The max and average kernel implementations for Xtensa fall back to the reference kernel when using 16-bit activations and 8-bit weights, instead of leveraging the available optimized Xtensa implementation.