Module ac 

Arithmetic Coder implementation for rwkvzip.

This implements a binary arithmetic coder with 32-bit precision, optimized for neural network probability distributions. The implementation is mathematically rigorous to ensure lossless compression.

Information-Theoretic Properties

• Uses base-2 arithmetic for bitstream output
• 32-bit precision prevents underflow for typical neural network distributions
• Integer CDF quantization uses 30-bit total (2^30) to minimize quantization error
• Probability floor ensures no symbol has zero probability (critical for lossless)

Structs

ArithmeticDecoder

Binary arithmetic decoder.

ArithmeticEncoder

Binary arithmetic encoder.

Constants

CDF_TOTAL

Total count for CDF quantization (2^30 for high precision)

Functions

p_min

Returns the minimum probability floor for symbols. P_MIN = 2 * 2^(-(PRECISION-2)) = 2^(-(PRECISION-3)) = 2^(-29)

quantize_pdf_to_cdf

Quantize probability distribution to integer CDF.

quantize_pdf_to_cdf_inplace

Quantize PDF to integer CDF using a reusable output buffer.

softmax_pdf

Compute softmax PDF without floor (for entropy calculation).

softmax_pdf_floor

Compute softmax PDF with probability floor.

softmax_pdf_floor_inplace

In-place version of softmax_pdf_floor to avoid allocations.

softmax_pdf_inplace