hpf

1. History of HPF:

Evolution: Fortran 77  Fortran 90  HPF

HPFF series of meetings during March 1992 – March 1993 lead to HPF version 1.0

2. Data parallelism: Exploitation of the concurrency that derives from the application of the

same operation to multiple elements of a data structure.  Task: each operation on each element  Granularity: small  Locality: implicit

3. Required capabilities of data parallel language:

 Express concurrency  Express data distribution

4. Main features of HPF covered in this course:

 Array assignment and array intrinsic functions (7.2)

Example:

Integer A(10,10), B(10,10), c A = B + c ……

A(1,1:7) = B(2,2:8) + 1.0 Example:

Before the execution of the following statement, X= [1.0, 20.0, 300.0, 4000.0, 50000.0], X(2:4) = X(1:3) + X(2:4) + X(3:5) What will X be after the execution?

Answer: X = [1.0, 321.0, 4320.0, 300.0, 50000.0]

Note that the execution is not the same as the execution of the following do loop: Do I = 2, 4

X(I) = X(I-1) + X(I) + X(I+1)

END DO

Constrains: all operant must be conformable.  FORALL statement and construct (7.4.1)

FORALL statement:

FORALL (triplet, …, triplet, mask) assignment Example:

FORALL (I=1:m, j=1:n) X(I,j) = I+j FORALL (I=1:100) X(I,I) = 0.0 FORALL construct:

FORALL (triplet, …, triplet, mask) assignment \\

…… | forall-body

/

END FORALL Constrains:

1. Evaluate all elements in any order 2. Evaluate mask in any order

3. Evaluate the statements in the forall-body in the order they appear.

 INDEPENDENT directive (7.4.2)

The guidance provides by the programmer to the compiler to assert that the iterations of a do-loop can be performed independently. It could be used in implementation of parallel tasks as shown in the example below. Example:

!HPF$ INDEPENDENT

Do I = 1, 4

CALL task (i) END DO ……

SUBROUTINE task(i) Integer I

IF (I.eq.1) THEN CALL action1 ELSEIF (I.eq.2) THEN CALL action2 ELSEIF (I.eq.3) THEN CALL action3 ELSE

CALL action4 ENDIF Return end

 Data distribution directives (7.3)

Examples:

!HPF$ PROCESSORS prec(4)

!HPF$ ALIGN array( : ) WITH D( : , * )

!HPF$ DISTRIBUTE array(BLOCK) ONTO proc

Constrains:

Must specify all dimensions.

Distribution applies to all aligned arrays

Three ways of distribution could be mixed, but need to be conformable.

5. Reading assignment: Chapter 7.8

6. Discussion: Based on your understand of Gaussian elimination algorithm, how does the data

distribution affect the performance? What about using MPI for implementation?