Refactor files: tiled vs multi-level, pt.3 - #535
Merged
Conversation
gemv and quantized_matmul move to multi_level/eval/; gemm_cpu_tiled, split_k and tile_quant_stage to tiled/eval/ (split_k and tile_quant_stage are written on the tile DSL, not multi-level as the plan's sketch assumed). gemm and gemm_cpu stay at the root: they exist to put both families on one table. gemm's strategies() is now a concatenator over tiled::eval::gemm::strategies() and multi_level::eval::gemm::strategies(), so each branch owns its own entries and the root only decides that both appear. gemm_cpu filters gemm's list by id and needs no change at all. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Both on by default; either compiles without the other, and a build with neither fails with a compile_error rather than a pile of missing items. `Strategy`'s two variants, their `From`/`Display`/dispatch arms and the `gemm` catalogue's two entry lists are cfg'd; `Auto` is SimpleCyclicCmma → SimpleUnit whenever multi-level is compiled, and Cmma → CpuGemm in a tiled-only build. Two items were on the wrong side of the seam, both found by the single-family builds warning they were dead: - `M`/`N`/`K`, `batch_axis` and `MatmulOperands` are tile-DSL vocabulary that only the tiled routines use. They move to `tiled/operands.rs`, which makes cubek-tile a tiled-only dependency (`tiled = ["dep:cubek-tile"]`). - `MatmulPrecision` / `MatrixPrecision` and `MatmulElems::new_deprecated` are a pure element-type mapping that both families' harnesses launch through. They move to the shared `definition/precision.rs`. `MatmulTypes`/`MatrixTypes` and the alias zoo stay multi-level. cubek-attention and cubek-convolution declare `features = ["multi-level"]`; the cubek facade gains `matmul-tiled` / `matmul-multi-level` passthroughs. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
tests/{harness,tiled,multi_level}, with the `matmul/` level dropped. harness/ holds
the launcher plumbing, the problem builders and the stride-0 broadcast helpers, and
carries no test of its own. Every tier (basic/extended/full/benchmarks) now crosses
the branch axis inside its family.
Three files had to be cut in two: basic/plane_accelerated.rs at the tile-DSL comment,
extended/stride_zero.rs (the strategy table is multi-level, the CpuGemm entry tiled,
the helpers now harness/broadcast.rs), and bench_catalog.rs, which becomes a
per-family correctness file plus multi_level/comparison.rs for the dsl-vs-legacy
probes — those exist only while both families do, so they die with the island and
are cfg'd on `tiled` inside it.
Deviation from the plan's "zero #[test] outside tiled/ and multi_level/": the
`Strategy::Auto` suite belongs to neither island and correctly survives deleting
either, so it stays at tests/auto/ (a directory module, since a bare tests/*.rs
would become its own test binary).
All six {tiled,multi-level} x {basic,extended,extended+benchmarks} combos compile
and clippy clean; the basic tier runs the same 47 tests as before.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
`cargo xtask test` runs cubek-matmul twice more per backend, once with each family
alone, and CI clippies both single-family builds at the extended tier. Without a job
that compiles them, the cfg attributes rot.
The first tiled-only run showed the seam was not actually being tested: the package's
own dev-dependency (`cubek-matmul = { path = "." }`, for `cpu-reference`) carried
default features, so every test build unified both families back in. With
`default-features = false` on it, a tiled-only test binary really holds only the
tiled and Auto tests: 9 + 9 against the 15 + 47 of a full build, and 15 + 39 for
multi-level alone. All three pass on cubecl/cpu.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Against metabolic's contributor book: - `mod.rs` holds no code: the Auto suite moves to `tests/auto/base.rs`. - No doc dates itself: the three re-export comments that said "back-compat", "have always used" and "keep working" now state what the re-export is for. - No analogy on my own initiative and no "legacy": the comparison probes name multi-level, and nothing in the code calls either family an island. - No em dashes in comments. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
`tiled` and `multi_level` are two architectures. "Family" is the other sense in this crate: a trait with implementations (`BatchMatmulFamily`, the gemm family), and one word cannot mean both. Renames the prose, the CI step names and the xtask loop; leaves every pre-existing `*Family` alone. The CI comment also said what the default is rather than what the steps do: every other job compiles both architectures, so nothing else notices a `#[cfg]` that stopped being true. TILED_SPLIT_PLAN.md is deleted. The split is done and the code is the record. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
# Conflicts: # crates/cubek-matmul/src/definition/mod.rs # crates/cubek-matmul/src/eval/benchmarks/gemm/strategy.rs
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.