bench: arithmetic operation micro-benchmarks#805
Open
FBumann wants to merge 2 commits into
Open
Conversation
An op-level tier alongside the whole-model builds: one benchmark per (operation, size profile), operands built outside the measured region so a run isolates a single op rather than a whole build. This attributes perf changes to a specific arithmetic path — a build benchmark says "kvl got heavier", an op benchmark says "expr+expr broadcast got heavier". - benchmarks/ops.py: op registry (OpSpec) + size profiles (small 1D×2000; large 3D×3×4×1000 — differ in element count *and* dim count; the asymmetric shape also catches dim-order bugs) + ~30 ops across scaling, var/expr arithmetic, quadratic, reductions and constraint construction. Binary labelled ops carry match/broadcast variants — the alignment-path axis where the interesting regressions live. - benchmarks/drivers/test_ops.py: parametrized driver, one benchmark per (op, profile). - conftest: add test_ops to CODSPEED_MODULES (tracked; memory advisory). 60 benchmarks, ~80s/run with memory. Signal validates: large ≈ 6× small, broadcast ≈ 5× match (the §9 cross-product).
Merging this PR will not alter performance
Performance Changes
Comparing 
Footnotes
|
Collapse to one 3-D profile (3×4×1000, ~12 K elements) — CodSpeed records time *and* memory per benchmark, so a second size wasn't buying a separate signal; one multi-dim profile keeps broadcast/alignment coverage with MB-scale ops above the noise floor, and halves the matrix. Benchmark ids drop the size suffix. Add three categories: absence/masking (expr.where / fillna / absence propagation — §4–§7, the semantics-heavy surface), groupby.sum, and an N-way merge (constraint-assembly cost). 35 ops, ~45 s/run with memory.
Collaborator
Author
|
Note The following content was generated by AI. CodSpeed cost check. This PR adds 35 arithmetic micro-benchmarks (one
The bare-metal walltime job remains gated to |
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.
I really want to make sure the v1 convention doesnt introduce regressions to linopy. And i think as soon as we decide on what to do with the multiindex stuff, there will be room for improval. THis should lay a nie foundation for that.
Codspeed cost is really small, but coverage (and granularity) really improves.
Note
AI-assisted (Claude Code): implementation and this description; reviewed by me.
Adds an op-level benchmark tier to
benchmarks/, alongside the whole-model build benchmarks. One benchmark per operation, with operands built outside the measured region so a run isolates a single arithmetic op.Why. Whole-build benchmarks catch a regression but can't attribute it — a build says "kvl got heavier", an op benchmark says "
expr+exprbroadcast got heavier". (Motivated by a real regression hunt where attribution needed exactly this granularity.)What.
benchmarks/ops.py— op registry (OpSpec) + a single 3-Dgridsize profile (dims 3×4×1000; the asymmetric shape also catches dim-order/transpose bugs) + 35 ops across scaling, var/expr arithmetic, quadratic, reductions, masking, groupby, merge, and constraint construction. Binary labelled ops carrymatch/broadcastvariants — the alignment-path axis where the interesting regressions live.benchmarks/drivers/test_ops.py— parametrized driver, one benchmark per op.conftest.py—test_opsadded toCODSPEED_MODULES(tracked; memory advisory).Cost. 35 benchmarks; the memory run stays ~2–2.5 min including the model builds that dominate the job — cheap.
Signal validated.
broadcast ≈ 5× matchon the alignment axis (the §9 cross-product) — well above the noise floor.Memory is report-only to start (op-scale memory can be noisy); op-time is the natural gate candidate once the signal proves stable.