]);CHKERRQ(ierr);
      ierr = MatMultAdd(aa[jrow],bmat->right,bmat->left,bmat->left);CHKERRQ(ierr);
      ierr = VecResetArray(bmat->right);CHKERRQ(ierr);
      jrow++;
      n--;
    }
    for (j=0; j<n; j++) {
      ierr = VecPlaceArray(bmat->right,xx + bs*aj[jrow]);CHKERRQ(ierr);            /* upper triangular part */
      ierr = MatMultAdd(aa[jrow],bmat->right,bmat->left,bmat->left);CHKERRQ(ierr);
      ierr = VecResetArray(bmat->right);CHKERRQ(ierr);

      ierr = VecPlaceArray(bmat->right,yy + bs*aj[jrow]);CHKERRQ(ierr);            /* lower triangular part */
      ierr = MatMultTransposeAdd(aa[jrow],bmat->middle,bmat->right,bmat->right);CHKERRQ(ierr);
      ierr = VecResetArray(bmat->right);CHKERRQ(ierr);
      jrow++;
    }
    ierr = VecResetArray(bmat->left);CHKERRQ(ierr);
    ierr = VecResetArray(bmat->middle);CHKERRQ(ierr);
  }
  ierr = VecRestoreArray(x,&xx);CHKERRQ(ierr);
  ierr = VecRestoreArray(y,&yy);CHKERRQ(ierr);
  CHKMEMQ;
  PetscFunctionReturn(0);
}

#undef __FUNCT__
#define __FUNCT__ "MatMultAdd_BlockMat"
PetscErrorCode MatMultAdd_BlockMat(Mat A,Vec x,Vec y,Vec z)
{
  PetscFunctionBegin;
  PetscFunctionReturn(0);
}

#undef __FUNCT__
#define __FUNCT__ "MatMultTranspose_BlockMat"
PetscErrorCode MatMultTranspose_BlockMat(Mat A,Vec x,Vec y)
{
  PetscFunctionBegin;
  PetscFunctionReturn(0);
}

#undef __FUNCT__
#define __FUNCT__ "MatMultTransposeAdd_BlockMat"
PetscErrorCode MatMultTransposeAdd_BlockMat(Mat A,Vec x,Vec y,Vec z)
{
  PetscFunctionBegin;
  PetscFunctionReturn(0);
}

/*
     Adds diagonal pointers to sparse matrix structure.
*/
#undef __FUNCT__
#define __FUNCT__ "MatMarkDiagonal_BlockMat"
PetscErrorCode MatMarkDiagonal_BlockMat(Mat A)
{
  Mat_BlockMat   *a = (Mat_BlockMat*)A->data;
  PetscErrorCode ierr;
  PetscInt       i,j,mbs = A->rmap.n/A->rmap.bs;

  PetscFunctionBegin;
  if (!a->diag) {
    ierr = PetscMalloc(mbs*sizeof(PetscInt),&a->diag);CHKERRQ(ierr);
  }
  for (i=0; i<mbs; i++) {
    a->diag[i] = a->i[i+1];
    for (j=a->i[i]; j<a->i[i+1]; j++) {
      if (a->j[j] == i) {
        a->diag[i] = j;
        break;
      }
    }
  }
  PetscFunctionReturn(0);
}

#undef __FUNCT__
#define __FUNCT__ "MatGetSubMatrix_BlockMat"
PetscErrorCode MatGetSubMatrix_BlockMat(Mat A,IS isrow,IS iscol,PetscInt csize,MatReuse scall,Mat *B)
{
  Mat_BlockMat   *a = (Mat_BlockMat*)A->data;
  Mat_SeqAIJ     *c;
  PetscErrorCode ierr;
  PetscInt       i,k,first,step,lensi,nrows,ncols;
  PetscInt       *j_new,*i_new,*aj = a->j,*ai = a->i,ii,*ailen = a->ilen;
  PetscScalar    *a_new,value;
  Mat            C,*aa = a->a;
  PetscTruth     stride,equal;

