simd: U64x8::transpose8 — the 8×8 u64 cross-lane transpose, all six backends

[AdaWorldAPI]

What

U64x8::transpose8(rows: [U64x8; 8]) -> [U64x8; 8], where out[i] lane j == rows[j] lane i. It is implemented on every backend:

backend	body
AVX-512	8 unpacklo/hi_epi64, then 16 vshufi64x2, for 24 shuffles
AVX2	four 4×4 blocks (unpack + vperm2i128, 8 each); the off-diagonal blocks trade places by choice of destination only
NEON	32 vtrn1q_u64/vtrn2q_u64 on 2×2 blocks; block placement is register renaming
wasm	32 i64x2_shuffle, same structure as NEON
scalar / nightly	the index map (this is the reference the other backends are tested against)

Why a transpose

Argon2's Block::compress needs exactly one cross-lane move. I counted how many words change lane between stages (8 lanes, choice of register free):

layout	load→row	in row pass	row→col	in col pass	col→store	total
vertical (lane = permutation)	112	0	112	0	112	336
horizontal (lane-rotation diagonalize)	64	96	112	96	112	480
greedy per stage	64	96	96	64	112	432

• In the vertical layout, BLAKE2b's diagonal step is register renaming, so no data moves inside a pass.
• Row→column is the one step that truly forces data across lanes: a column G reads rows 0, 2, 4 and 6, which sit in different lanes. This primitive does that move in registers.
• The load and store columns come from keeping blocks in canonical word order. Storing Argon2's blocks transposed removes them, which leaves 224. That is a follow-up in the password-hashes fork.

Verification

• simd::tests::u64x8_transpose8_matches_the_index_map:

  • uses 64 distinct words;
  • asserts the fixture is not symmetric, so an identity "transpose" would fail;
  • checks that transposing twice returns the input.

  It passes on native AVX-512 and on pinned x86-64-v3.

• The NEON parity arm (qemu) and the wasm parity arm (node) pass with the new check 0x310.

• Disable runs. I broke each backend in turn and confirmed its check fails:

  • AVX-512: changed the chunk selector from 0x88 to 0xDD;
  • AVX2: changed vperm2i128 from 0x20 to 0x31;
  • NEON: swapped vtrn2q for vtrn1q (rc 784);
  • wasm: swapped shuffle <1,3> for <0,2> (rc 784).

• Full cargo test --lib on native: 2483 passed. clippy -D warnings is clean on native and on v3.

• Not run: the nightly-simd arm.

🤖 Generated with Claude Code

https://claude.ai/code/session_019HnekoM1EidTwQLS3oFVFm

---

Generated by Claude Code

[coderabbitai]

Navigate logical layers of code changes, visualize relationships, and explore their blast radius.

Note

Currently processing new changes in this PR. This may take a few minutes, please wait...

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Essentials

Run ID: 16315b19-ad23-4092-92a0-1ec323d2d2a4

📥 Commits

Reviewing files that changed from the base of the PR and between 90cc821 and 873835d.

📒 Files selected for processing (10)

• .claude/blackboard.md
• crates/neon-simd-parity/src/main.rs
• crates/wasm-simd-parity/src/lib.rs
• src/simd.rs
• src/simd_avx2.rs
• src/simd_avx512.rs
• src/simd_neon.rs
• src/simd_nightly/u_word_types.rs
• src/simd_scalar.rs
• src/simd_wasm.rs

 ________________________________________________________
< Congrats on shipping. Unfortunately, you shipped *it*. >
 --------------------------------------------------------
  \
   \   (\__/)
       (•ㅅ•)
       / 　 づ

✨ Finishing Touches

📝 Generate docstrings

• Commit to this branch
• Create a new PR

---

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

[cursor]

Bugbot couldn't run - usage limit reached

Bugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit.

A user or team admin can review and increase usage limits in the Cursor dashboard.

(requestId: serverGenReqId_3c24c383-cb16-45c8-a83f-78f58e9ab85b)