D-RMAP-2 probe: BF16 16×16 tile GEMM vs bit tile, first measured cell by AdaWorldAPI · Pull Request #1623 · AdaWorldAPI/lance-graph

AdaWorldAPI · 2026-10-20T01:09:31Z

This PR was split out of #1622, which now carries only the D-RST plan. It is probe code only, in crates/mork-pathmap-probe; nothing is adopted.

What it measures

On one 256×256 focus map it computes Boolean composition, walk counts and directed triangle counts two ways:

through ndarray::simd::bf16_tile_gemm_16x16_packed, via the facade (no new kernel);
through a bit matrix.

Both are checked against a triple-loop oracle. Fixtures: density 1%, 10% and 50%, plus a skewed one with 4 dense hub rows and columns.

src/tile_gemm.rs, tests/tile_gemm.rs: 4 tests. Five guards were each disabled once and each made a test fail (transpose in compose, threshold, triangle transpose, B-block column offset, count transpose).
examples/tile_gemm_cost.rs: release build, medians of 7, every timed answer asserted first, inputs passed through black_box.
- black_box mattered. An earlier run without it reported a 0.2 µs triangle count, which was the compiler hoisting the pure call out of the timing loop.

Results, with two conclusions kept apart

Measured on this host's F32x16 polyfill tier.
- BF16 walk counts equal the oracle exactly, and count > 0 equals Boolean composition.
- The bit arm wins every cell:
  - composition: 2.6–25 µs;
  - walk counts by popcount: 165–192 µs;
  - triangles: 1.9–72 µs.
- BF16 A·A takes about 19 ms, and staging bits to BF16 alone costs 150–400 µs.
Not measured: anything about AMX or VDPBF16PS hardware.
- This host lacks both. Checked directly per ndarray's AMX Gotcha 4: CPUID AMX and AVX512-BF16 bits are 0, XCR0 = 0xe7, and arch_prctl(ARCH_REQ_XCOMP_PERM) on syscall 158 returns -95.
- The polyfill allocates on every output element (ISS-BF16-POLYFILL-PER-ELEMENT-ALLOC). Its cost is therefore not even a lower bound on tile-hardware cost.

Board

D-RMAP-2 §5a.
STATUS_BOARD D-RMAP-2 row: PARTIAL.
ISS-BF16-POLYFILL-PER-ELEMENT-ALLOC in ISSUES.md.
Board entry 2026-10-20-d-rmap-2-tile-gemm-cell.md.
Entries index regenerated.

The three commits are cherry-picks of #1622's original commits, so their provenance survives. The cherry-pick dropped one two-line hunk that touched the D-RST plan, which this branch does not carry.

Checks run locally

Probe cargo test --release, cargo clippy --all-targets -D warnings and cargo fmt --check are clean.
Append-only gate passes; no new citation-decay findings.

Generated by Claude Code

Boolean composition, walk counts and triangle counts on one 256 x 256 focus, through ndarray::simd::bf16_tile_gemm_16x16_packed (the facade, not a new kernel) and through a bit matrix, both against a triple-loop oracle. 4 tests; 5 guards disable-verified. D-RMAP-2 §3's exactness claim holds: BF16 walk counts equal the oracle exactly and count > 0 equals Boolean composition at 1 %, 10 %, 50 % and with skewed hubs. On this host only the F32x16 polyfill tier runs (no AMX, no avx512_bf16): A*A takes about 19 ms, while the bit arm composes in 2.6-25 us and gives the same walk counts by popcount in 165-192 us. The polyfill allocates per output element, recorded as ISS-BF16-POLYFILL-PER-ELEMENT-ALLOC; tile-tier hosts are unmeasured. Recorded in D-RMAP-2 §5a, the D-RST plan §5a, STATUS_BOARD, ISSUES and a board entry. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VyWZYKycy6MoYDxHRVnV5K

CPUID, XCR0 and arch_prctl(ARCH_REQ_XCOMP_PERM) on syscall 158 checked directly, per ndarray AMX_GOTCHAS Gotcha 4: no AMX, no AVX512-BF16; the permission request returns EOPNOTSUPP. The polyfill tier is the host's, not a detection bug. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VyWZYKycy6MoYDxHRVnV5K

