Leverage retworkx for apply_operation* inner loop #5267
Conversation
This commit migrates the inner loop for apply_operation_back and apply_operation_front to leverage retworkx's insert_node_between_multiple PyDiGraph method. This retworkx method performs essentially the same loop internally but executes faster. Depends on Qiskit/rustworkx#181 being included in a release.
|
Just to show some initial numbers on this. It's showing a ~2X speed improvement for To this with this PR applied (and retworkx patched): The interesting things are after this PR the bottlenecks for apply operation back are a 3 way tie between calling I'm thinking that after this and #5272 we might be able to speed up the other 2 bottlenecks by changing the |
|
I added the |
|
retworkx 0.6.0 was released this morning, so this is no longer blocked. |
In sabre_swap the comprehension in _score_heuristic() with the basic heuristic starts to become a bottleneck as the size of the coupling map increases. This probably becomes the top bottleneck in a transpilation after Qiskit#5316, Qiskit#5183, Qiskit#5294, Qiskit#5267, and Qiskit#5272 are merged. This commit is an attempt to try and mitigate that a bit by using numpy native operations instead of a python comprehension in sum(). If this is insufficient we'll likely have to either avoid doing a sum like this or drop down to a lower level with cython or rust and operate on the array directly to remove this bottleneck.
* Use numpy sum instead of comprehension in _score_heuristic In sabre_swap the comprehension in _score_heuristic() with the basic heuristic starts to become a bottleneck as the size of the coupling map increases. This probably becomes the top bottleneck in a transpilation after #5316, #5183, #5294, #5267, and #5272 are merged. This commit is an attempt to try and mitigate that a bit by using numpy native operations instead of a python comprehension in sum(). If this is insufficient we'll likely have to either avoid doing a sum like this or drop down to a lower level with cython or rust and operate on the array directly to remove this bottleneck. * Make distance matrix a coupling map property * Fix tests * Fix lint Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
Summary
This commit migrates the inner loop for apply_operation_back and
apply_operation_front to leverage retworkx's
insert_node_between_multiple PyDiGraph method. This retworkx method
performs essentially the same loop internally but executes faster.
Details and comments
Depends on Qiskit/rustworkx#181 being included in
a release.