  PetscFunctionBegin;
  ierr = ISEqual(isrow,iscol,&equal);CHKERRQ(ierr);
  if (!equal) SETERRQ(PETSC_ERR_SUP,"Only for idential column and row indices");
  ierr = ISStride(iscol,&stride);CHKERRQ(ierr);
  if (!stride) SETERRQ(PETSC_ERR_SUP,"Only for stride indices");
  ierr = ISStrideGetInfo(iscol,&first,&step);CHKERRQ(ierr);
  if (step != A->rmap.bs) SETERRQ(PETSC_ERR_SUP,"Can only select one entry from each block");

  ierr = ISGetLocalSize(isrow,&nrows);CHKERRQ(ierr);
  ncols = nrows;

  /* create submatrix */
  if (scall == MAT_REUSE_MATRIX) {
    PetscInt n_cols,n_rows;
    C = *B;
    ierr = MatGetSize(C,&n_rows,&n_cols);CHKERRQ(ierr);
    if (n_rows != nrows || n_cols != ncols) SETERRQ(PETSC_ERR_ARG_SIZ,"Reused submatrix wrong size");
    ierr = MatZeroEntries(C);CHKERRQ(ierr);
  } else {
    ierr = MatCreate(A->comm,&C);CHKERRQ(ierr);
    ierr = MatSetSizes(C,nrows,ncols,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr);
    if (A->symmetric) {
      ierr = MatSetType(C,MATSEQSBAIJ);CHKERRQ(ierr);
    } else {
      ierr = MatSetType(C,MATSEQAIJ);CHKERRQ(ierr);
    }
    ierr = MatSeqAIJSetPreallocation(C,0,ailen);CHKERRQ(ierr);
    ierr = MatSeqSBAIJSetPreallocation(C,1,0,ailen);CHKERRQ(ierr);
  }
  c = (Mat_SeqAIJ*)C->data;

  /* loop over rows inserting into submatrix */
  a_new    = c->a;
  j_new    = c->j;
  i_new    = c->i;

  for (i=0; i<nrows; i++) {
    ii    = ai[i];
    lensi = ailen[i];
    for (k=0; k<lensi; k++) {
      *j_new++ = *aj++;
      ierr     = MatGetValue(*aa++,first,first,value);CHKERRQ(ierr);
      *a_new++ = value;
    }
    i_new[i+1]  = i_new[i] + lensi;
    c->ilen[i]  = lensi;
  }

  ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
  ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
  *B = C;
  PetscFunctionReturn(0);
}

#undef __FUNCT__
#define __FUNCT__ "MatAssemblyEnd_BlockMat"
PetscErrorCode MatAssemblyEnd_BlockMat(Mat A,MatAssemblyType mode)
{
  Mat_BlockMat   *a = (Mat_BlockMat*)A->data;
  PetscErrorCode ierr;
  PetscInt       fshift = 0,i,j,*ai = a->i,*aj = a->j,*imax = a->imax;
  PetscInt       m = a->mbs,*ip,N,*ailen = a->ilen,rmax = 0;
  Mat            *aa = a->a,*ap;

  PetscFunctionBegin;
  if (mode == MAT_FLUSH_ASSEMBLY) PetscFunctionReturn(0);

  if (m) rmax = ailen[0]; /* determine row with most nonzeros */
  for (i=1; i<m; i++) {
    /* move each row back by the amount of empty slots (fshift) before it*/
    fshift += imax[i-1] - ailen[i-1];
    rmax   = PetscMax(rmax,ailen[i]);
    if (fshift) {
      ip = aj + ai[i] ;
      ap = aa + ai[i] ;
      N  = ailen[i];
      for (j=0; j<N; j++) {
        ip[j-fshift] = ip[j];
        ap[j-fshift] = ap[j];
      }
    }
    ai[i] = ai[i-1] + ailen[i-1];
  }
  if (m) {
    fshift += imax[m-1] - ailen[m-1];
    ai[m]  = ai[m-1] + ailen[m-1];
  }
  /* reset ilen and imax for each row */
  for (i=0; i<m; i++) {
    ailen[i] = imax[i] = ai[i+1] - ai[i];
  }
  a->nz = ai[m];
  for (i=0; i<a->nz; i++) {
#if defined(PETSC_USE_DEBUG)
    if (!aa[i]) SETERRQ3(PETSC_ERR_PLIB,"Null matrix at location %D column %D nz %D",i,aj[i],a->nz);
#endif
    ierr = MatAssemblyBegin(aa[i],MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
    ierr = MatAssemblyEnd(aa[i],MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
  }
  CHKMEMQ;
  ierr = PetscInfo4(A,"Matrix size: %D X %D; storage space: %D unneeded,%D used\n",m,A->cmap.n/A->cmap.bs,fshift,a->nz);CHKERRQ(ierr);
  ierr = PetscInfo1(A,"Number of mallocs during MatSetValues() is %D\n",a->reallocs);CHKERRQ(ierr);
  ierr = PetscInfo1(A,"Maximum nonzeros in any row is %D\n",rmax);CHKERRQ(ierr);
  a->reallocs          = 0;
  A->info.nz_unneeded  = (double)fshift;
  a->rmax              = rmax;

