Solution

marks 5 (1 for each line except processors declaration)

   !HPF$ ALIGN START(:) WITH A(*,:)
   !HPF$ ALIGN SWAP(:)  WITH A(*,:)
   !HPF$ ALIGN Y(:)     WITH A(:,*)
   !HPF$ ALIGN X(:)     WITH A(*,:)
   !HPF$ PROCESSORS P(16,16) ! say
   !HPF$ DISTRIBUTE A(CYCLIC,CYCLIC) ONTO P

marks 18 (2 per line)

Assume here that scalars are replicated:

• location_array = MAXLOC( ABS(A(i:n,i)) ) reduction operation (therefore communication between rows to establish where the largest element is) and broadcast (to all processors) with MAXLOC, (assume location_array is replicated),
• verb+location = ..+ no communications as RHS and LHS are both replicated
• SWAP is replicated along dimension 1 so information must be broadcast between rows.
• broadcast of TEMP to all processors,
• A(location,i:n) = A(i,i:n) communication in first dimension of A
• A(i,i:n) = SWAP(i:n} no comms
• Y(location) = Y(i) gives a one element communication,
• Y(i) = TEMP no comms as TEMP is replicated.