Computing Integer Bit Widths for Video Codec Implementations

Video codecs are commonly implemented using efficient integer arithmetic. For correct operation, all integers must be sized to fit the full dynamic range of all signals. Despite most codecs being based on linear transforms, their implementations contain nonlinearities, such as quantization, making exactly computing their dynamic range intractable. While loose upper bounds are more easily calculated, these can significantly over estimate the number of bits required, leading to higher implementation costs. Instead implementers often resort to empirically determining necessary bit widths using test footage and noise signals. We found that this method can lead to under estimates of necessary bit widths of between one and two bits in practice. We propose a heuristic for creating synthetic test patterns which produce near worst case signals. These signals may be used to more robustly and efficiently determine bit width requirements for codec implementations.