Skip to content

Phase 7e-b (partial) — Tier 1 matmul_blocked pure-Li IKJ tiles

Summary

benchmarks/tier1_micro/matmul_blocked/li/main.li runs 512× (64×64, block=16) IKJ micro-tiles (≈512³ flops) with no LI_EXTRA_C; harness marks matmul_blocked as li_pure=True.

Agent continuation

  1. Read benchmarks/tier1_micro/matmul_blocked/li/main.li and matmul_naive/li/main.li for tile/rep patterns.
  2. Run ./li-tests/run_all.sh and python3 benchmarks/harness/bench.py --tier 1 (expect matmul_blocked verify ok (pure Li)).
  3. Then merge stack #138→#144→this; benchmarks dashboard ingest after tag; optional full 512×512 when stack arrays scale.
  4. Blocked on slice syntax for true 512³ single allocation (v1 uses tiled equivalence).

Changed

Path What
benchmarks/tier1_micro/matmul_blocked/li/main.li IKJ blocked loops + C[i][j] = C[i][j] + aik * B[k][j]
benchmarks/harness/bench.py li_pure=True for matmul_blocked
benchmarks/results/README.md pure_li lists matmul_blocked

Not changed

  • Full 512×512 Li arrays on stack (still 64×64 tiles × 512 reps).
  • simd_dot / ArrayBinOpF64 SIMD codegen.
  • Compiler MIR (uses existing 2d load/store + binops).
  • benchmarks org catalog.toml ingest.

Breaking

N/A — Li column label pure lic; cpp/rust/julia still use shared C kernel.

Security

N/A.

Performance

Smoke: pure-Li ~0.013s vs native ~0.011s on tier-1 verify (not parity-gated). GFLOPS uses flops_per_run=2×512³.

Downstream

  • Master plan 7e-b — matmul_blocked partial alongside matmul_naive.