* For input float32 normalized range (1.0,-1.0) -> output int[16,24,32]:
* scale it by multiplying filter coefficients by 2^31
* For input int[16, 24, 32] -> output float32
* For input float32 normalized range (1.0,-1.0) -> output int[16,24,32]:
* scale it by multiplying filter coefficients by 2^31
* For input int[16, 24, 32] -> output float32