Macroblock tree overhaul/optimization
Move the second core part of macroblock tree into an assembly function; SIMD-optimize roughly half of it (for x86). Roughly ~25-65% faster mbtree, depending on content. Slightly change how mbtree handles the tradeoff between range and precision for propagation. Overall a slight (but mostly negligible) effect on SSIM and ~2% faster.
Showing with 391 additions and 92 deletions