C	!!====================== mpi_scatter_reduce01.f ======================
C	!!
C	!! Master distributes (scatters) an array across proccesses.
C	!! Processes add their elements, then combine sum in master
C	!! through a reduction operation
C	!!====================================================================
	PROGRAM parsum
	INCLUDE "mpif.h"

C	MyPE: My logical process number
C	NumPE: Number of processes
C	i: loop index variable
C	A(16): Global array
C	MyA(4): My portion of A
C	MySum: My partial sum
C	Sum: Total sum

	INTEGER MyPE, NumPE, i,	error_code
	DOUBLE PRECISION A(16), MyA(4), MySum, Sum

	CALL MPI_INIT(error_code)
	CALL MPI_COMM_RANK(MPI_COMM_WORLD, MyPE, error_code)
	CALL MPI_COMM_SIZE(MPI_COMM_WORLD, NumPE, error_code)

C	Master fill array (analogous to a read operation)
	IF(MyPE == 0) THEN
		DO i=0, 16
			A(i) = i
		END DO
	END IF

C	Scatter array among all processes - each process gets
C	4 contiguous entries of A
	CALL MPI_SCATTER(A, 4, MPI_DOUBLE_PRECISION, MyA, 4, MPI_DOUBLE_PRECISION, 0, MPI_COMM_WORLD, error_code)
	WRITE(6,*) "PE ", MyPE, ": MyA is - ", (MyA(i),i=1,4)

C	Calculate sums, then reduce to master process
	MySum = 0.0
	DO i=1,4
		MySum = MySum+MyA(i)
	END DO

	CALL MPI_REDUCE(MySum, Sum, 1, MPI_DOUBLE_PRECISION, MPI_SUM, 0, MPI_COMM_WORLD, error_code)
	IF(MyPE == 0) THEN
		WRITE(6,*) "Total Sum is: ",Sum
	ENDIF
	CALL MPI_FINALIZE(error_code)
	END Program parsum