test(bench): establish reproducible Criterion host baseline (#22)

[tembo]

Establishes a reproducible Criterion benchmark baseline on main for the host/Rayon platform, ahead of the CubeCL HAL migration. Host-only, built against the current public API with --no-default-features. No public API or default features are changed. Draft PR per the issue execution contract (draft_pr_only: true, completion authority: human).

Changed files

• Cargo.toml — add criterion dev-dependency (default-features = false, cargo_bench_support) and the host_benchmarks [[bench]] target (harness = false).
• benches/host_benchmarks.rs — the six benchmark categories over the documented shape matrix (harness = false, custom main prints the environment block).
• tests/benchmark_support/mod.rs — shared deterministic fixture + environment-recording utility (included via #[path] by both the bench and the determinism test).
• tests/bench_determinism.rs — verifies the fixture is deterministic, in range, seed-distinct, and that environment probing succeeds.
• docs/benchmarking/README.md — inventory, local execution, environment recording, comparison rules, and how Introduce backend abstraction and minimal CubeCL backend #34–[Validation] Add reproducible CubeCL CPU and accelerator conformance runner #50 should add comparable CubeCL groups without changing these host baselines.
• docs/benchmarking/initial-results.md — initial result artifact + exact environment identity.
• .github/workflows/benchmarks.yml — CI step compiling benchmarks with cargo bench --no-run --no-default-features (no timing on shared CI), plus fmt check and host-only tests.
• README.md — replaced the stale benchmark-branch instructions with a pointer to the new baseline.

Cargo.lock is intentionally not committed (the repo gitignores it for a library).

Benchmark inventory and operation mapping

ha-ndarray ops are lazy; operation-only benchmarks build deterministic input data once outside the measured region, build a cheap borrowed view per iteration in the untimed iter_batched setup, and measure op construction + materialization via NDArrayRead::buffer. criterion::black_box is used throughout. construct_convert is the only intentionally end-to-end group (includes allocation + copy).

#	Category	Public API	Benchmark group
1	construction	ArrayBuf::convert(&data, shape) (alloc + copy)	construct_convert
2	unary elementwise	NDArrayUnary::exp	unary_exp
3	binary elementwise	NDArrayMath::add (same-shape)	binary_add
3	binary (broadcast)	NDArrayTransform::broadcast + NDArrayMath::add	binary_add_broadcast
4	reduction	NDArrayReduceAll::sum_all	reduce_sum_all
5	matrix multiply	MatrixDual::matmul	matmul
6	non-contiguous view	NDArrayTransform::transpose (materialized)	transpose

Substitution note: axis-reduction (NDArrayReduce::sum) requires an owned 'static accessor via the public reduce_axes path (Accessor::from(A) has a B: 'static bound), so sum_all (available on borrowed inputs) is used as the representative reduction. No product functionality was added to make a benchmark possible.

Shape matrix: elementwise/reduction/transform use [16,16] (latency), [256,256] (medium), [1024,1024] (throughput); matmul uses [16,16], [96,96], [256,256]. IDs encode dtype/shape (e.g. matmul/f32/256x256).

Exact commands executed

cargo fmt --all -- --check
cargo test --no-default-features
cargo bench --no-run --no-default-features
cargo bench --no-default-features

Environment / toolchain identity

rustc: rustc 1.96.0 (ac68faa20 2026-05-25)
cargo: cargo 1.96.0 (30a34c682 2026-05-25)
os: linux
arch: x86_64
cpu_model: Intel(R) Xeon(R) Platinum 8275CL CPU @ 3.00GHz
logical_cores: 2
physical_cores: 2
feature_set: no-default-features (host-only, Rayon)

Benchmark compilation result

cargo bench --no-run --no-default-features succeeds in a clean checkout (Finished \bench` profile [optimized]`).

Initial Criterion result artifact

All 21 benchmarks; Criterion reports [lower mean upper]. Full machine-readable output is in target/criterion/<group>/<bench>/new/estimates.json (not committed); see docs/benchmarking/initial-results.md for the complete table. Highlights:

Benchmark	Mean
construct_convert/f32/1024x1024	345.59 µs
unary_exp/f32/1024x1024	2.3161 ms
binary_add/f32/1024x1024	857.92 µs
binary_add_broadcast/f32/1024x1024	9.8600 ms
reduce_sum_all/f32/1024x1024	623.98 µs
matmul/f32/256x256	14.574 ms
transpose/f32/1024x1024	12.435 ms

These numbers are specific to the machine above and are a baseline, not a threshold.

Total benchmark duration

cargo bench --no-default-features (21 benchmarks, sample_size=10, warm_up_time=1s, measurement_time=3s): 103 s wall time.

Test results

cargo test --no-default-features — all host-only tests pass, including the 5 new bench_determinism tests (arithmetic 2, bench_determinism 5, compare 7, cond 1, construct 3, linalg 2, reduce 3, transform 9).

Known limitations / unavailable validation

• No OpenCL / all-feature validation (explicit non-goal; left to Introduce backend abstraction and minimal CubeCL backend #34–[Validation] Add reproducible CubeCL CPU and accelerator conformance runner #50).
• Axis-reduction not benchmarked (owned 'static accessor requirement); sum_all used instead.
• Batched/broadcasted matmul not in the baseline shape matrix.
• Results are machine-specific and must not be compared across machines.
• No performance threshold, supported-hardware claim, or backend-retirement decision is introduced.

Implements #22.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: c80103fe81

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Pin the host benchmarks to the host platform

When this target is compiled with --features opencl or --all-features—which the documented future accelerator groups will require—ArrayBuf uses the top-level Platform, whose select implementation in src/platform.rs dispatches inputs of at least GPU_MIN_SIZE to OpenCL. Consequently, these supposedly stable host groups (including construction at line 87) will benchmark OpenCL for the medium and large cases while still reporting host-only, Rayon, corrupting the intended host-versus-accelerator comparison. Use explicitly host-backed array types or otherwise prevent these groups from being built on the feature-dependent top-level platform.

Useful? React with 👍 / 👎.