  A->same_nonzero = PETSC_TRUE;
  ierr = MatMarkDiagonal_BlockMat(A);CHKERRQ(ierr);
  PetscFunctionReturn(0);
}

#undef __FUNCT__
#define __FUNCT__ "MatSetOption_BlockMat"
PetscErrorCode MatSetOption_BlockMat(Mat A,MatOption opt)
{
  PetscFunctionBegin;
  if (opt == MAT_SYMMETRIC) {
    A->ops->relax = MatRelax_BlockMat_Symmetric;
    A->ops->mult  = MatMult_BlockMat_Symmetric;
  } else {
    PetscInfo1(A,"Unused matrix option %s\n",MatOptions[opt]);
  }
  PetscFunctionReturn(0);
}


static struct _MatOps MatOps_Values = {MatSetValues_BlockMat,
       0,
       0,
       MatMult_BlockMat,
/* 4*/ MatMultAdd_BlockMat,
       MatMultTranspose_BlockMat,
       MatMultTransposeAdd_BlockMat,
       0,
       0,
       0,
/*10*/ 0,
       0,
       0,
       MatRelax_BlockMat,
       0,
/*15*/ 0,
       0,
       0,
       0,
       0,
/*20*/ 0,
       MatAssemblyEnd_BlockMat,
       0,
       MatSetOption_BlockMat,
       0,
/*25*/ 0,
       0,
       0,
       0,
       0,
/*30*/ 0,
       0,
       0,
       0,
       0,
/*35*/ 0,
       0,
       0,
       0,
       0,
/*40*/ 0,
       0,
       0,
       0,
       0,
/*45*/ 0,
       0,
       0,
       0,
       0,
/*50*/ 0,
       0,
       0,
       0,
       0,
/*55*/ 0,
       0,
       0,
       0,
       0,
/*60*/ MatGetSubMatrix_BlockMat,
       MatDestroy_BlockMat,
       MatView_BlockMat,
       0,
       0,
/*65*/ 0,
       0,
       0,
       0,
       0,
/*70*/ 0,
       0,
       0,
       0,
       0,
/*75*/ 0,
       0,
       0,
       0,
       0,
/*80*/ 0,
       0,
       0,
       0,
       MatLoad_BlockMat,
/*85*/ 0,
       0,
       0,
       0,
       0,
/*90*/ 0,
       0,
       0,
       0,
       0,
/*95*/ 0,
       0,
       0,
       0};

#undef __FUNCT__
#define __FUNCT__ "MatBlockMatSetPreallocation"
/*@C
   MatBlockMatSetPreallocation - For good matrix assembly performance
   the user should preallocate the matrix storage by setting the parameter nz
   (or the array nnz).  By setting these parameters accurately, performance
   during matrix assembly can be increased by more than a factor of 50.

   Collective on MPI_Comm

   Input Parameters:
+  B - The matrix
.  bs - size of each block in matrix
.  nz - number of nonzeros per block row (same for all rows)
-  nnz - array containing the number of nonzeros in the various block rows
         (possibly different for each row) or PETSC_NULL

   Notes:
     If nnz is given then nz is ignored

   Specify the preallocated storage with either nz or nnz (not both).
   Set nz=PETSC_DEFAULT and nnz=PETSC_NULL for PETSc to control dynamic memory
   allocation.  For large problems you MUST preallocate memory or you
   will get TERRIBLE performance, see the users' manual chapter on matrices.

