A 64-bit word contains four 16-bit signed fixed-point values. This is the fixed 16-bit data format.
A 64-bit word contains two 8-bit signed fixed-point values. This is the fixed 32-bit data format.
Enough precision and dynamic range (for filtering and simple image computations on pixel values) can be provided by an intermediate format of fixed data values. Pixel multiplication is used to convert from pixel data to fixed data. Pack instructions are used to convert from fixed data to pixel data (clip and truncate to an 8-bit unsigned value). The FPACKFIX instruction supports conversion from 32-bit fixed to 16-bit fixed. Rounding is done by adding one to the rounding bit position. You should use floating-point data to perform complex calculations needing more precision or dynamic range.