The cosine terms are constants. These are multiplied with fyx for 2D transform. When this is implemented in hardware/software we deal with following issues:
- fixed point vs floating point representation of operands.
- quantization noise in operands, I am referring to cosine terms and intermediate results
- limited precision, especially for cosine terms
- rounding errors throughout calculatiom
I am curious to know, how does one out carry error analysis on a 2D DCT taking the above into consideration? Since 2D FDCT is separable, it can be implemented as 2 1-D FDCT in sequence.
All I want to learn is how to carry out error analysis for such algorithm for hardware and also software implementations!