Technical data

Optimizing with MACRO-32

– Combine vectorization with decomposition.

– Consider a solution at a higher level hierarchy.

– Retest the program compiling with /VECTOR (or any

combination with a parallel qualiﬁer) and return to the start

of step 3.

3.1.2 Combining Decomposition with Vectorization

To produce code that executes in parallel and on vector processors, compile

/VECTOR with a parallel qualiﬁer. Table 3–1 lists the compilation

combinations and their interpretations.

Table 3–1 Qualiﬁer Combinations for Parallel Vector Processing

Combination Interpretation

/VECTOR/PARALLEL=AUTOMATIC

Performs a dependence analysis on suitable loops and optimizes

them for parallel-vector processing; chooses loops and prepares

them for vector or parallel processing based on whether they

will execute efﬁciently and produce correct results. In a nested

structure, decomposition and vectorization may occur for multiple

loops — but no loop is decomposed inside a decomposed loop

and no loop is vectorized inside a vectorized loop.

/VECTOR/PARALLEL=MANUAL

Performs a dependence analysis, optimization, and vectorization

only; disqualiﬁes from vectorization any loops preceded by the

CPAR$ DO_PARALLEL directive from vectorization; in these

loops, parses the user-supplied directives.

/VECTOR/PARALLEL

(Same as VECTOR/PARALLEL=MANUAL)

/VECTOR/PARALLEL=(MANUAL,AUTOMATIC)

Same as /VECTOR/PARALLEL=AUTOMATIC except disqualiﬁes

loops preceded by CPAR$ DO_PARALLEL. In those loops,

only user-supplied directives are parsed. Any loops contained

in a manually decomposed loop are disqualiﬁed from

autodecomposition but not vectorization.

Both parallel and vector processing have certain tradeoff qualities, which

affect the aggregate speedup of vector and parallel processing. The

combined vector-parallel processing will be somewhat less than the

3–3