   Level: intermediate

.seealso: MatCreate(), MatCreateBlockMat(), MatSetValues()

@*/
PetscErrorCode PETSCMAT_DLLEXPORT MatBlockMatSetPreallocation(Mat B,PetscInt bs,PetscInt nz,const PetscInt nnz[])
{
  PetscErrorCode ierr,(*f)(Mat,PetscInt,PetscInt,const PetscInt[]);

  PetscFunctionBegin;
  ierr = PetscObjectQueryFunction((PetscObject)B,"MatBlockMatSetPreallocation_C",(void (**)(void))&f);CHKERRQ(ierr);
  if (f) {
    ierr = (*f)(B,bs,nz,nnz);CHKERRQ(ierr);
  }
  PetscFunctionReturn(0);
}

EXTERN_C_BEGIN
#undef __FUNCT__
#define __FUNCT__ "MatBlockMatSetPreallocation_BlockMat"
PetscErrorCode PETSCMAT_DLLEXPORT MatBlockMatSetPreallocation_BlockMat(Mat A,PetscInt bs,PetscInt nz,PetscInt *nnz)
{
  Mat_BlockMat   *bmat = (Mat_BlockMat*)A->data;
  PetscErrorCode ierr;
  PetscInt       i;

  PetscFunctionBegin;
  if (bs < 1) SETERRQ1(PETSC_ERR_ARG_OUTOFRANGE,"Block size given %D must be great than zero",bs);
  if (A->rmap.n % bs) SETERRQ2(PETSC_ERR_ARG_INCOMP,"Blocksize %D does not divide number of rows %D",bs,A->rmap.n);
  if (A->cmap.n % bs) SETERRQ2(PETSC_ERR_ARG_INCOMP,"Blocksize %D does not divide number of columns %D",bs,A->cmap.n);
  if (nz == PETSC_DEFAULT || nz == PETSC_DECIDE) nz = 5;
  if (nz < 0) SETERRQ1(PETSC_ERR_ARG_OUTOFRANGE,"nz cannot be less than 0: value %d",nz);
  if (nnz) {
    for (i=0; i<A->rmap.n/bs; i++) {
      if (nnz[i] < 0) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"nnz cannot be less than 0: local row %d value %d",i,nnz[i]);
      if (nnz[i] > A->cmap.n/bs) SETERRQ3(PETSC_ERR_ARG_OUTOFRANGE,"nnz cannot be greater than row length: local row %d value %d rowlength %d",i,nnz[i],A->cmap.n/bs);
    }
  }
  A->rmap.bs = A->cmap.bs = bs;
  bmat->mbs  = A->rmap.n/bs;

  ierr = VecCreateSeqWithArray(PETSC_COMM_SELF,bs,PETSC_NULL,&bmat->right);CHKERRQ(ierr);
  ierr = VecCreateSeqWithArray(PETSC_COMM_SELF,bs,PETSC_NULL,&bmat->middle);CHKERRQ(ierr);
  ierr = VecCreateSeq(PETSC_COMM_SELF,bs,&bmat->left);CHKERRQ(ierr);

  ierr = PetscMalloc2(A->rmap.n,PetscInt,&bmat->imax,A->rmap.n,PetscInt,&bmat->ilen);CHKERRQ(ierr);
  if (nnz) {
    nz = 0;
    for (i=0; i<A->rmap.n/A->rmap.bs; i++) {
      bmat->imax[i] = nnz[i];
      nz           += nnz[i];
    }
  } else {
    SETERRQ(PETSC_ERR_SUP,"Currently requires block row by row preallocation");
  }

  /* bmat->ilen will count nonzeros in each row so far. */
  for (i=0; i<bmat->mbs; i++) { bmat->ilen[i] = 0;}

