! Tuned F90 version  NCCS = 19
! - Stencil put in subroutine, called 3 times to cycle arrays
! - Rely on compiler to produce an efficient tiling

SUBROUTINE echo_f90_tuned(N, niters, check)
  INTEGER, INTENT( IN ) :: N, niters
  REAL, INTENT( OUT ) :: check

  REAL, DIMENSION (N,N) :: P1, P2, P3, c
  INTEGER iter

  CALL echo_f90_tuned_setup(c, P1, P2, P3, N)

  IF (MODULO(niters, 3) > 0) THEN
    PRINT *, 'niters should be divisible by 3, results will be off'
  ENDIF

  DO iter=1, niters, 3
    CALL stencil_f90(c, P1, P2, P3, N)
    CALL stencil_f90(c, P2, P3, P1, N)
    CALL stencil_f90(c, P3, P1, P2, N)
  END DO

  check = P1(N/2,7*N/8)

  RETURN
END



SUBROUTINE stencil_f90(c, P1, P2, P3, N)
  INTEGER, INTENT( IN ) :: N
  REAL, DIMENSION(N,N), INTENT( INOUT ) :: c, P1, P2, P3

  P3(2:N-1,2:N-1) = (2-4*c(2:N-1,2:N-1)) * P2(2:N-1,2:N-1) &
      + c(2:N-1,2:N-1)*(P2(1:N-2,2:N-1) + P2(3:N,2:N-1)    &
      + P2(2:N-1,1:N-2) + P2(2:N-1,3:N)) - P1(2:N-1,2:N-1)

  RETURN
END