Safe Context-adaptive binary arithmetic coding (CABAC) encoder / decoder in Rust

Implementation of CABAC using H.264/265, VP8, rANS and Fpaq0 encoders.

The coder is designed to encode binary values in an efficient manner, taking into account the bits that were previously seen. The previous state is stored in the Context object, which is updated each time a bit is seen. Normally multi-bit symbols are converted to binary form, and each bit is assigned a corresponding bin (which gets its own context). Bits that are expected to be random can use the "bypass" mode, which are very efficiently added to the bitstream without a context.

In order to decode the bistream, the same contexts need to be used in the exact same sequence or you will get back garbage. This also include any bypass bits that must also be read in the same exact order.

There are four encoders included:

• h264/h265 CABAC which uses a 6 bit state to track previously
• VP8 CABAC which uses a 16-bit state to track what it has seen.
• rANS encoder (based on ryg_rans and dropbox/lepton) that uses the VP8 state to track probability
• Fpaq0 arithmetic encoder which has some nice properties since it is fast, carryless and can be run in parallel similar to the rANS. The parallel mode allows for interleving of arbitary bitstreams as long as the bitstreams are written in the same order as the bits are encoded.

Performance notes:

• Criterion bench tests included
• No unsafe code
• rANS has not yet been significantly optimized although it outperforms the other encoders, encoding uses division although it could use an inverse multiple and decoding is done one value at a time, although it could be done in parallel.
• Fpaq0 has a parallel SIMD version using the wide crate, but it is not yet faster than the non-SIMD version. It is feature config off by default. The Fpaq0 parallel version requires the parallel streams to be somewhat balanced, otherwise encoding performance may suffer. Decoding performance is not affected.

Here is the relative performance (in microseconds, lower is better) for encoding, decoding as measured on Intel i9-12900K (compiled with -Ctarget-cpu=native):