  /* allocate the matrix space */
  ierr = PetscMalloc3(nz,Mat,&bmat->a,nz,PetscInt,&bmat->j,A->rmap.n+1,PetscInt,&bmat->i);CHKERRQ(ierr);
  bmat->i[0] = 0;
  for (i=1; i<bmat->mbs+1; i++) {
    bmat->i[i] = bmat->i[i-1] + bmat->imax[i-1];
  }
  bmat->singlemalloc = PETSC_TRUE;
  bmat->free_a       = PETSC_TRUE;
  bmat->free_ij      = PETSC_TRUE;

  bmat->nz                = 0;
  bmat->maxnz             = nz;
  A->info.nz_unneeded  = (double)bmat->maxnz;

  PetscFunctionReturn(0);
}
EXTERN_C_END

/*MC
   MATBLOCKMAT - A matrix that is defined by a set of Mat's that represents a sparse block matrix
                 consisting of (usually) sparse blocks.

  Level: advanced

.seealso: MatCreateBlockMat()

M*/

EXTERN_C_BEGIN
#undef __FUNCT__
#define __FUNCT__ "MatCreate_BlockMat"
PetscErrorCode PETSCMAT_DLLEXPORT MatCreate_BlockMat(Mat A)
{
  Mat_BlockMat   *b;
  PetscErrorCode ierr;

  PetscFunctionBegin;
  ierr = PetscMemcpy(A->ops,&MatOps_Values,sizeof(struct _MatOps));CHKERRQ(ierr);
  ierr = PetscNew(Mat_BlockMat,&b);CHKERRQ(ierr);

  A->data = (void*)b;

  ierr = PetscMapInitialize(A->comm,&A->rmap);CHKERRQ(ierr);
  ierr = PetscMapInitialize(A->comm,&A->cmap);CHKERRQ(ierr);

  A->assembled     = PETSC_TRUE;
  A->preallocated  = PETSC_FALSE;
  ierr = PetscObjectChangeTypeName((PetscObject)A,MATBLOCKMAT);CHKERRQ(ierr);

  ierr = PetscOptionsBegin(A->comm,A->prefix,"Matrix Option","Mat");CHKERRQ(ierr);
  ierr = PetscOptionsEnd();

  ierr = PetscObjectComposeFunctionDynamic((PetscObject)A,"MatBlockMatSetPreallocation_C",
                                     "MatBlockMatSetPreallocation_BlockMat",
                                      MatBlockMatSetPreallocation_BlockMat);CHKERRQ(ierr);

  PetscFunctionReturn(0);
}
EXTERN_C_END

#undef __FUNCT__
#define __FUNCT__ "MatCreateBlockMat"
/*@C
   MatCreateBlockMat - Creates a new matrix based sparse Mat storage

  Collective on MPI_Comm

   Input Parameters:
+  comm - MPI communicator
.  m - number of rows
.  n  - number of columns
.  bs - size of each submatrix
.  nz  - expected maximum number of nonzero blocks in row (use PETSC_DEFAULT if not known)
-  nnz - expected number of nonzers per block row if known (use PETSC_NULL otherwise)


   Output Parameter:
.  A - the matrix

   Level: intermediate

   PETSc requires that matrices and vectors being used for certain
   operations are partitioned accordingly.  For example, when
   creating a bmat matrix, A, that supports parallel matrix-vector
   products using MatMult(A,x,y) the user should set the number
   of local matrix rows to be the number of local elements of the
   corresponding result vector, y. Note that this is information is
   required for use of the matrix interface routines, even though
   the bmat matrix may not actually be physically partitioned.
   For example,

.keywords: matrix, bmat, create

.seealso: MATBLOCKMAT
@*/
PetscErrorCode PETSCMAT_DLLEXPORT MatCreateBlockMat(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt bs,PetscInt nz,PetscInt *nnz, Mat *A)
{
  PetscErrorCode ierr;

  PetscFunctionBegin;
  ierr = MatCreate(comm,A);CHKERRQ(ierr);
  ierr = MatSetSizes(*A,m,n,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr);
  ierr = MatSetType(*A,MATBLOCKMAT);CHKERRQ(ierr);
  ierr = MatBlockMatSetPreallocation(*A,bs,nz,nnz);CHKERRQ(ierr);
  PetscFunctionReturn(0);
}