… apart The measured polyfill-tier result is not evidence about AMX or VDPBF16PS performance, not even a lower bound, since the polyfill's cost is dominated by allocation. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VyWZYKycy6MoYDxHRVnV5K

coderabbitai · 2026-10-20T01:09:39Z

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: b64674a8-61f2-4a27-ab89-43ae71873d1d

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

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

Reverts b5672a5 and 3329eeb on this branch. Their content now lives in #1623 as cherry-picks, so its measurements, disable-verified guards and provenance are kept there. This PR returns to D-RST-0, the plan, and gains one pointer to #1623 in §5a. No protected history is rewritten: the append-only gate compares against the merge-base with main, and the reverted ISSUES.md and board-entry additions were never on main. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VyWZYKycy6MoYDxHRVnV5K

States the operator clarification: preparation, execution and maintenance are measured apart; break-even comes from measured values; §5a's cell is read in those terms (execution on prepared operands only, with BF16 staging as part of preparation); Boolean existence vs BF16 multiplicity and the BF16 exactness bound. No kernel change. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VyWZYKycy6MoYDxHRVnV5K

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VyWZYKycy6MoYDxHRVnV5K

chatgpt-codex-connector · 2026-10-20T08:20:29Z

Codex Review Summary

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

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-20T08:23:17.874818Z	`23f0998`	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: 23f0998902

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".

Conflict only in the generated .claude/board/entries/README.md; regenerated with entries_index.py --write (378 rows, --check passes). STATUS_BOARD auto-merged to this branch's measured D-RMAP-2 row. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VyWZYKycy6MoYDxHRVnV5K

…fter #1623 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VncKoyG6JkpsoqhqTGTv3f

claude added 3 commits October 20, 2026 01:06

AdaWorldAPI mentioned this pull request Oct 20, 2026

plan (lab only): relational geometry amortization (D-RST) and recursive perspective rebinding (D-RPR) #1622

Merged

claude added 2 commits October 20, 2026 01:32

D-RMAP-2: mark lab exploration only, not admitted to production

23f0998

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VyWZYKycy6MoYDxHRVnV5K

AdaWorldAPI marked this pull request as ready for review October 20, 2026 08:20

chatgpt-codex-connector Bot reviewed Oct 20, 2026

View reviewed changes

Comment thread crates/mork-pathmap-probe/src/tile_gemm.rs

AdaWorldAPI merged commit 8b570ed into main Oct 20, 2026
13 checks passed

AdaWorldAPI mentioned this pull request Oct 20, 2026

D-RMAP-2: count the row-major B blocks in the BF16 byte total #1625

Merged

AdaWorldAPI mentioned this pull request Oct 20, 2026

Consolidate into #1622: D-RPR-1 sub-cases, plan links for #1601–#1617, and the handoff lab probe #1624

Merged

Navigation Menu

Repository navigation

D-RMAP-2 probe: BF16 16×16 tile GEMM vs bit tile, first measured cell - #1623

D-RMAP-2 probe: BF16 16×16 tile GEMM vs bit tile, first measured cell#1623
AdaWorldAPI merged 6 commits into
mainfrom
ccr-08f48cec-lbv9l4-tile-gemm

AdaWorldAPI commented Oct 20, 2026

Uh oh!

coderabbitai Bot commented Oct 20, 2026 •

edited

Loading

Uh oh!

chatgpt-codex-connector Bot commented Oct 20, 2026 •

edited

Loading

Uh oh!

chatgpt-codex-connector Bot left a comment

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

AdaWorldAPI commented Oct 20, 2026

What it measures

Results, with two conclusions kept apart

Board

Checks run locally

Uh oh!

coderabbitai Bot commented Oct 20, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

chatgpt-codex-connector Bot commented Oct 20, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codex Review Summary

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

coderabbitai Bot commented Oct 20, 2026 •

edited

Loading

chatgpt-codex-connector Bot commented Oct 20, 2026 •

edited

Loading