Code Version 3 - Streamed

M-- 12 ------- DO 131 j=1,Njj

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

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

M 123v tmp1(i,j,k)=conjG(GDn(i,j,k))*HDn(i,j,k)

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

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

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

M 123v tmp2(i,j,k)=conjG(XDn(i,j,k))*XDn(i,j,k)

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

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

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

M-- 123v ====> 131 CONTINUE

1 ANUM=SUM(tmp1)

1 ADEN=SUM(tmp2)

Previous slide

Next slide

Back to first slide

View graphic version

• -Here is the modified code after the 3rd round of optimization

• -By adding in dummy arrays for the partial sums and moving the global reduction outside of the loop, the patrial sum can now be streamed