95% GPU usage but only x2 faster than the reference SIMD encoder/decoder