In an earlier study, we inserted prefetches by hand into MP3D, LU, and
PTHOR [61]. In the case of PTHOR, the access patterns are
irregular and difficult to prefetch even by hand-we will discuss that
case later in Section 
. In contrast, the other two
cases (MP3D and LU) have regular and predictable access patterns. Therefore
inserting prefetches by hand was relatively easy, and we saw large
performance gains. Our compiler algorithm is also successful at prefetching
these two cases, which we examine in more detail in this subsection.