refactor(traits): device-portable PauliBackend abstraction (RFC scaffold) - #145
Closed
Roger-luo wants to merge 1 commit into
Closed
refactor(traits): device-portable PauliBackend abstraction (RFC scaffold)#145Roger-luo wants to merge 1 commit into
Roger-luo wants to merge 1 commit into
Conversation
Introduce a de-biased backend trait so a CUDA-backed map can be a drop-in `Config::Map`, enabling `CuPauliSum = PauliSum<CudaConfig>` without a second Rust type. The current `ACMap` family is host-biased in two ways that block a GPU backend: every transform takes a host Rust closure (`Fn(&W,&V)->...`), and iteration yields host references (`(&W,&V)`). Neither can cross to a CUDA kernel / device-resident map. `PauliBackend` instead reifies the (small, closed) Pauli-propagation vocabulary — Clifford 1q/2q, Pauli rotations, scale/merge/truncate/overlap — as data (gate + qubit indices + scalar coeffs), so each backend supplies a CPU loop or a GPU kernel. This commit is additive (nothing rewired yet): - docs/design/device-portable-pauli-backend.md: problem, design, CPU+CUDA impl table, migration plan, open questions. - ppvm-traits: PauliBackend trait + Clifford1q/Clifford2q enums + PackedWord (the canonical packed encoding shared with the CUDA backend). - reference::HashMapBackend: a packed-key CPU reference impl proving the shape, with the Clifford-1q family + scale/merge/truncate/overlap fully implemented and tested; branching ops marked ITERATE for follow-up. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
|
Collaborator
Author
|
hmm the problem is real here, but the abstraction is not the right shape, I prob need to redesign this myself instead of asking claude doing it... let me open an issue first |
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.
What & why
Scaffold + design RFC for a de-biased map abstraction so a CUDA-backed map can be a drop-in
Config::Map, givingCuPauliSum = PauliSum<CudaConfig>— one Rust type, no parallel device type.The current
ACMapfamily is biased toward host (CPU) semantics in two ways that block a GPU backend:Fn(&W, &V) -> …(map_add_assign,map_insert*,scale,retain,contains_with). The gate math lives in these closures inppvm-pauli-sum(sum/clifford.rs,rot1.rs,rot2.rs). You can't ship an arbitrary Rust closure to a CUDA kernel.ACMapIter::Item = (&W, &V); a device-resident map can't hand out&V, and per-term host iteration would erase the GPU win.PauliBackendinstead reifies the (small, closed) Pauli-propagation vocabulary — Clifford 1q/2q, Pauli rotations,scale/merge/truncate/overlap— as data (gate + qubit indices + scalar coeffs). Each backend supplies a CPU loop or a GPU kernel.What's in this PR (additive — nothing rewired yet)
docs/design/device-portable-pauli-backend.md— problem, design, CPU-vs-CUDA impl table, 7-step migration plan, open questions.ppvm-traits:PauliBackendtrait +Clifford1q/Clifford2qenums +PackedWord(canonical(u64,u64)encoding, shared with the CUDA backend).reference::HashMapBackend— packed-key CPU reference proving the shape: Clifford-1q (10 gates) +scale_all/merge/truncate_abs/overlap/exportimplemented and tested (6 tests); rotations + 2q Cliffords markedITERATE.cargo test -p ppvm-traitsandcargo check --workspaceare green;mainis untouched.How to iterate
This is intentionally a scaffold so it can be picked up async. The migration steps and open questions (Complex coeffs on cuco's 8-byte path, where to hoist
levi_civita/comm_2, truncation reification) are itemized in the design doc — grab a row.🤖 Generated with Claude Code