'Peephole' optimization - or: collecting and optimizing two-qubit blocks - before routing (#12727)

* init

* up

* up

* Update builtin_plugins.py

* Update builtin_plugins.py

* reno

* Update builtin_plugins.py

* Update builtin_plugins.py

* Update peephole-before-routing-c3d184b740bb7a8b.yaml

* neko check

* check neko

* Update builtin_plugins.py

* test neko

* Update builtin_plugins.py

* Update builtin_plugins.py

* Update builtin_plugins.py

* lint

* tests and format

* remove FakeTorino test

* Update peephole-before-routing-c3d184b740bb7a8b.yaml

* Apply suggestions from code review

Co-authored-by: Matthew Treinish <mtreinish@kortar.org>

* comments from code review

* fix precision

* up

* up

* update

* up

* .

* cyclic import

* cycl import

* cyl import

* .

* circular import

* .

* lint

* Include new pass in docs

* Fix Split2QUnitaries dag manipulation

This commit fixes the dag handling to do the 1q unitary insertion.
Previously the dag manipulation was being done manually using the
insert_node_on_in_edges() rustworkx method. However as the original node
had 2 incoming edges for each qubit this caused the dag after running
the pass to become corrupted. Each of the new 1q unitary nodes would end
up with 2 incident edges and they would be in a sequence. This would result
in later passes not being able to correctly understand the state of the
circuit correctly. This was causing the unit tests to fail. This commit
fixes this by just using `substitute_node_with_dag()` to handle the
node substition, while doing it manually to avoid the overhead of
checking is probably possible, the case where a unitary is the product
of two 1q gates is not very common so optimizing it isn't super
critical.

* Update releasenotes/notes/peephole-before-routing-c3d184b740bb7a8b.yaml

* stricter check for doing split2q

* Update qiskit/transpiler/preset_passmanagers/builtin_plugins.py

Co-authored-by: Matthew Treinish <mtreinish@kortar.org>

* code review

* Update qiskit/transpiler/passes/optimization/split_2q_unitaries.py

Co-authored-by: Matthew Treinish <mtreinish@kortar.org>

* new tests

* typo

* lint

* lint

---------

Co-authored-by: Matthew Treinish <mtreinish@kortar.org>
(cherry picked from commit 1214d51)

Author: sbrandhsn

qiskit/transpiler/passes/__init__.py

@@ -91,6 +91,7 @@
    ElidePermutations
    NormalizeRXAngle
    OptimizeAnnotated
+   Split2QUnitaries
 
 Calibration
 =============
@@ -244,6 +245,7 @@
 from .optimization import ElidePermutations
 from .optimization import NormalizeRXAngle
 from .optimization import OptimizeAnnotated
+from .optimization import Split2QUnitaries
 
 # circuit analysis
 from .analysis import ResourceEstimation

qiskit/transpiler/passes/optimization/__init__.py

@@ -38,3 +38,4 @@
 from .elide_permutations import ElidePermutations
 from .normalize_rx_angle import NormalizeRXAngle
 from .optimize_annotated import OptimizeAnnotated
+from .split_2q_unitaries import Split2QUnitaries

qiskit/transpiler/passes/optimization/split_2q_unitaries.py

@@ -0,0 +1,83 @@
+# This code is part of Qiskit.
+#
+# (C) Copyright IBM 2017, 2024.
+#
+# This code is licensed under the Apache License, Version 2.0. You may
+# obtain a copy of this license in the LICENSE.txt file in the root directory
+# of this source tree or at http://www.apache.org/licenses/LICENSE-2.0.
+#
+# Any modifications or derivative works of this code must retain this
+# copyright notice, and modified files need to carry a notice indicating
+# that they have been altered from the originals.
+"""Splits each two-qubit gate in the `dag` into two single-qubit gates, if possible without error."""
+from typing import Optional
+
+from qiskit.transpiler.basepasses import TransformationPass
+from qiskit.circuit.quantumcircuitdata import CircuitInstruction
+from qiskit.dagcircuit.dagcircuit import DAGCircuit, DAGOpNode
+from qiskit.circuit.library.generalized_gates import UnitaryGate
+from qiskit.synthesis.two_qubit.two_qubit_decompose import TwoQubitWeylDecomposition
+
+
+class Split2QUnitaries(TransformationPass):
+    """Attempt to splits two-qubit gates in a :class:`.DAGCircuit` into two single-qubit gates
+
+    This pass will analyze all the two qubit gates in the circuit and analyze the gate's unitary
+    matrix to determine if the gate is actually a product of 2 single qubit gates. In these
+    cases the 2q gate can be simplified into two single qubit gates and this pass will
+    perform this optimization and will replace the two qubit gate with two single qubit
+    :class:`.UnitaryGate`.
+    """
+
+    def __init__(self, fidelity: Optional[float] = 1.0 - 1e-16):
+        """Split2QUnitaries initializer.
+
+        Args:
+            fidelity (float): Allowed tolerance for splitting two-qubit unitaries and gate decompositions
+        """
+        super().__init__()
+        self.requested_fidelity = fidelity
+
+    def run(self, dag: DAGCircuit):
+        """Run the Split2QUnitaries pass on `dag`."""
+        for node in dag.topological_op_nodes():
+            # skip operations without two-qubits and for which we can not determine a potential 1q split
+            if (
+                len(node.cargs) > 0
+                or len(node.qargs) != 2
+                or node.matrix is None
+                or node.is_parameterized()
+            ):
+                continue
+
+            decomp = TwoQubitWeylDecomposition(node.op, fidelity=self.requested_fidelity)
+            if (
+                decomp._inner_decomposition.specialization
+                == TwoQubitWeylDecomposition._specializations.IdEquiv
+            ):
+                new_dag = DAGCircuit()
+                new_dag.add_qubits(node.qargs)
+
+                ur = decomp.K1r
+                ur_node = DAGOpNode.from_instruction(
+                    CircuitInstruction(UnitaryGate(ur), qubits=(node.qargs[0],)), dag=new_dag
+                )
+
+                ul = decomp.K1l
+                ul_node = DAGOpNode.from_instruction(
+                    CircuitInstruction(UnitaryGate(ul), qubits=(node.qargs[1],)), dag=new_dag
+                )
+                new_dag._apply_op_node_back(ur_node)
+                new_dag._apply_op_node_back(ul_node)
+                new_dag.global_phase = decomp.global_phase
+                dag.substitute_node_with_dag(node, new_dag)
+            elif (
+                decomp._inner_decomposition.specialization
+                == TwoQubitWeylDecomposition._specializations.SWAPEquiv
+            ):
+                # TODO maybe also look into swap-gate-like gates? Things to consider:
+                #   * As the qubit mapping may change, we'll always need to build a new dag in this pass
+                #   * There may not be many swap-gate-like gates in an arbitrary input circuit
+                #   * Removing swap gates from a user-routed input circuit here is unexpected
+                pass
+        return dag

qiskit/transpiler/preset_passmanagers/builtin_plugins.py

@@ -14,6 +14,8 @@
 
 import os
 
+from qiskit.circuit import Instruction
+from qiskit.transpiler.passes.optimization.split_2q_unitaries import Split2QUnitaries
 from qiskit.transpiler.passmanager import PassManager
 from qiskit.transpiler.exceptions import TranspilerError
 from qiskit.transpiler.passes import BasicSwap
@@ -64,12 +66,23 @@
     CYGate,
     SXGate,
     SXdgGate,
+    get_standard_gate_name_mapping,
 )
 from qiskit.utils.parallel import CPU_COUNT
 from qiskit import user_config
 
 CONFIG = user_config.get_config()
 
+_discrete_skipped_ops = {
+    "delay",
+    "reset",
+    "measure",
+    "switch_case",
+    "if_else",
+    "for_loop",
+    "while_loop",
+}
+
 
 class DefaultInitPassManager(PassManagerStagePlugin):
     """Plugin class for default init stage."""
@@ -160,6 +173,58 @@ def pass_manager(self, pass_manager_config, optimization_level=None) -> PassMana
                 )
             )
             init.append(CommutativeCancellation())
+            # skip peephole optimization before routing if target basis gate set is discrete,
+            # i.e. only consists of Cliffords that an user might want to keep
+            # use rz, sx, x, cx as basis, rely on physical optimziation to fix everything later one
+            stdgates = get_standard_gate_name_mapping()
+
+            def _is_one_op_non_discrete(ops):
+                """Checks if one operation in `ops` is not discrete, i.e. is parameterizable
+                Args:
+                    ops (List(Operation)): list of operations to check
+                Returns
+                    True if at least one operation in `ops` is not discrete, False otherwise
+                """
+                found_one_continuous_gate = False
+                for op in ops:
+                    if isinstance(op, str):
+                        if op in _discrete_skipped_ops:
+                            continue
+                        op = stdgates.get(op, None)
+
+                    if op is not None and op.name in _discrete_skipped_ops:
+                        continue
+
+                    if op is None or not isinstance(op, Instruction):
+                        return False
+
+                    if len(op.params) > 0:
+                        found_one_continuous_gate = True
+                return found_one_continuous_gate
+
+            target = pass_manager_config.target
+            basis = pass_manager_config.basis_gates
+            # consolidate gates before routing if the user did not specify a discrete basis gate, i.e.
+            # * no target or basis gate set has been specified
+            # * target has been specified, and we have one non-discrete gate in the target's spec
+            # * basis gates have been specified, and we have one non-discrete gate in that set
+            do_consolidate_blocks_init = target is None and basis is None
+            do_consolidate_blocks_init |= target is not None and _is_one_op_non_discrete(
+                target.operations
+            )
+            do_consolidate_blocks_init |= basis is not None and _is_one_op_non_discrete(basis)
+
+            if do_consolidate_blocks_init:
+                init.append(Collect2qBlocks())
+                init.append(ConsolidateBlocks())
+                # If approximation degree is None that indicates a request to approximate up to the
+                # error rates in the target. However, in the init stage we don't yet know the target
+                # qubits being used to figure out the fidelity so just use the default fidelity parameter
+                # in this case.
+                if pass_manager_config.approximation_degree is not None:
+                    init.append(Split2QUnitaries(pass_manager_config.approximation_degree))
+                else:
+                    init.append(Split2QUnitaries())
         else:
             raise TranspilerError(f"Invalid optimization level {optimization_level}")
         return init

releasenotes/notes/peephole-before-routing-c3d184b740bb7a8b.yaml

@@ -0,0 +1,20 @@
+---
+features_transpiler:
+  - |
+    Added a new pass :class:`.Split2QUnitaries` that iterates over all two-qubit gates or unitaries in a
+    circuit and replaces them with two single-qubit unitaries, if possible without introducing errors, i.e.
+    the two-qubit gate/unitary is actually a (kronecker) product of single-qubit unitaries.
+  - |
+    The passes :class:`.Collect2qBlocks`, :class:`.ConsolidateBlocks` and :class:`.Split2QUnitaries` have been
+    added to the ``init`` stage of the preset pass managers with optimization level 2 and optimization level 3.
+    The modification of the `init` stage should allow for a more efficient routing for quantum circuits that either:
+    
+      * contain two-qubit unitaries/gates that are actually a product of single-qubit gates
+      * contain multiple two-qubit gates in a continuous block of two-qubit gates.
+      
+    In the former case, the routing of the two-qubit gate can simply be skipped as no real interaction
+    between a pair of qubits occurs. In the latter case, the lookahead space of routing algorithms is not
+    'polluted' by superfluous two-qubit gates, i.e. for routing it is sufficient to only consider one single
+    two-qubit gate per continuous block of two-qubit gates. These passes are not run if the pass
+    managers target a :class:`.Target` that has a discrete basis gate set, i.e. all basis gates have are not
+    parameterized.

test/python/compiler/test_transpiler.py

@@ -84,6 +84,8 @@
 from qiskit.transpiler import CouplingMap, Layout, PassManager, TransformationPass
 from qiskit.transpiler.exceptions import TranspilerError, CircuitTooWideForTarget
 from qiskit.transpiler.passes import BarrierBeforeFinalMeasurements, GateDirection, VF2PostLayout
+from qiskit.transpiler.passes.optimization.split_2q_unitaries import Split2QUnitaries
+
 from qiskit.transpiler.passmanager_config import PassManagerConfig
 from qiskit.transpiler.preset_passmanagers import generate_preset_pass_manager, level_0_pass_manager
 from qiskit.transpiler.target import (
@@ -862,6 +864,42 @@ def test_do_not_run_gatedirection_with_symmetric_cm(self):
             transpile(circ, coupling_map=coupling_map, initial_layout=layout)
             self.assertFalse(mock_pass.called)
 
+    def tests_conditional_run_split_2q_unitaries(self):
+        """Tests running `Split2QUnitaries` when basis gate set is (non-) discrete"""
+        qc = QuantumCircuit(3)
+        qc.sx(0)
+        qc.t(0)
+        qc.cx(0, 1)
+        qc.cx(1, 2)
+
+        orig_pass = Split2QUnitaries()
+        with patch.object(Split2QUnitaries, "run", wraps=orig_pass.run) as mock_pass:
+            basis = ["t", "sx", "cx"]
+            backend = GenericBackendV2(3, basis_gates=basis)
+            transpile(qc, backend=backend)
+            transpile(qc, basis_gates=basis)
+            transpile(qc, target=backend.target)
+            self.assertFalse(mock_pass.called)
+
+        orig_pass = Split2QUnitaries()
+        with patch.object(Split2QUnitaries, "run", wraps=orig_pass.run) as mock_pass:
+            basis = ["rz", "sx", "cx"]
+            backend = GenericBackendV2(3, basis_gates=basis)
+            transpile(qc, backend=backend, optimization_level=2)
+            self.assertTrue(mock_pass.called)
+            mock_pass.called = False
+            transpile(qc, basis_gates=basis, optimization_level=2)
+            self.assertTrue(mock_pass.called)
+            mock_pass.called = False
+            transpile(qc, target=backend.target, optimization_level=2)
+            self.assertTrue(mock_pass.called)
+            mock_pass.called = False
+            transpile(qc, backend=backend, optimization_level=3)
+            self.assertTrue(mock_pass.called)
+            mock_pass.called = False
+            transpile(qc, basis_gates=basis, optimization_level=3)
+            self.assertTrue(mock_pass.called)
+
     def test_optimize_to_nothing(self):
         """Optimize gates up to fixed point in the default pipeline
         See https://github.com/Qiskit/qiskit-terra/issues/2035

test/python/transpiler/test_preset_passmanagers.py

@@ -14,6 +14,7 @@
 
 import unittest
 
+
 from test import combine
 from ddt import ddt, data
 
@@ -279,7 +280,7 @@ def counting_callback_func(pass_, dag, time, property_set, count):
             callback=counting_callback_func,
             translation_method="synthesis",
         )
-        self.assertEqual(gates_in_basis_true_count + 1, collect_2q_blocks_count)
+        self.assertEqual(gates_in_basis_true_count + 2, collect_2q_blocks_count)
 
 
 @ddt