C******************************************************************************
C FILE: loop10.f
C DESCRIPTION:
C   Demonstrates loop collapse
C******************************************************************************

	program loop10

	IMPLICIT NONE
	integer ARRAY_SIZE, DIM1, DIM2, DIM3
	parameter(ARRAY_SIZE = 1000)
	parameter(DIM1 = 500)
	parameter(DIM2 = 80)
	parameter(DIM3 = 40)
	integer	I, J, K
        real A(DIM1, DIM2, DIM3), AA(DIM1, DIM2, DIM3)
        real B(DIM1, DIM2, DIM3), BB(DIM1, DIM2, DIM3)
        real C(DIM1, DIM2, DIM3), CC(DIM1, DIM2, DIM3)
        integer(8) T1, T2, IRTC

C*****************************************************************************
C*				Uncollapsed Loop			     *
C*****************************************************************************
        T1 = IRTC()

        DO K=1, DIM3
          DO J=1, DIM2
            DO I=1, DIM1
              A(I, J, K) = A(I, J, K) + B(I, J, K) * C(I, J, K)
            ENDDO
          ENDDO
        ENDDO

        T2 = IRTC()
        write(*,*)'Untuned loop took', (T2-T1)/1000000, 'msec'

C*****************************************************************************
C*				Collapsed Loop				     *
C*****************************************************************************
        T1 = IRTC()

        DO I=1, (DIM1 * DIM2 * DIM3)
          AA(I, 1, 1) = AA(I, 1, 1) + BB(I, 1, 1) * CC(I, 1, 1)
        ENDDO

        T2 = IRTC()
        write(*,*)'Tuned loop took', (T2-T1)/1000000, 'msec'

C       Need to actually use the results of the calculations or else the
C       optimizing compiler may just skip doing them...so we'll just
C       print them.  This will force the optimizer to actually do the work.
        print *, A(DIM1,DIM2,DIM3), AA(DIM1,DIM2,DIM3)

999   end