Fixed bug with finding the exit iteration of the inner loops, -1 and +1 to value. Replaced vec_load with vec_unaligned_load for GCC. Pad cache correctly when the number of iterations of the inner loop is not a multiple of the vector length.