perf: 5.4x faster recompression (11.6x single-worker), byte-identical output - #1

Merged

Snesnopic merged 19 commits into

Snesnopic:mainfrom

W-Floyd:perf/huffman-region-search

Aug 1, 2026

Commits on Jul 30, 2026

perf(huffman): search only the region costs big_values can move

W-Floyd
and
Claude Opus 5 (1M context)
committed
8a82d7d
perf(bitstream): one 64-bit access per field instead of a bit at a time

W-Floyd
and
Claude Opus 5 (1M context)
committed
9692a24
perf(packer): stop copying and re-reading what is already in memory

W-Floyd
and
Claude Opus 5 (1M context)
committed
7f9989e
perf(packer): assemble the output in memory and write it once

W-Floyd
and
Claude Opus 5 (1M context)
committed
fe3bdb4
perf(huffman): let the span search vectorise

W-Floyd
and
Claude Opus 5 (1M context)
committed
bb9ca1d
perf(huffman): hand-write the two cost kernels

W-Floyd
and
Claude Opus 5 (1M context)
committed
03b079f
perf(huffman): enumerate candidates by shape, not by region counts

W-Floyd
and
Claude Opus 5 (1M context)
committed
65b82c1
perf(huffman): decode a count1 quadruple from one lookup

W-Floyd
and
Claude Opus 5 (1M context)
committed
0d35edf
perf(packer): copy scalefactors rather than decode and re-emit them

W-Floyd
and
Claude Opus 5 (1M context)
committed
25815d6
perf(huffman): skip the region a zero table codes, and bound it once

W-Floyd
and
Claude Opus 5 (1M context)
committed
621737e
refactor(huffman): drop the unused pair decode table

W-Floyd
and
Claude Opus 5 (1M context)
committed
06047e1
perf(packer): do not spawn workers when there is no search to share

W-Floyd
and
Claude Opus 5 (1M context)
committed
0b25783
fix: build the vector kernels only where their instructions exist

W-Floyd
and
Claude Opus 5 (1M context)
committed
02484fc
fix(huffman): do not treat an uncodable tail as a free one

W-Floyd
and
Claude Opus 5 (1M context)
committed
d7363eb
perf(huffman): resolve a big-value pair from one ten-bit probe

W-Floyd
and
Claude Opus 5 (1M context)
committed
5b83a07
perf(packer): describe the new reservoir instead of building it

W-Floyd
and
Claude Opus 5 (1M context)
committed
030417e
perf(packer): let frames describe their bytes instead of copying them

W-Floyd
and
Claude Opus 5 (1M context)
committed
de55c23
perf(packer): write new side info into the frame, not into an array beside it

W-Floyd
and
Claude Opus 5 (1M context)
committed
e22ae1b

Commits on Jul 31, 2026

fix: add missing #include <string> to mp3_reader.hpp
W-Floyd
committed
55634aa

Navigation Menu

perf: 5.4x faster recompression (11.6x single-worker), byte-identical output - #1

perf: 5.4x faster recompression (11.6x single-worker), byte-identical output#1
Snesnopic merged 19 commits into
Snesnopic:mainfrom
W-Floyd:perf/huffman-region-search

Commits on Jul 30, 2026

perf(huffman): search only the region costs big_values can move

perf(bitstream): one 64-bit access per field instead of a bit at a time

perf(packer): stop copying and re-reading what is already in memory

perf(packer): assemble the output in memory and write it once

perf(huffman): let the span search vectorise

perf(huffman): hand-write the two cost kernels

perf(huffman): enumerate candidates by shape, not by region counts

perf(huffman): decode a count1 quadruple from one lookup

perf(packer): copy scalefactors rather than decode and re-emit them

perf(huffman): skip the region a zero table codes, and bound it once

refactor(huffman): drop the unused pair decode table

perf(packer): do not spawn workers when there is no search to share

fix: build the vector kernels only where their instructions exist

fix(huffman): do not treat an uncodable tail as a free one

perf(huffman): resolve a big-value pair from one ten-bit probe

perf(packer): describe the new reservoir instead of building it

perf(packer): let frames describe their bytes instead of copying them

perf(packer): write new side info into the frame, not into an array beside it

Commits on Jul 31, 2026

fix: add missing #include <string> to mp3_reader.hpp