Integrate array_non_zero with the CAVLC 8x8dct interleave function.
Roughly 1.5-2x faster than the original separate array_non_zero method.