perf(dfast): close level 3 parity gap — compress 11.0× → 5.5× of donor#145

Closed

polaz wants to merge 30 commits into

mainfrom

perf/#111-phase7c-dfast-level3-memory-tail

Commits on May 16, 2026

perf(encoding): #111 Phase 7c — split dfast hash table sizing to donor parity
polaz
committed
perf(encoding): #111 Phase 7c — port donor single-slot storage + sparse insertion
polaz
committed
perf(encoding): #111 Phase 7c step 3 — drop dfast window-Vec byte duplication
polaz
committed
perf(encoding): preallocate compress_to_vec output, add compress_slice_to_vec
polaz
committed
perf(encoding): cap compress_slice_to_vec initial capacity at one output block
polaz
committed
perf(dfast): tighten insert_positions inner loop with hoisted invariants
polaz
committed
perf(dfast): peek repcode at ip+1 to match donor exactly
polaz
committed
perf(dfast): lower DFAST_MIN_MATCH_LEN from 6 to 5 to match donor mls
polaz
committed
perf(dfast): switch hash to donor scalar mul, drop CRC32d kernel dispatch
polaz
committed
perf(dfast): inline(always) on best_match → hash_candidate → probe_slot_match chain
polaz
committed
perf(build): enable fat LTO and codegen-units=1 for release/bench
polaz
committed
perf(dfast): grow skip-step by scan distance to match donor
polaz
committed
perf(dfast): rewrite start_matching_fast_loop in donor outer/inner shape
polaz
committed
perf(dfast): inline rep1-only peek at ip+1 (donor parity, drop rep2/rep3)
polaz
committed
chore: ignore local helper docs in .local/
polaz
committed
perf(encoding): scale initial all_blocks capacity by source-size hint
polaz
committed
fix(dfast): tighten fast_loop guards to HASH_READ_SIZE to stop OOB load
polaz
committed
chore: rescope LTO to bench, fix stale docs, sync bench API call
polaz
committed
fix(dfast): rebase before fast-loop slot pack + compact on trim
polaz
committed
fix(dfast): carry tail seed cursor + simplify committed + scrub dead params
polaz
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

perf(dfast): close level 3 parity gap — compress 11.0× → 5.5× of donor#145

perf(dfast): close level 3 parity gap — compress 11.0× → 5.5× of donor#145
polaz wants to merge 30 commits into
mainfrom
perf/#111-phase7c-dfast-level3-memory-tail

Commits on May 16, 2026

perf(encoding): #111 Phase 7c — split dfast hash table sizing to donor parity

perf(encoding): #111 Phase 7c — port donor single-slot storage + sparse insertion

perf(encoding): #111 Phase 7c step 3 — drop dfast window-Vec byte duplication

perf(encoding): preallocate compress_to_vec output, add compress_slice_to_vec

perf(encoding): cap compress_slice_to_vec initial capacity at one output block

perf(dfast): tighten insert_positions inner loop with hoisted invariants

perf(dfast): peek repcode at ip+1 to match donor exactly

perf(dfast): lower DFAST_MIN_MATCH_LEN from 6 to 5 to match donor mls

perf(dfast): switch hash to donor scalar mul, drop CRC32d kernel dispatch

perf(dfast): inline(always) on best_match → hash_candidate → probe_slot_match chain

perf(build): enable fat LTO and codegen-units=1 for release/bench

perf(dfast): grow skip-step by scan distance to match donor

perf(dfast): rewrite start_matching_fast_loop in donor outer/inner shape

perf(dfast): inline rep1-only peek at ip+1 (donor parity, drop rep2/rep3)

chore: ignore local helper docs in .local/

perf(encoding): scale initial all_blocks capacity by source-size hint

fix(dfast): tighten fast_loop guards to HASH_READ_SIZE to stop OOB load

chore: rescope LTO to bench, fix stale docs, sync bench API call

fix(dfast): rebase before fast-loop slot pack + compact on trim

fix(dfast): carry tail seed cursor + simplify committed + scrub dead params

Commits on May 17, 2026

fix: empty-block guard, compress_bound saturating add, gitignore pattern

docs(dfast): correct collapse-pair, drop tracked interop crash, expect on collect

fix(dfast): free dead prefix on trim, guard ensure_room_for, hoist headroom check

fix(dfast): rep1 floor=4, probe gate width, empty-block guards

refactor(dfast): drop history_start alias + clarify fuzz test signal

refactor(dfast): inner-loop exit enum + API symmetry + doc cleanup

perf(fse): cache predefined LL/ML/OF default tables in AtomicPtr

fix(dfast): probe last hashable position + enforce short-hash floor

fix(dfast,fse): drop unused trim arg, harden FSE cache slot routing

perf(fse): return &'static FSETable from cache + dfast cleanup