SABRE: get transpiler statistics

[geexie]

I'm toying with SABRE transpiler and wondering, Is it possible to get statistics about SWAPS added and other transpiler debug statistics, like number of remappings, etc? I can deduce some information indirectly comparing circuit summaries, but it might be useful to let SABRE to be a bit more verbose about what is going on. Probably, store circuit after each optimisation pass...

[alecandido]

Agreed, it could be useful.

But no, with just the present code, you can not. I'm sorry 😞

[alecandido]

Btw, the relevant code is here

qibo/src/qibo/transpiler/router.py

Lines 630 to 668 in de4cc77

	def __call__(self, circuit: Circuit):
	"""Route the circuit.
	Args:
	circuit (:class:`qibo.models.circuit.Circuit`): Circuit to be routed.
	Returns:
	(:class:`qibo.models.circuit.Circuit`, dict): Routed circuit and final {logical: physical} qubit mapping.
	"""
	assert_placement(circuit, self.connectivity)
	self._preprocessing(circuit=circuit)
	longest_path = np.max(self._dist_matrix)
	while self._dag.number_of_nodes() != 0:
	execute_block_list = self._check_execution()
	if execute_block_list is not None:
	self._execute_blocks(execute_block_list)
	else:
	self._find_new_mapping()
	# If the number of added swaps is too high, the algorithm is stuck.
	# Reset the circuit to the last saved state and make the nearest gate executable by manually adding SWAPs.
	if (
	len(self._temp_added_swaps) > self.swap_threshold * longest_path
	): # threshold is arbitrary
	while self._temp_added_swaps:
	self._temp_added_swaps.pop()
	self.circuit_map.undo()
	self._temp_added_swaps = []
	self._shortest_path_routing()
	circuit_kwargs = circuit.init_kwargs
	routed_circuit = self.circuit_map.routed_circuit(circuit_kwargs=circuit_kwargs)
	if self._final_measurements is not None:
	routed_circuit = self._append_final_measurements(
	routed_circuit=routed_circuit
	)
	return routed_circuit, self.circuit_map.final_layout()

Which is the SABRE implementation. And it is only the routed circuit and the global permutation (to map final measurements).

So, without changing the code (or hooking in somehow, e.g. with a debugger), there is no more information you can get.

[geexie]

@alecandido, thank you for clarification. As for global permutation map, do I need to use it to reinterpret the results? I skip it for now, but based on you comment I was just lucky. What is the proper way to use it?

[alecandido]

Sorry for the late reply @geexie.

As said, it passed some time, and I was not implementing it myself. I should read the code again to make sure about it.

However, I believe that the general idea assumes the generation of a final state vector, or to get shots only from measurements at the end of the circuit.
In this sense, if you place your circuit or introduce swaps to route your gates, the idea is that you keep applying permutations. And you need to undo those to reinterpret results.

E.g. assume you want to execute the following circuit

0: ─o───o───
1: ─X─o─|─M─
2: ───X─X───

but the (0, 2) pair is not connected. Then you will route it to

0: ─o───x───M─
1: ─X─o─x─o───
2: ───X───X───

But now, in the final configuration, you also swapped the measurement gate. Which is applying onto the physical qubit 0, but virtually it is acting on 1.
That's why you need to take into account the permutation.

However, that is still a poor man solution, since it is disregarding the swap of mid-circuit measurements. So, we will need something more anyhow...

Regarding whether you should use it explicitly or not, I would have hoped that this is done automatically by the transpiler. But it is actually fully ignored if you use it automatically...

qibo/src/qibo/models/circuit.py

Line 1084 in 534b95f

transpiled_circuit, _ = transpiler(self) # pylint: disable=E1102

We definitely need to dedicate some time to it...