[aferaudo]

Summary

• The codelets build on arm64. With -target bpf, clang does not search the multiarch include directory, where Ubuntu keeps glibc's bits/ headers (/usr/include/aarch64-linux-gnu on arm64). Every codelet failed with 'bits/libc-header-start.h' file not found. codelets/Makefile.defs now adds the host's multiarch directory, as jbpf's own test codelets do.
• NEON path in e3_slot_convert.h. The shared convert header had only an AVX2+F16C loop. Its build guard let aarch64 through as if a NEON loop existed, so on ARM the gNB-side helper silently ran the scalar convert. The new loop gives the same bytes as AVX2+F16C. The microbenchmark builds on arm64 and hides NEON in its scalar mode.
• Paired with wineslab/ocudu-e3-private#5: the convert header is shared byte for byte, so merge both together. The controller, libe3, the verifier and the adaptive_cpu dApp already built on arm64 unchanged.

Type of change

• Bug fix (scalar convert on aarch64)
• New feature / enhancement (arm64 build)
• New Service Model
• Refactor (no behavior change)
• Documentation
• Test / CI / packaging
• Other (explain):

Linked issue

None.

Mandatory test checklist

• arm64 build on NVIDIA Grace (Neoverse-V2), Ubuntu 24.04, GCC 13.3, clang 18:
  • ./build.sh and the verifier build;
  • make -C codelets builds both codelets and they pass the verifier (632 and 345 instructions);
  • no AVX/F16C/scalar warnings.
• NEON = AVX2+F16C: for all 65,536 bf16 values at 6 scales, the NEON output is byte-identical to the AVX2+F16C output.
• Microbenchmark on Grace: vector vs scalar agreement … IDENTICAL. For 4 ports, NEON takes 33.9 µs and the scalar convert 331 µs.
• Over the air on aarch64, with the paired ocudu PR:
  • Setup: Foxconn RU, n78, 100 MHz, 4T4R, a COTS phone, the adaptive_cpu dApp over JSON.
  • The codelet loads through jbpf's arm64 JIT, and the gNB helper logs convert path=NEON.
  • Slots 8, 9, 18 and 19 reach the dApp at 100/s each. dApp IQ age p50 is 0.1 ms (0.7 ms with scalar). The dApp check passed.
• x86: not re-run on hardware. On x86, codelets/Makefile.defs now also adds /usr/include/x86_64-linux-gnu, and the AVX2 loop is unchanged apart from the macro that selects it. CI builds x86.
• Provenance: tested on top of Send the gNB's uplink symbol mask in every L1-KPM indication #11. Rebased onto main with no conflicts; the changed files are identical.
• VERSION, libe3.pc, new dependencies: N/A, none changed.

CI checklist

• build and measure is green for the head commit (to check once CI has run).
• Commit policy: check_commit_trailers.py --base origin/main --head HEAD passes (Assisted-by trailer).

Twin-repo coordination

• This PR does not change the E3 wire protocol or public ABI, OR a paired PR exists in each affected twin repo (link below).

codelets/include/e3_slot_convert.h must stay byte-identical to ocudu's ocudu_janus/e3/e3_slot_convert.h: ocudu's configure compares them. Paired PR: wineslab/ocudu-e3-private#5.

Workflow confirmation

• Linear, fast-forward-able descendant of main (fbdf325), with no merge commits.
• Two commits, each with an imperative subject and a what/why body.
• I have read and followed CONTRIBUTING.md (libe3's, applied to this repository).

🤖 Generated with Claude Code