Prefetch Data from Memory

• Reordering instructions in unrolled loop leads to effective (pseudo-) prefetch of the data

• • no instruction overhead; compiler does this optimization automatically.

• Explicit (manual) prefetch for memory:

  • prefetch to 1st level cache should be done in form of pseudo-prefetch
  • compiler will insert prefetch to 2nd level cache automatically (LNO)
  • manual prefetch to 2nd level cache can be done with compiler directive:
  • same in C with the corresponding #pragma directive

for(i=0; i<n; i+=4){

t = b[i+3];

a += b[i+0];

a += b[i+1];

a += b[i+2];

a += t;

}

C*$* prefetch_ref=a(1)

c*$* prefetch_ref=a(1+16)

do I=1,n

c*$* prefetch_ref=a(I+32),stride=16,kind=rd

sum = sum + a(I)

enddo

Previous slide

Next slide

Back to first slide

View graphic version