> >We prefetch into L1 or L2, depending on what's going on. In this case > >L1 is per core and L2 is shared, so you're suggesting we give up > >prefetching/blocking in L2? > How do you do that? By compiler? Manually? Our compiler does this automatically. -- greg