The hard part aside... Some cool optimizations are possible if the size of each row doesn't change. * Rows could be done in parallel. * The simplest method wou

The hard part aside... Some cool optimizations are possible if the size of each row doesn't change. * Rows could be done in parallel. * The simplest method would be to copy the row to a scratch buffer, and then copy over data to their final position. * Furthermore, with some fancy permutation, you only need a register of scratch space. Although, with more random memory access, that could result in worse performance.