Technical data

Optimizing with MACRO-32

3.1 VECTORIZATION

Many loops that VAX FORTRAN decomposes can also be vectorized. VAX

FORTRAN performs vectorization automatically whenever /VECTOR is

speciﬁed at compilation. VAX FORTRAN can vectorize any FORTRAN–77

Standard-conforming program; and vectorized programs can freely call

and be called by other programs (vectorized, not vectorized, and non-

FORTRAN) as long as both abide by the VAX Calling Standard.

The VAX vector architecture supports most FORTRAN language features,

as follows:

• Data types: LOGICAL*4, INTEGER*4, REAL*4, REAL*8,

COMPLEX*8, and COMPLEX*16

• Operators: +, -, *, /(ﬂoating point), and **

• All VAX FORTRAN intrinsic functions

Although no VAX vector form exists for integer divide operations,

VAX FORTRAN vectorizes them by converting them to ﬂoating-point

operations.

3.1.1 Using Vectorization Alone

Vectorize a program using the following iterative process:

1 Using /CHECK=BOUNDS, compile, debug, and run a scalar version of

the program.

2 Compile and run the program using /VECTOR and the suitable

vector-related qualiﬁers.

3 Evaluate execution time and results. The results should be

algebraically equivalent to the scalar results; if not, check the

DUMMY_ALIASES or ACCURACY_SENSITIVE settings.

• If performance is adequate, stop.

• If performance is inadequate, you have similar options as with

autodecomposition:

– Check the /SHOW=LOOPS output to see if CPU-intensive

loops vectorized. To vectorize effectively, source code must

not contain certain inhibiting constructs such as unknown

dependencies. However, you can use LSE diagnostics and add

assertions to the source code to overcome dependencies.

3–2