✨ Synthesize two-qubit blocks in the native basis

[burgholzer]

🤖 AI text below 🤖

Description

Both compile_for_target and synthesize_for_target now resynthesize constant two-qubit blocks in the selected native basis while preserving cheaper native operations. For example, non-native RZZ(0.3); RXX(0.4) needs two CZ gates together instead of four individually; a native SWAP/RZZ pair stays at two gates when block synthesis would need three. Unrestricted targets also optimize their blocks.

Both pipelines reuse the existing fusion implementation at two stages:

• Before placement, replace a block only when both its two-qubit IR operation count and native cost decrease. This removes cancelled interactions before routing and keeps compact gate pairs intact. Native support is checked without assigning physical sites.
• After placement/routing, resynthesize only when native two-qubit cost decreases, using the assigned sites and operand direction. Individual lowering preserves supported native gates and supplies the comparison cost.

Run scanning, Weyl decomposition, phase handling, and emission remain shared. Numerical block failures fall back to individual lowering; the standalone target-independent fusion factory remains available. The selected synthesis basis uses one entangler family and does not search arbitrary native mixtures or calibration costs. No new dependencies or block IR are added.

Validation and measurements

Release CTest: 3,537 passed, with one existing QDMI skip. All 90 Python MLIR tests, whole-file C++ lint, repository lint, and complete executable documentation/internal link checks passed. The final target-synthesis suite passes all 61 tests. Coverage includes full-unitary and phase preservation, pre-routing cancellation, native preservation, reversed connectivity, unrestricted optimization, numerical fallback, and routed measurement results.

Benchpress Small compares upstream main 8d9bcfb03 with this change: 42 circuits across four abstract topologies, five alternating samples per version and case (1,680 successful samples). Both builds use GCC 13.3, LLVM/MLIR 23.1.0, the release preset with LTO disabled, CPU 0 affinity, one routing trial, and seed 42. Targets use the id/sx/x/rz/cz basis with measurement, reset, and global phase. Timing includes the compiler pipeline, QC/QCO conversion, cleanup, and OpenQASM export; file import and target construction are excluded.

Topology	Compile time	Native two-qubit count	Two-qubit depth
All-to-all	+0.4%	unchanged	unchanged
Square	+0.2%	-2.0%	-2.2%
Heavy-hex	-0.6%	-5.4%	-5.9%
Linear	-0.4%	-4.8%	-5.1%

Changes are geometric means of paired ratios, using median compile times; lower is better. Native count improves in 41 cases, ties in 127, and never regresses. Depth improves in 38 of 152 comparable cases and never regresses; structured-control-flow cases have no depth metric. Zero values are excluded only from geometric means. Output hashes and quality metrics are deterministic within each version.

Runtime is broadly unchanged: summed case medians increase 0.86%, and the largest per-case median slowdown is 3.60% (adder_n10, all-to-all, 17.916 to 18.562 ms). Separate long-block probes preserve full matrices, including phase, and reduce 1,024-gate RZZ/RXX synthesis from 2,048 native entanglers to two. These results cover the stated workloads and settings, not a general speedup or global synthesis optimum. The benchmark drivers, exact build metadata, raw measurements, and PNG/PDF plots remain in the local experiment archive.

Codex assisted with implementation, review, validation, and this description.

Checklist

• The pull request only contains commits that are focused and relevant to this change.
• I have added appropriate tests that cover the new/changed functionality.
• I have updated the documentation to reflect these changes.
• I have added entries to the changelog for any noteworthy additions, changes, fixes, or removals.
• I have added migration instructions to the upgrade guide (if needed).
• The changes follow the project's style guidelines and introduce no new warnings.
• The changes are fully tested and pass the CI checks.
• I have reviewed my own code changes.

If PR contains AI-assisted content:

• Any agent that created, edited, or submitted GitHub content was explicitly authorized for that scope, as required by our AI Usage Guidelines.
• Every agent-authored or agent-edited public text body begins with the visible disclosure 🤖 *AI text below* 🤖 (titles are exempt).
• I have disclosed AI assistance in the PR description.
• I confirm that I have personally reviewed and understood all AI-generated content, and accept full responsibility for it.

[codecov]

Codecov Report

❌ Patch coverage is 97.11538% with 3 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
...QCO/Transforms/NativeSynthesis/TargetSynthesis.cpp	97.0%	3 Missing ⚠️

📢 Thoughts on this report? Let us know!