Skip to content

Investigation — K5 gather vs scalar loads on AVX2 (REQ-K5-004, REQ-KERNEL-007)

Status: OPEN — decision held by prior, no measurements yet. This page contains no performance numbers: none have been measured on registered hardware (Charter T2 — no invented data).

Question

Should the AVX2 take backend use hardware gather (vpgatherdd/vpgatherdq/vpgatherps/ vpgatherpd) or the scalar 4×-unrolled independent-load path?

Current state (v0.2, gate M4)

  • Both paths are compiled in src/kernels/take/take_avx2.cpp behind the kUseGatherTake constexpr switch (PRD 08 §5 K5 requires both to exist).
  • The shipped default is scalar unrolled loads, held by the Survey §4.2 prior: on Haswell through Zen 3, gather is microcoded or issue-limited and does not beat scalar independent loads for cache-resident batches; the OoO core already extracts the MLP.
  • The evidence-gated measurement could not be run at M4: the only machine available to the project is an Apple M-series (ARM64) host, and x86 emulation timing is not valid evidence. Recorded as a deferral in docs/releases/gates/M4.md.

Domain constraint discovered during implementation

vpgatherd* sign-extends its 32-bit indices, so the gather path is only correct while every idx < 2^31. The scalar default has no such limit (indices are uint32_t up to values.len). Any future flip to gather must either keep a guard for large-index batches or restrict the gather path to values.len ≤ 2^31. Debug-asserted in the gather path.

Measurement protocol (to run on first registered x86 machine — M5 ledger work or later)

  1. Machine registered per REQ-LEDGER-013 (pinned frequency governor, isolated core set).
  2. bench_take dict-size sweep (L1/L2/L3/DRAM-resident) × {i32, i64, f32, f64} × index patterns (sequential, random, clustered), gather build vs default build (-DQUIVER_K5_TAKE_GATHER toggle to be added with the experiment).
  3. Decision rule (pre-registered): flip the default only if gather wins by ≥ 10% geomean across the dict-size sweep on ≥ 2 distinct µarchs, with no regression > 5% on any swept point; otherwise keep scalar and record the verdict in the ledger.

Reopening criteria

Registered access to any of: Intel Skylake-SP/Ice Lake+ (gather improved), AMD Zen 4+, or any µarch where the ledger shows K5 far below its MLP roofline.

Traceability: REQ-K5-004, REQ-KERNEL-007; Survey §4.2; gate M4 deferral.