Datasets:

---

arm-bench

/

arm-bench-trace

Like 2

Follow

CMU l3 3

Summary

Adds 18 KleidiAI definitions, each with workloads, a reference-scalar solution and a baseline-kleidiai-arm solution, built with kgen (add definition -> add workload -> add solution -> calibrate). The SME2 tier now has 20 KleidiAI kernels. Every new definition passes all five kgen check legs (static, reference, reference-scalar, baseline, diagnosis) on its target hardware.

It also moves two existing quantized definitions from a hand-set min_sqnr_db to the baseline-measured baseline_sqnr_db gate, and moves gemm_bf16_n512_k512 from SQNR to an elementwise comparison.

What is in this PR

Graviton tier (5 definitions, N/K as in the name, 13 workloads each, M = 1..2048)

SME2 tier (13 definitions, 13 workloads each, M = 1..2048; conv2d has 16)

gemm_i32_u8 and gemm_u8_f32 use the new-style kai_matmul_uker API; the others use the per-kernel API.

Changes to existing files

• workloads/gemm/gemm_qs4cx_sme2_n256_k512.jsonl: added baseline_sqnr_db (44.3-50.1 dB, measured on Apple M4). Its gate was a hand-set 40 dB min_sqnr_db.
• workloads/gemm/gemm_qs4c32_n256_k512.jsonl: added baseline_sqnr_db (48.2-49.6 dB, measured on Graviton3). Its gate was a hand-set 30 dB min_sqnr_db.
• definitions/gemm/gemm_bf16_n512_k512.json: dropped the correctness:sqnr tag; it is judged elementwise (abs_tol 0.09, worst measured baseline error 0.0835). The rest of the diff is kgen reformatting the JSON.

How it was built

• baseline_sqnr_db comes from kgen calibrate --author baseline-kleidiai-arm --baseline-author baseline-kleidiai-arm on the machine matching each baseline's ISA (Graviton4 for the four quantized Graviton definitions, M4 for SME2, Graviton3 for gemm_qs4c32). The gate is the tag minus the 1 dB margin. Calibrating with reference-scalar as the author gives ~128 dB and is not meaningful.
• Output is float32 in the raw ABI: fp16 results are widened exactly, int32 and int8 results are converted. fp16/bf16 inputs are just raw 16-bit patterns. Quantization parameters that would be scalars are one-element tensors (axis S = 1), since the raw ABI passes no scalars.
• Baselines dispatch the dedicated GEMV kernel for M = 1 on the Graviton tier (the SME2 mopa kernels take any M in one call). Weight packs are cached by content comparison (issue #80); activation quantization and packing are redone on every call.

How it was verified

• kgen check (all five legs) passes for every definition above, on the target hardware: Apple M4 (mac-m4.metal) for the SME2 tier and the two SME2 changes, Graviton3 (c7g, Neoverse-V1) for the Graviton tier and gemm_qs4c32. Reference-scalar and baseline both pass all workloads, and the diagnosis leg reports them consistent.
• kgen autovec passes for every definition above with --max-slowdown 0.5, where the geomean speedup of baseline kernels is at least 2x faster than the autovec version of reference-scalar kernels

Notes for reviewers

• Companion code PR: the elementwise tolerance overrides (gemm_qai8_qsi8cx_sme2, gemm_u8_f32_sme2, gemm_bf16, gemm_f32_bf16) and the removal of the two hand-set min_sqnr_db entries live in config/kernel_contracts.yaml in the code repository. Without that change these definitions fall back to the gemm default tolerance and gemm_qai8_qsi8cx_sme2 would fail. Code PR: .
• Gates tighten for two existing definitions (gemm_qs4cx_sme2: 40 dB -> about 43-49 dB, gemm_qs4c32: 30 dB -> about 47-48 dB). Verdicts from earlier agent runs on these two were judged against the looser gates.
• Known limitation: the weight-side memcmp that guards the pack cache runs inside the timed region (issue #85, the untimed setup hook, is not applied here). gemm_f16act_qs4c32_sme2 also widens its fp16 activation to fp32 inside the timed region, because KleidiAI's LHS packer only reads fp32.
• Nothing under runs/ is touched.