dr-isa-stream-lab: G2L falsifier — one witnessed translation reused across GEMMs (D-DT64-G2L-1) by AdaWorldAPI · Pull Request #1994 · AdaWorldAPI/lance-graph

AdaWorldAPI · 2026-11-10T18:18:45Z

#1992 (its former base) and AdaWorldAPI/ndarray#418 are both merged; this PR now targets main directly. Persistence of the map table is #1993.

What

The consumer names a selection in global ids; the source holds its u8 lanes in its own row order (SPACE_B), described by a DR MapTable. Reading across the two spaces directly is a gather and stays refused. Per ADMISSION B (#1640):

translate locates every selected id once, under one map-table frame (counted lookups), and expresses the selection in the source's order. A map whose coordinate revision is not the source's is refused (stale mapping).
The object is bound on the source, in the source's space.
B is prepared once through TranslatedB. Admission resolves every selected position to the identity it reads (BValues::highest_row) and checks it against the inner operand's extent; packing then makes one map read per row. The operand identity names the translation.
Every run reads the source's lanes with zero G2L lookups and zero map reads.

Measured (one AMX VM, three runs; each quantity on its own)

64 fields × 64 columns, 200 global ids, int8_tile_gemm_tier() = AMX TDPBUSD.

Step	Time	Movement
translate once	83–85 µs	3200 G2L lookups
prepare B once (admit + pack)	46–48 µs	200 admission + 200 packing map reads; 16,384 B bytes packed
stage A, per warm run	11–14 µs	12,800 useful bytes into a 16,384-byte padded buffer
GEMM only	0.71–0.86 µs	—
warm run (resolve + stage A + GEMM)	20–24 µs	0 lookups, 0 map reads, 12,800 A bytes staged
re-translate per operation	157–186 µs	3200 lookups + 400 map reads + B pack + A stage

Zero lookups is not zero data movement. The selection is scattered in the source's order, so every warm run still gathers it into a dense K-padded A, at about 15× the cost of the GEMM. The translation removes the global → local resolution, not the gather.

K order. Translation reorders K. That is validated here only for exact integer GEMM (the sum does not depend on order, absent overflow). It is not a result for order-sensitive operators: floating-point reductions, saturation, scans, or anything that reads K by position.

Operand extent (review fix)

TranslatedB::extent() used to report the source's position range. An identity past a short inner BTable was admitted, and packing indexed past the table (panic). It now reports the inner extent, and BValues::highest_row maps positions to identities, so admit_operand refuses with ShapeRefusal::OperandShape { last: 3000, rows: 100, .. } before packing reads anything (#1990's contract). Test: a_translated_identity_past_the_inner_table_is_refused_before_packing, with a silence twin where every identity fits.

A translation is also refused when it was made on another source in the same order (G2lRefusal::Foreign): a_translation_from_another_source_in_the_same_order_is_refused.

Falsifiers (`tests/integration/g2l.rs`), each disable-verified red

Test	Disable that turns it red
exact vs an oracle in global ids alone; 5 runs, 0 further lookups or map reads	per-row packing cache removed
a source reorder refuses the old binding, translation and operand	staleness check in `TranslatedB::new` removed
a stale mapping (lanes reordered, map not) is refused	map-vs-source coordinate check removed
a prepared B from one translation is refused under another	translated identity ignoring the translation
an identity past the inner B table is refused before packing	`TranslatedB::highest_row` removed
a translation from another source in the same order is refused	the frame check in `Translation::check` removed
the untranslated cross-space read stays refused; unknown and repeated ids refused	—

55 lab tests pass in release mode; clippy -D warnings clean.

Still open: who owns the translation in production, and frames across a restore (#1993).

🤖 Generated with Claude Code

https://claude.ai/code/session_01XAqKMMWgnUU9g7Ldyn6sbk

Adds int8::direct and int8::PreparedI8B. A is read in place at the u8 lanes' pitch and B is prepared once as PackedI8B bands, through ndarray::simd::int8_gemm_packed_strided (AdaWorldAPI/ndarray#418). The prepared-B reuse rule moves into shape::ReuseKey, shared by the BF16 and INT8 prepared operands so the two cannot drift. Measured on one AMX VM (median of 41 warm runs, three runs), 256 fields x 256 columns: the direct GEMM takes 45.5-48.1 us against 78.3-78.8 us for backend::gemm_i8, end to end 47.5-57.9 us against 88.9-92.2 us; at 64 fields 0.98-1.01 us against 3.4-3.6 us. No A copied, no allocation. Needs ndarray#418. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XAqKMMWgnUU9g7Ldyn6sbk

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XAqKMMWgnUU9g7Ldyn6sbk

The benchmark's pack-timing row called ndarray::hpc::amx_matmul::vnni_pack_i8 directly. hpc is reached only through the ndarray::simd polyfill, so the row now times PackedI8B::pack_from, which packs into the same VNNI layout. The measured share is unchanged (23.1 us, 39% of the gemm_i8 row at 256 fields). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XAqKMMWgnUU9g7Ldyn6sbk

…n reused across GEMMs D-DT64-G2L-1. A selection named in global ids is translated once under one frame of the source's DR MapTable into the source's own row order (counted lookups). The object is bound in the source's space, B is prepared once through the translation (one map read per row), and every run reads the source's lanes in place with zero lookups. Cross-space reads without a translation stay refused (ADMISSION B, #1640). One AMX VM, three runs: translate once 68-85 us (3200 lookups), prepare B 37-47 us, warm run 16-22 us with 0 lookups; re-translating per operation 128-166 us (about 8x). Six falsifiers in tests/integration/g2l.rs, four disable runs red. Contract doc G2L row moves from OPEN to measured (lab). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XAqKMMWgnUU9g7Ldyn6sbk

coderabbitai · 2026-11-10T18:18:53Z

Important

Trigger review

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

Run configuration

Source mode: Auto
Repository: Unavailable
Branch: Unavailable
File: No configuration file from this repository applied
Commit: Unavailable
Layers: Organization UI
Review profile: CHILL
Plan: Advanced
Run ID: ce354266-566f-4b49-8dcd-9d1758b8e46d

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

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

chatgpt-codex-connector · 2026-11-10T18:26:38Z

Codex Review Summary: ✅ Completed

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

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-11-10T18:27:45.347802Z	`041ca3a`	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: 041ca3a405

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".

