Code Version 2 - Streamed

1 ANUM = 0.0

1 ADEN = 0.0

1v ------- DO 131 j=1,Njj

123 ------ DO 131 i=1,Nii

123v ----- DO 131 k=1,Nkk

123v ANUM=ANUM+conjG(GDn(i,j,k))*HDn(i,j,k)

123v & +conjG(GVz(i,j,k)) *HVz(i,j,k)

123v & +conjG(GVr(i,j,k)) *HVr(i,j,k)

123v & +conjG(GVa(i,j,k)) *HVa(i,j,k)

123v ADEN=ADEN+conjG(XDn(i,j,k))*XDn(i,j,k)

123v & +conjG(XVz(i,j,k)) *XVz(i,j,k)

123v & +conjG(XVr(i,j,k)) *XVr(i,j,k)

123v & +conjG(XVa(i,j,k)) *XVa(i,j,k)

123v ====> 131 CONTINUE

Previous slide

Next slide

Back to first slide

View graphic version

• -Here is a code segment that will show what is ment by serial dependencies.

• -Two global sums are computed and we see that after some optimization the inner loop vectorizes well

• -The loop will not task or stream, however, due to the global reduction operation

• -To optimize for an MSP, this serial depedency can be removed