Yep. There's a ton of data dependencies, where e.g. you can't begin decoding the next bit until you've finished decoding the current bit. It's all about progressive decoding so there's multiple passes over each group of wavelet coefficients, accessed in a data-dependent sequence. Each wavelet transform involves accessing and interleaving the same 2D arrays in both row- and column-oriented fashions.
These design decisions all made sense when clock rates were exponentiating, but they're all nightmares now that we rely on branch prediction and memory prefetching and superscalar execution units. The codec is simply not a good fit for the computing architectures we have today.
> These design decisions all made sense when clock rates were exponentiating, but they're all nightmares now that we rely on branch prediction and memory prefetching and superscalar execution units. The codec is simply not a good fit for the computing architectures we have today.
Arguably not a good choice for the year 2000, either, considering that all high performance CPUs at that time were out-of-order, superscalar and deeply pipelined.
These design decisions all made sense when clock rates were exponentiating, but they're all nightmares now that we rely on branch prediction and memory prefetching and superscalar execution units. The codec is simply not a good fit for the computing architectures we have today.