Skip to content

docs: refresh BENCHMARK.md division/parse/tostr tables post PRs #82–#89#91

Merged
mmurshed merged 1 commit into
mainfrom
docs/benchmark-refresh
Jun 12, 2026
Merged

docs: refresh BENCHMARK.md division/parse/tostr tables post PRs #82–#89#91
mmurshed merged 1 commit into
mainfrom
docs/benchmark-refresh

Conversation

@mmurshed

Copy link
Copy Markdown
Owner

Re-measured the division (skewed + balanced), parse, and ToString tables on the 2026-05-30 size grids with a focused paired harness (GMP re-timed in the same run). Multiplication tables left at 2026-05-30 values — mul code untouched by the session. 100M/200M division rows annotated as not re-measured.

table headline
div skewed 5M×1M: 200.6 → 111.4 ms (2.87× → 1.59×); 50M×10M: 1.13× — near parity
tostr 1M: 224.8 → 170.2 ms (4.48× → 3.31×); 20M: 2.52× → 1.78×
parse 1M: 49.0 → 43.3 ms; 20M: 1.56× → 1.50×

Also adds a "2026-06-11 division & decimal-IO session" summary section — one row per PR (#82#89) with headline wins, the post-session dispatch state (no known shape holes), and the two process lessons (CPU-time ≠ wall-clock on the threaded stack; Newton tower drift ~B^6 ulps needs B^16 residue windows).

🤖 Generated with Claude Code

Re-measured 2026-06-11 with a focused paired BigMath-vs-GMP harness on
the same size grids (GMP re-timed in the same run; mul tables left as
2026-05-30 values since multiplication code was untouched; the 100M/200M
division rows are annotated as not re-measured).

Headline movements:
- skewed div 5M x 1M: 200.6 -> 111.4 ms (2.87x -> 1.59x vs GMP)
- skewed div 50M x 10M: 2464 -> 1461 ms (1.86x -> 1.13x, near parity)
- tostr 1M digits: 224.8 -> 170.2 ms (4.48x -> 3.31x)
- tostr 20M: 5433 -> 3785 ms (2.52x -> 1.78x)
- parse 1M: 49.0 -> 43.3 ms (2.33x -> 2.11x)

Adds a session summary section (PRs #82-#89, one row per PR with
headline wins, dispatch-state summary, and the two process lessons),
linking to DIVISION.md / STRING_CONVERSION.md for the deep dives.
ToString methodology matches the original table: warm best-of-N below
100k digits, single cold run (chain build included) at 100k+.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
@mmurshed mmurshed merged commit bafdc18 into main Jun 12, 2026
@mmurshed mmurshed deleted the docs/benchmark-refresh branch June 12, 2026 05:04
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant