Switch default routing/layout method to sabre for opt level 3

This commit switches the default routing and layout method for
optimization level 3 to use SabreSwap and Sabre layout passes. The
quality of results is typically better with sabre than the default
stochastic swap and dense layout we're using now. For optimization
level 3 where we try to produce the best quality result and runtime
is of secondary concern using sabre makes the most sense (especially
after #7036 which improves the runtime performance). Also for
optimization level 3 currently sabre is typically faster because we
increase the number of trials for stochastic swap (which is generally
significantly faster) which slows it down as it's doing more work.
This should improve the quality and speed of the results in general
when running with optimization level 3. In the future we can do more
work to improve the runtime performance of the sabre passes and
hopefully make it fast enough to use for all the optimization levels
which have more constraints on the run time performance than level 3.

Related to #7112 and #7200

qiskit/transpiler/preset_passmanagers/level3.py

@@ -96,8 +96,8 @@ def level_3_pass_manager(pass_manager_config: PassManagerConfig) -> PassManager:
     inst_map = pass_manager_config.inst_map
     coupling_map = pass_manager_config.coupling_map
     initial_layout = pass_manager_config.initial_layout
-    layout_method = pass_manager_config.layout_method or "dense"
-    routing_method = pass_manager_config.routing_method or "stochastic"
+    layout_method = pass_manager_config.layout_method or "sabre"
+    routing_method = pass_manager_config.routing_method or "sabre"
     translation_method = pass_manager_config.translation_method or "translator"
     scheduling_method = pass_manager_config.scheduling_method
     instruction_durations = pass_manager_config.instruction_durations

releasenotes/notes/sabre-opt-lvl3-default-550924470d683112.yaml

@@ -0,0 +1,18 @@
+---
+upgrade:
+  - |
+    The default routing pass and layout pass for optimization level 3 has
+    changed to use :class:`~qiskit.transpiler.passes.SabreSwap` and
+    :class:`~qiskit.transpiler.passes.SabreLayout` respectively. This
+    was done to improve the quality of the output result as using the sabre
+    passes produces better results than using
+    :class:`~qiskit.transpiler.passes.StochasticSwap` and
+    :class:`~qiskit.transpiler.passes.DenseLayout` which were used as the
+    defaults in prior releases. This change will improve the quality of the
+    results when running :func:`~qiskit.compiler.transpile` or
+    :func:`~qiskit.execute_function.execute` functions with the
+    ``optimization_level`` kwarg set to ``3``. While this is generally an
+    improvement if you need to retain the previous behavior for any reason
+    you can do this by explicitly setting the ``routing_method="stochastic"``
+    and ``layout_method="dense"`` when calling
+    :func:`~qiskit.compiler.transpile` with ``optimization_level=3``.

test/python/transpiler/test_preset_passmanagers.py

@@ -674,10 +674,33 @@ def test_layout_tokyo_fully_connected_cx(self, level):
             19: ancilla[14],
         }
 
+        sabre_layout = {
+            0: ancilla[0],
+            1: ancilla[1],
+            2: ancilla[2],
+            3: ancilla[3],
+            4: ancilla[4],
+            5: ancilla[5],
+            6: ancilla[6],
+            7: qr[4],
+            8: qr[1],
+            9: ancilla[7],
+            10: ancilla[8],
+            11: ancilla[9],
+            12: qr[2],
+            13: qr[0],
+            14: ancilla[10],
+            15: ancilla[11],
+            16: ancilla[12],
+            17: ancilla[13],
+            18: ancilla[14],
+            19: qr[3],
+        }
+
         expected_layout_level0 = trivial_layout
         expected_layout_level1 = dense_layout
         expected_layout_level2 = dense_layout
-        expected_layout_level3 = dense_layout
+        expected_layout_level3 = sabre_layout
 
         expected_layouts = [
             expected_layout_level0,