[AdaWorldAPI]

This is phase 1 of D-CMR-2b, a lab measurement only. ir.rs, exec.rs and the rest of the library are unchanged.

Background. fused_ternlog declines Pred, so a residual comparison runs on the tiled path through scratch. That scratch is already caller-owned and one tile wide (slots × 256 words), allocated once and never sized to the population. So the question was not heap memory. It was whether per-op tile write-back and the separate And and count passes cost anything measurable.

Probe: crates/lance-graph-benches/examples/pred_scratch_probe.rs. Run it with cargo run --release -p lance-graph-benches --example pred_scratch_probe. It lives in the benches crate, so mask-risc gains no ndarray dev-dependency.

• Shapes, over i32 lanes:
  • Count(a)
  • Count(a∧b)
  • b gated under a
  • Count(a∧b∧plane)
• Selectivity: about 1 %, 50 % and 99 %.
• Workloads:
  • whole: one extent 0..n, with n = 16k, 1M or 4M.
  • runs: set_lower's shape. Over 1M rows, K runs of 200, 3,000 or 40,000 rows at odd starts, together about a quarter of the rows.
• Arms:
  • A: execute_extent per extent, with reused scratch.
  • A+alloc: Scratch::for_program called once per call.
  • Ac: the program compiled once, then execute_compiled per extent.
  • B: compare into fresh chunks, then one mask_ternlog_popcount.
  • B′: the same fold over reused chunks.
• Before any timing, each configuration's answer is checked against a bit-serial oracle. The arms are interleaved within each rep in a rotating order, and the reported figure is the median of 101 per-rep ratios. A null arm (A timed twice) measures the method's own noise.

Result, from 3 runs. Each range covers the three selectivities across all three runs.

workload	null	A+alloc/A	Ac/A	B′/A
whole 16k	0.989–0.999	1.026–1.090	0.979–1.005	1.06–1.46
whole 1M	0.957–1.012	0.971–1.027	0.971–1.033	0.95–1.20
whole 4M	0.984–1.014	0.989–1.023	0.987–1.011	0.970–1.028
1,310 runs × 200	0.952–1.033	0.981–1.023	0.821–0.948	0.236–0.524
87 runs × 3,000	0.971–0.997	0.989–1.023	0.905–1.027	0.67–1.06
6 runs × 40,000	0.971–0.999	0.985–1.029	0.993–1.029	1.00–1.45

Reading.

• On whole extents and long runs the fold does not win; it is equal or slower. A Pred lowering arm for whole extents is not supported.
• On 200-row runs, per-run execute_extent costs about 170–520 ns a run, and the fold is 2.5–4× faster. Compiling once cuts A by only 5–18 %.
• One possible cause, from reading the code but not attributed by measurement: an unaligned extent runs as a head-edge word, a body and a tail-edge word, each running every op. Every call also validates.
• The remedy is not decided and not built. The options are fewer edge-tile passes, a run-batched entry point, or a short-run fold.
• Open: this version's fold is slower at full tiles than the first version of the probe (B′/A 1.06–1.46 against 0.95–1.05 at 16k). The cause isn't found. 64-byte chunk alignment was tested and is not it.

Board: entry entries/2026-10-21-d-cmr-2b-scratch-probe.md; D-CMR-2b rows in STATUS and in the plan; LATEST_STATE. Contract inventory is unchanged.

Coordination: claimed on EXO-50 as session_013z2e1YXkm6o6HNrDn5BPWA/023. Dependencies were confirmed there: ndarray at origin/master 37a3b59, and mask-risc's 47 tests pass against it.

🤖 Generated with Claude Code

https://claude.ai/code/session_013z2e1YXkm6o6HNrDn5BPWA

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

• Configuration used: Organization UI
• Review profile: CHILL
• Plan: Advanced
• Run ID: aa2e5fb2-6bd4-4a65-bd3b-b47d461a997e

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Warning

Billing warning: we have not been able to collect payment for this subscription for more than 72 hours. Please update the payment method or pay any pending invoices in Billing to avoid service interruption.

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-21T08:19:57.353071Z	b39cbac	Draft marked ready

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: b39cbac1f7

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".