Replacing Graphix transpiler with graphix-jcz-transpiler

[emlynsg]

This PR takes the causal flow-based transpilation method detailed in the graphix-jcz-transpiler and adds it to Graphix, replacing the previous transpilation approach into one that is more principled and consistent with the literature [1].

Major changes include:

• Full rework of the transpiler.py module to implement $J,, CZ$ gate decomposition and use the causal-flow-based approach to construct patterns
• Addition of a $J(\alpha)$ gate to circuits, with equivalent additions to random circuits, ops and management to codebase components related to OpenQASM
• Updates to many tests to account for the new transpiler generating BlochMeasurement commands which need to be inferred as Pauli (usually for preprocessing)

[1] The Measurement Calculus
Vincent Danos, Elham Kashefi, Prakash Panangaden

[thierry-martinez]

Veriphix tests are likely being killed by the OOM killer. When running the Veriphix test suite locally, I observed that it hangs on the very first test of test_client.py. This test runs a delegated random circuit that only standardizes the pattern, without using Pauli presimulation or space minimization. There is a difficulty that I don't know how to address here: the new transpiler cannot match the performance of the original one without Pauli presimulation, but we do not yet know how to implement Pauli presimulation without setting the input.

[emlynsg]

Veriphix tests are likely being killed by the OOM killer. When running the Veriphix test suite locally, I observed that it hangs on the very first test of test_client.py. This test runs a delegated random circuit that only standardizes the pattern, without using Pauli presimulation or space minimization. There is a difficulty that I don't know how to address here: the new transpiler cannot match the performance of the original one without Pauli presimulation, but we do not yet know how to implement Pauli presimulation without setting the input.

I agree that seems to be the case. I can imagine it should work if either minimize_space or to_space_minimal_pattern are applied -- let's discuss tomorrow morning.

[codecov]

Codecov Report

❌ Patch coverage is 94.44444% with 14 lines in your changes missing coverage. Please review.
✅ Project coverage is 88.73%. Comparing base (8bf1ebe) to head (e067d7a).
⚠️ Report is 2 commits behind head on master.

Files with missing lines	Patch %	Lines
graphix/transpiler.py	95.96%	9 Missing ⚠️
graphix/qasm3_exporter.py	72.72%	3 Missing ⚠️
graphix/sim/density_matrix.py	50.00%	1 Missing ⚠️
graphix/sim/statevec.py	50.00%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #484      +/-   ##
==========================================
- Coverage   88.85%   88.73%   -0.12%     
==========================================
  Files          49       49              
  Lines        7135     7202      +67     
==========================================
+ Hits         6340     6391      +51     
- Misses        795      811      +16     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[thierry-martinez]

LGTM! Thank you very much for this big change.

[matulni]

Thanks @emlynsg for this nice work! Unfortunately I don't know the details about the JCZ transpiler so I only made comments about the code.

On a general note, your decompose_... functions lend themselves to return iterators (check the old _transpile_rzz), which are evaluated lazily and avoid allocating lists in the intermediate steps (you may also want to check map). I didn't check, so I won't make any claims on its performance (my guess is that it's not significant for the circuit sizes that we deal with). I will not bother you with it, we can leave it as is 😬

[matulni]

Thank you, very nice work. We should be able recycle some of your ideas in TranspileSwapResult in a more general API for Circuit !

I made some minor comments (and there remain a few comments from a previous review I made to resolve).

[matulni]

I'm not sure if we need this check, we are already checking in instruction_to_qasm3.
By the same logic, we should also check that there are not non-Z measurements (which we do at instruction_to_qasm3).

If we decide to keep it:

Suggested change

	if any(instr.kind == InstructionKind.J for instr in circuit.instruction):
	elif any(instr.kind == InstructionKind.J for instr in circuit.instruction):

[matulni]

We are transpiling j and measurements in circuit_to_qasm_lines, so I think there's no need to transpile twice.

Suggested change

	if transpile:
	circuit = circuit.transpile_j_to_rzh().transpile_measurements_to_z_axis()

[matulni]

Suggested change

	raise ValueError("J gates must be decomposed before QASM3 export using `Circuit.transpile_j_to_rzh`.")
	raise ValueError("J gates must be decomposed before QASM3 export using `Circuit.transpile_j_to_rzh`, or setting `transpile=True`.")

[matulni]

Done below

Suggested change

p_test = p_test.infer_pauli_measurements()

[matulni]

Maybe it's an overkill, but do you think it would be worth to introduce a property classical_outputs to narrow the type _CO depending on the type of _R?

[matulni]

I think the logic in this method deserves to be a method of Statevec (or even DenseState).
The reason is twofold:

• It relies on the internal representation of statevec.psi (in fact, it will fail with the new jit backend).
• We could use it in DenseStateBackend.sort_qubits instead of repeated calls to swap (outside of this PR).