Skip to content

Claude/continue session 0m a va#27

Merged
AdaWorldAPI merged 6 commits intomasterfrom
claude/continue-session-0mAVa
Mar 24, 2026
Merged

Claude/continue session 0m a va#27
AdaWorldAPI merged 6 commits intomasterfrom
claude/continue-session-0mAVa

Conversation

@AdaWorldAPI
Copy link
Copy Markdown
Owner

No description provided.

claude added 6 commits March 23, 2026 12:03
Minecraft-style variable bit width encoding for palette indices (1–8 bits).
Smaller palettes use fewer bits: 4-entry palette → 2 bits → 4× compression.

Core API:
  - bits_for_palette_size(n) → minimum bits needed
  - pack_indices(&[u8], bits) → Vec<u64>
  - unpack_indices(&[u64], bits, count) → Vec<u8>
  - pack/unpack_indices_bytes() — serialization-friendly byte variants
  - transcode(packed, old_bits, new_bits, count) — in-place width change
  - compression_ratio(bits) → f32

PackedPaletteArray wrapper:
  - from_indices() — auto-selects bit width from palette size
  - get(i)/set(i, val) — O(1) random access into packed data
  - grow_palette(new_size) — widen bit width when palette grows
  - compression_ratio() — actual compression achieved

18 tests covering: all bit widths (1-8), non-aligned boundaries,
byte roundtrip, transcode, grow, compression, empty/single edge cases.

https://claude.ai/code/session_01CdqyUTUfjKZuk8YGJzv6LB
- packed.rs: add // SAFETY: comments to both prefetch unsafe blocks
- packed.rs: fix prefetch_t1 doc (was "non-temporal", correct is "L2 temporal")
- vml.rs: document vspow domain restriction (a[i] > 0 required for SIMD path)

Sentinel-QA verdict: BLOCK → PASS after these fixes.

https://claude.ai/code/session_01CdqyUTUfjKZuk8YGJzv6LB
- simd.rs: allow unused_imports on re-exports (intentional public API surface)
- simd_avx512.rs: allow dead_code (types for consumer use, not yet all consumed)
- hpc/mod.rs: remove duplicated #[allow(missing_docs)] attribute
- hpc/zeck.rs: remove unused popcount_raw import
- hpc/compression_curves.rs: prefix unused gt_top10 with underscore, allow dead_code
- hpc/cyclic_bundle.rs: allow dead_code (benchmark utilities)
- hpc/blas_level2.rs: use saturating_sub and abs_diff per clippy
- backend/kernels_avx512.rs: use += assign operators per clippy
- hpc/parallel_search.rs: remove redundant closure

https://claude.ai/code/session_01CdqyUTUfjKZuk8YGJzv6LB
Three patterns (SIMD slices, Copy microcopies, gated write-back)
and the no-&mut-self-during-computation rule.

https://claude.ai/code/session_01CdqyUTUfjKZuk8YGJzv6LB
Replace manual index arithmetic with chunks_exact/chunks_exact_mut zip
patterns. Eliminates bounds-check overhead for smaller MKL-sized arrays.
Packed database uses chunks_exact for row iteration.

https://claude.ai/code/session_01CdqyUTUfjKZuk8YGJzv6LB
Port jitson (JSON config → native code via Cranelift JIT) into hpc::jitson.

Module structure:
- ir.rs: ScanParams, PhilosopherIR, RecipeIR, JitError (always available)
- detect.rs: CpuCaps with Copy derive (always available)
- engine.rs: JitEngine with RwLock cache, &self compile path (jit feature)
- scan.rs: ScanKernel + Cranelift IR generation (jit feature)

Key design decisions:
- Cranelift deps behind optional `jit` feature flag (heavy deps)
- JitEngine::compile_scan takes &self (not &mut self) per data-flow rules
- Cache uses RwLock with double-checked locking for thread safety
- CpuCaps is Copy; ScanParams remains Clone (contains Vec)
- Cranelift 0.116 (crates.io published) instead of wasmtime fork path deps

8 new tests (ir + detect), 1036 total lib tests passing.

https://claude.ai/code/session_01CdqyUTUfjKZuk8YGJzv6LB
@chatgpt-codex-connector
Copy link
Copy Markdown

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, you can upgrade your account or add credits to your account and enable them for code reviews in your settings.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants