Outer Loop Unrolling

• • the unroll factor should match the cache line size
  • mostly 1st level cache optimization
  • if the data fits into the 2nd level cache, this is good optimization to use

DO I=1,N

DO J=1,N

A(I)=A(I)+B(I,J)*C(J)

ENDDO ENDDO

Problem:

A(I) is constant for the inner loop J

C(J) is traversed each I iteration

B(I,J) is traversed poorly

Unrolling the outer loop

will load the complete cache

line of B in to the registers

-> data re-use

one 1st level cache line

-LNO:outer_unroll=n

Previous slide

Next slide

Back to first slide

View graphic version