[AdaWorldAPI]

Stacked on #1983, and based on its branch. Retarget to main once #1983 merges. Both need AdaWorldAPI/ndarray#414.

What this adds

A virtual CSR (vcsr.rs). Graph adjacency in the lanes' own 256 × 256 cube coordinates.

• The cube is the whole 64K universe: one lane is one 256 × 256 matrix.
• The edge relation is identity × identity, so it is 256 × 256 tile pairs, each a 256 × 256 bitmask block.
• Each row tile lists its nonempty column tiles, and an edge is addressed by (tile_r, tile_c, slot_r, slot_c).
• There are no row_ptr or col_idx arrays. ExplicitCsr is the materialized form of the same edges, kept only as the oracle.

Neighbour aggregation (spmm.rs). Y[p][i] = Σ_{j ∈ row(i)} lane_p[j] (plus-times) over 16 DT64-bound lanes, as BF16 GEMMs per block:

• ndarray strided tile GEMM (AMX TDPBF16PS on this host). A is read in place from the slab; B is staged per 16-row group.
• ndarray's own backend::cblas_gemm_bf16bf16f32. This is the CBLAS-compatible entry (AMX, then AVX-512 BF16, then scalar), with no external library. It takes dense &[u16] and no lda, so A is staged per block.
• Staged-A tile path, for comparison.

bench_all. Every arm of #1983 and of this PR, oracle-checked, in one table. The result is committed as crates/dr-isa-stream-lab/BENCH.md.

blasgraph's CsrStorage (BitVec values, HDR semirings) is not realized by a floating-point GEMM. Its adapter is a named seam, not wired.

Benchmark highlights (AMX host, n = 65,536, 16 lanes)

Workload	Best arm	Time	ndarray CBLAS	Scalar CSR
SUM WHERE, 16 × 16 masks, whole 64K	tile GEMM, A in place	172 µs	3.4 ms	—
Graph, degree 8, local	scalar CSR	11.6 ms	581 ms	11.6 ms
Graph, degree 64 within the tile	tile GEMM	12.5–15.3 ms	51 ms	75 ms

• The WHERE terminal is the only arm inside the µs gate.
• ndarray's CBLAS entry repacks on every call. matmul_bf16_to_f32 copies both operands, and bf16_tile_gemm_16x16 VNNI-packs B and allocates each call, so it runs about 20× slower than the strided tile path on the same problem.
• Block GEMM cost is set by block area, not by edges. It loses to a scalar CSR loop at about 0.15% fill and wins 5–6× at 25% fill.
• The block count depends on locality in the G2L order. One long-range edge per node grows it about 8×.

Decision: no production crate yet

Every graph arm misses the gate. Even the dense case is bounded at about 1.4 ms by its own 268M MACs. The best arm on the sparse graph is the scalar loop.

The WHERE terminal meets the gate. If it is promoted, it belongs in the Quack / mask-risc terminal set behind DT64 binding, not in a new crate.

Revisit conditions are in .claude/board/entries/2026-11-08-dr-virtual-csr-bf16-gemm.md.

Correctness

• Virtual equals explicit: the virtual CSR matches the explicit CSR row by row.
• Exact against the oracle:
  • tile GEMM (direct and staged) at n = 256, 513, 4096 and 65,536;
  • ndarray CBLAS at 513 and 65,536.
• Lane borrows: 16 per aggregation, whatever n is.
• Refusals: a graph in another row order, a stale frame, and fewer than 16 lanes for direct mode.
• Orientation: a one-edge test checks that row i gets lane[j], not lane[i].
• Disable runs, each going red:
  • the graph-space check;
  • the empty-group skip;
  • B orientation;
  • the tail K;
  • the CBLAS accumulator.
• Totals: 27 tests pass, and clippy with -D warnings is clean.

🤖 Generated with Claude Code

https://claude.ai/code/session_01XAqKMMWgnUU9g7Ldyn6sbk

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

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: c9ef7e73-6412-47d9-97e0-3eac77b35aeb

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• Trigger review

• 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

• X
• Mastodon
• Reddit
• LinkedIn

[chatgpt-codex-connector]

Codex Review Summary: ✅ Completed

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

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-11-08T00:42:45.347922Z	686c443	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.

[AdaWorldAPI]

excluded-crate fmt, clippy and tests is red for the same reason as #1983. On a runner without AMX, ndarray master's polyfill fallback rejects a strided A longer than 16·K (A must be 16xK BF16 row-major). The fix is AdaWorldAPI/ndarray#415, and I'll re-run this job after it merges. This branch also now carries #1983's bf16 exactness fix (merge f252621).

---

Generated by Claude Code