From http://ols.108.redhat.com/2007/GCC-Reprints/GCC2007-Proceedings.pdf . and more speci?cally across iterations of a loop, classic vectorization techniques have traditionally focused on exploiting exactly that kind of parallelism ([4], [24]). This is also the kind of parallelism that G.C.C vectorizer has been originally designed to target ([13], [15]). However, there are many kinds of important computations that cannot be vectorized if considering only the context of the loop. Figure 1a shows, for example, an unrolled loop, hand-optimized for a sequential machine, which is typical for multimedia kernels [9]. Other possible examples include structure accesses, such as R.G.B.A in image
Duration : 0:1:53
From http://ols.108.redhat.com/2007/GCC-Reprints/GCC2007-Proceedings.pdf . just needed to be 