PRACE NPT

#include "mpiargs.h"

void btrans( int n1, int n2, double *a, double **t, int dir )

// ---------------------------------------------------------------------

// --- 'btrans' performs a blockwise local transposition. Input is a

//      1-D array 'a', output is a 2-D array 't'

// ---------------------------------------------------------------------

{

   int i, j, k, l;

   int dim1, dim2, m1, m2, shift;

// ---------------------------------------------------------------------

   if ( dir == 0 ) {

      dim1 = 0;

      dim2 = 1;

   }

   else {

      dim1 = 1;

      dim2 = 0;

   }

   l = 0; m2 = sizes[me][dim2];

   for( k = 0; k < nodes; k++ ) {

      shift = offset[k][dim1];

      m1 = sizes[k][dim1];

      for( j = 0; j < m2; j++ ) {

         for( i = 0; i < m1; i++ ) {

            t[j][i+shift] = a[l];

            l++;

         }

      }

   }

}