STREAM_UNROLL

Purpose

The STREAM_UNROLL directive instructs the compiler to apply the combined functionality of software prefetch and loop unrolling to DO loops with a large iteration count. Stream unrolling functionality optimizes DO loops to use multiple streams. You can specify the STREAM_UNROLL directive for both inner and outer DO loops, and the compiler will use an optimal number of streams to perform stream unrolling where applicable. Applying the STREAM_UNROLL directive to a loop with dependencies will produce unexpected results.

Syntax


>>---STREAM_UNROLL--+---------------------+--------------------><
                    '-(--unroll_factor--)-'

unroll_factor: The unroll_factor must be a positive scalar integer constant expression. An unroll_factor of 1 disables loop unrolling. If you do not specify an unroll_factor, the compiler determines the optimal number to perform stream unrolling.

Rules

You must specify one of the following compiler options to enable loop unrolling:

–O3 or higher optimization level
-qhot compiler option
-qsmp compiler option

Note that if the -qstrict option is in effect, no stream unrolling will occur. If you want to enable stream unrolling with the -qhot option alone, you must also specify -qstrict=none.

The STREAM_UNROLL directive must immediately precede a DO loop.

You must not specify the STREAM_UNROLL directive more than once, or combine the directive with UNROLL, NOUNROLL, UNROLL_AND_FUSE, or NOUNROLL_AND_FUSE directives for the same DO construct.

You must not specify the STREAM_UNROLL directive for a DO WHILE loop or an infinite DO loop.

Examples

The following is an example of how STREAM_UNROLL can increase performance.

     integer, dimension(1000) :: a, b, c
     integer i, m, n

!IBM* stream_unroll(4)
      do i =1, n
        a(i) = b(i) + c(i)
      enddo
     end

An unroll factor reduces the number of iterations from n to n/4, as follows:

m = n/4
do  i =1, n/4
    a(i) = b(i) + c(i)
    a(i+m) = b(i+m) + c(i+m)
    a(i+2*m) = b(i+2*m) + c(i+2*m)
    a(i+3*m) = b(i+3*m) + c(i+3*m)
enddo

The increased number of read and store operations are distributed among a number of streams determined by the compiler, reducing computation time and boosting performance.

Related information

For further information on using prefetch techniques in XL Fortran see the PREFETCH directive set.
For additional methods on optimizing loops, see the BLOCK_LOOP, LOOPID, UNROLL and the UNROLL_AND FUSE directives.