TranslatedB::extent() reported the source's position range, so an identity past the inner table passed admit_operand and packing indexed past the table. extent() now reports the inner operand's own extent, and BValues gains highest_row(selection), which TranslatedB overrides to resolve each selected position to the identity it reads. admit_operand then refuses with ShapeRefusal::OperandShape before any packing (#1990's contract). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XAqKMMWgnUU9g7Ldyn6sbk

The G2L bench now reports separately: G2L lookups and translation time; admission and packing map reads during B preparation; B bytes packed and preparation time; A staging bytes and time per warm run; GEMM-only time; and the complete warm run. GEMM alone is 0.71-0.86 us; staging the sparse selection into a padded A is 11-14 us of a 20-24 us warm run, so zero lookups is not zero data movement. BENCH-SHAPE, the board entry and the contract doc record that K order changes under translation and is validated only for exact integer GEMM. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XAqKMMWgnUU9g7Ldyn6sbk

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XAqKMMWgnUU9g7Ldyn6sbk

…z9-g2l-lab

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XAqKMMWgnUU9g7Ldyn6sbk

Two sources built in the same order have equal coordinate digests, so Translation::check admitted a translation from the other one and packing then panicked reading a table with a frame it never issued (Codex, #1994). check now asks the source's own map table whether it issued the translation's frame, and refuses with G2lRefusal::Foreign before packing. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XAqKMMWgnUU9g7Ldyn6sbk

AdaWorldAPI · 2026-11-10T18:55:46Z

CI: excluded-crate fmt, clippy and tests is red until ndarray#418 merges. The only error is E0432: unresolved imports ndarray::simd::{int8_gemm_packed_strided, PackedI8B} in src/int8.rs. CI builds the lab against ndarray master, and both items arrive with #418. #1992 is blocked by the same thing.

Locally, against #418's head (4b87d85), all 55 lab tests pass in release mode and clippy -D warnings is clean. Once #418 merges, a re-run of this check should go green with no change here.

Generated by Claude Code

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XAqKMMWgnUU9g7Ldyn6sbk

claude added 5 commits November 10, 2026 17:51

dr-isa-stream-lab: record the INT8 direct arm results

d024630

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XAqKMMWgnUU9g7Ldyn6sbk

wip: G2L lab, one witnessed translation reused across GEMMs (untested)

ecb9d99

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XAqKMMWgnUU9g7Ldyn6sbk

AdaWorldAPI marked this pull request as ready for review November 10, 2026 18:26

chatgpt-codex-connector Bot reviewed Nov 10, 2026

View reviewed changes

Comment thread crates/dr-isa-stream-lab/src/g2l.rs Resolved

Comment thread crates/dr-isa-stream-lab/src/g2l.rs Resolved

claude added 6 commits November 10, 2026 18:42

Merge main into the INT8 direct arm branch

b37ebac

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XAqKMMWgnUU9g7Ldyn6sbk

Merge branch 'ccr-8062bcc7-l7lkz9-int8-direct' into ccr-8062bcc7-l7lk…

49da4cd

…z9-g2l-lab

board: regenerate entries index after merging main

2969164

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XAqKMMWgnUU9g7Ldyn6sbk

Merge main (with #1992 squashed) into the G2L lab branch

34b6ef3

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XAqKMMWgnUU9g7Ldyn6sbk

AdaWorldAPI changed the base branch from ccr-8062bcc7-l7lkz9-int8-direct to main November 10, 2026 19:15

AdaWorldAPI merged commit 5ffca72 into main Nov 10, 2026
13 checks passed

Navigation Menu

Repository navigation

dr-isa-stream-lab: G2L falsifier — one witnessed translation reused across GEMMs (D-DT64-G2L-1) - #1994

dr-isa-stream-lab: G2L falsifier — one witnessed translation reused across GEMMs (D-DT64-G2L-1)#1994
AdaWorldAPI merged 12 commits into
mainfrom
ccr-8062bcc7-l7lkz9-g2l-lab

AdaWorldAPI commented Nov 10, 2026 •

edited

Loading

Uh oh!

coderabbitai Bot commented Nov 10, 2026 •

edited

Loading

Uh oh!

chatgpt-codex-connector Bot commented Nov 10, 2026 •

edited

Loading

Uh oh!

chatgpt-codex-connector Bot left a comment

Uh oh!

Uh oh!

Uh oh!

AdaWorldAPI commented Nov 10, 2026

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

AdaWorldAPI commented Nov 10, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What

Measured (one AMX VM, three runs; each quantity on its own)

Operand extent (review fix)

Falsifiers (tests/integration/g2l.rs), each disable-verified red

Uh oh!

coderabbitai Bot commented Nov 10, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

chatgpt-codex-connector Bot commented Nov 10, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

Uh oh!

AdaWorldAPI commented Nov 10, 2026

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

AdaWorldAPI commented Nov 10, 2026 •

edited

Loading

Falsifiers (`tests/integration/g2l.rs`), each disable-verified red

coderabbitai Bot commented Nov 10, 2026 •

edited

Loading

chatgpt-codex-connector Bot commented Nov 10, 2026 •

edited

Loading