cvmg

Date:

10-12-2011

NAME

CVMGM, CVMGN, CVMGP, CVMGT, CVMGZ - (Deferred implementation) Conditional vector merge functions

SYNOPSIS

CVMGM ([I=]i, [J=]j, [K=]k)
CVMGN ([I=]i, [J=]j, [K=]k)
CVMGP ([I=]i, [J=]j, [K=]k)
CVMGT ([I=]i, [J=]j, [K=]k)
CVMGZ ([I=]i, [J=]j, [K=]k)

IMPLEMENTATION

Cray Linux Environment (CLE)

STANDARDS

Fortran extension

DESCRIPTION

At run time, k is tested. You can use the conditional vector merge (CVMG) functions when an IF statement involving arrays prevents vectorization of a loop. The compiler can vectorize almost all such loops, but these functions can be used in older codes. Scalar arguments can also be used with these functions.

CVMG functions cannot be passed as arguments. They are elemental functions.

These functions test for the following:

  • CVMGM tests for minus (negative). i is returned if k < 0. j is returned if k >= 0.

  • CVMGN tests for nonzero. i is returned if k is not equal to 0. j is returned if k = 0.

  • CVMGP tests for positive or zero. i is returned if k >= 0. j is returned if k < 0.

  • CVMGT tests for true. i is returned if k is true. j is returned if k is false.

  • CVMGZ tests for zero. i is returned if k = 0. j is returned if k is not equal to 0.

These functions accept the following arguments:

i

Can be of type logical, Boolean, integer, real, or Cray pointer.

j

Can be of type logical, Boolean, integer, real, or Cray pointer.

k

Can be of type logical, Boolean, integer, real, or Cray pointer.

If k satisfies the condition tested by the function (for example, in CVMGP, if k is positive), the first argument (i) is returned as the function result. If k does not satisfy the condition tested by the function, j is returned as the function result.

For operands other than type logical, the vector merge functions are not generic with respect to data typing. They accept arithmetic values of different types but interpret them as Boolean type (that is, as bit patterns). The returned function value is also Boolean. Therefore, the function reference CVMGT(1.0,2.0,.TRUE.) is not type real, with a value of 1.0, but a Boolean value that acts the same as 1.0 in a floating-point operation.

A problem can arise if you assume that the function reference is type real and use it in an expression or assignment that causes automatic type conversion. For example, if you use the function reference in a context where an integer is needed, the result is not valid because 1.0 and 1 have different bit patterns.

Because CVMG function values are Boolean, a binary operation involving two CVMG functions uses integer arithmetic. This can produce unexpected results in assignments such as the following in which A, B, C, and D are real:

X = CVMGT(A,B,LOGIC1) + CVMGT(C,D,LOCIC2)
                     ! Integer arithmetic, invalid results

However, when used in an expression with another operand, a CVMG function value takes on the type of the other operand, without any explicit type conversion. For example, the following expression uses real arithmetic:

X = 1.0 + CVMGT(2.0,3.0,LEXP)  ! Valid (types agree)

The following suggestions explain how to prevent bugs when using these functions:

  • Use only one CVMG function in a given expression. If you use more than one CVMG function in an expression, use explicit type conversion. This does not generate any additional code. For example:

X = REAL(CVMGT(1.0,2.0,LTEST)) + REAL(CVMGT(X,Y,LTEST2))

  • Ensure that the assignment type matches the function argument. For example:

X = CVMGT(1.0,2.0,LTEST) ! Valid
X = REAL(CVMGT(1,2,LTEST)) ! Valid (uses explicit type conversion)
X = CVMGT(1,2,LTEST) ! Invalid (type mismatch)

  • Never use mixed typing in the first two arguments of a CVMGT function. Example:

CVMGT(1,2.0,LTEST)  ! DO NOT DO THIS

The names of these intrinsics cannot be passed as arguments.

EXAMPLES

Consider the following code:

DO I = N,M
   X(I) = A(I)
   IF (B(I) .GT. C(I)) X(I) = D(I)
END DO

This could be rewritten as follows:

DO  I = N,M
   X(I) = CVMGT(D(I), A(I), B(I).GT.C(I))
END DO

The following rewrite would also be valid:

DO I = N,M
   X(I) = CVMGP( D(I), A(I), B(I) - C(I))
END DO