eliminate calls to getColumn() and putColumn() in GMRES solver

[alecjohnson]

In GMRES(), a matrix V is defined with dimensions xkrylovlen by (m+1), where
m is the number of iterations per restart and is chosen in the code to be 20.
The columns of this matrix are accessed by calls to getColumn() and putColumn(), which gather/scatter the data from/to a vector. This is likely to be expensive, since the large size of the Krylov vectors means that these accesses will usually be cache misses; such scatter/gather can also be a computational bottleneck when a vector unit is available. Transposing V would eliminate the need for any copying of data. We can then make v and w simple pointers:

  // don't need this any more:
  //double *v = new double[xkrylovlen];
  //double *w = new double[xkrylovlen];
  double *v,*w;

and change the getColumn() calls to set the pointer:

      // don't need this any more:
      //getColumn(v, V, k, xkrylovlen);
      v = V[k]; // use this instead

and simply delete the putColumn() calls.

This change also allows us to vectorize the use of V to modify the Krylov vector at the end of the loop:

    // the new code
    //
    for (register int j = 0; j < k; j++)
    {
      const double yj = y[j];
      double* Vj = V[j];
      // this will vectorize nicely
      for (int i = 0; i < xkrylovlen; i++)
        xkrylov[i] += yj * Vj[i];
    }

    // this was the old code.
    //
    //for (int jj = 0; jj < xkrylovlen; jj++) {
    //  tmp = 0.0;
    //  for (register int l = 0; l < k; l++)
    //    tmp += y[l] * V[l][jj];
    //  xkrylov[jj] += tmp;
    //}

[alecjohnson]

On my laptop this accelerated the computation part of the field solve by about 14 percent.