Compilation passes¶
Download this notebook - compilation_example.ipynb
There are numerous ways to optimize circuits in pytket
. In this notebook we will introduce the basics of compilation passes and how to combine and apply them.
We assume familiarity with the pytket
Circuit
class. The objective is to transform one Circuit
into another, equivalent, Circuit
, that:
satisfies the connectivity constraints of a given architecture;
satisfies some further user-defined constraints (such as restricted gate sets);
minimizes some cost function (such as CX count).
Passes¶
The basic mechanism of compilation is the ‘pass’, which is a transform that can be applied to a circuit. There is an extensive library of passes in pytket
, and several standard ways in which they can be combined to form new passes. For example:
from pytket.passes import DecomposeMultiQubitsCX
pass1 = DecomposeMultiQubitsCX()
This pass converts all multi-qubit gates into CX and single-qubit gates. So let’s create a circuit containing some non-CX multi-qubit gates:
from pytket.circuit import Circuit
circ = Circuit(3)
circ.CRz(0.5, 0, 1)
circ.T(2)
circ.CSWAP(2, 0, 1)
[CRz(0.5) q[0], q[1]; T q[2]; CSWAP q[2], q[0], q[1]; ]
In order to apply a pass to a circuit, we must first create a CompilationUnit
from it. We can think of this as a ‘bridge’ between the circuit and the pass. The CompilationUnit
is constructed from the circuit; the pass is applied to the CompilationUnit
; and the transformed circuit is extracted from the CompilationUnit
:
from pytket.predicates import CompilationUnit
cu = CompilationUnit(circ)
pass1.apply(cu)
circ1 = cu.circuit
Let’s have a look at the result of the transformation:
print(circ1.get_commands())
[Rz(0.25) q[1];, T q[2];, CX q[0], q[1];, Rz(3.75) q[1];, CX q[0], q[1];, CX q[1], q[0];, H q[1];, CX q[0], q[1];, Tdg q[1];, CX q[2], q[1];, T q[1];, CX q[0], q[1];, T q[0];, Tdg q[1];, CX q[2], q[1];, CX q[2], q[0];, T q[1];, Tdg q[0];, H q[1];, T q[2];, CX q[2], q[0];, CX q[1], q[0];]
Predicates¶
Every CompilationUnit
has associated with it a set of ‘predicates’, which describe target properties that can be checked against the circuit. There are many types of predicates available in pytket
. For example, the GateSetPredicate
checks whether all gates in a circuit belong to a particular set:
from pytket.predicates import GateSetPredicate
from pytket.circuit import OpType
pred1 = GateSetPredicate({OpType.Rz, OpType.T, OpType.Tdg, OpType.H, OpType.CX})
When we construct a CompilationUnit
, we may pass a list of target predicates as well as the circuit:
cu = CompilationUnit(circ, [pred1])
To check whether the circuit associated to a CompilationUnit
satisfies its target predicates, we can call the check_all_predicates()
method:
cu.check_all_predicates()
False
pass1.apply(cu)
cu.check_all_predicates()
True
We can also directly check whether a given circuit satisfies a given predicate, using the predicate’s verify()
method:
pred1.verify(circ1)
True
In-place compilation¶
The example above produced a new circuit, leaving the original circuit untouched. It is also possible to apply a pass to a circuit in-place:
DecomposeMultiQubitsCX().apply(circ)
print(circ.get_commands())
[Rz(0.25) q[1];, T q[2];, CX q[0], q[1];, Rz(3.75) q[1];, CX q[0], q[1];, CX q[1], q[0];, H q[1];, CX q[0], q[1];, Tdg q[1];, CX q[2], q[1];, T q[1];, CX q[0], q[1];, T q[0];, Tdg q[1];, CX q[2], q[1];, CX q[2], q[0];, T q[1];, Tdg q[0];, H q[1];, T q[2];, CX q[2], q[0];, CX q[1], q[0];]
Combining passes¶
There are various ways to combine the elementary passes into more complex ones.
To combine several passes in sequence, we use a SequencePass
:
from pytket.passes import SequencePass, OptimisePhaseGadgets
seqpass = SequencePass([DecomposeMultiQubitsCX(), OptimisePhaseGadgets()])
This pass will apply the two transforms in succession:
cu = CompilationUnit(circ)
seqpass.apply(cu)
circ1 = cu.circuit
print(circ1.get_commands())
[TK1(0, 0, 0.25) q[1];, TK1(0, 0, 0.5) q[2];, CX q[1], q[0];, TK1(0.5, 0.5, 0.5) q[1];, CX q[0], q[1];, TK1(0, 0, 3.75) q[1];, CX q[2], q[1];, TK1(0, 0, 0.25) q[1];, CX q[0], q[1];, TK1(0, 0, 3.75) q[1];, CX q[2], q[1];, CX q[2], q[0];, TK1(0.5, 0.5, 0.75) q[1];, TK1(0, 0, 3.75) q[0];, CX q[2], q[0];, CX q[1], q[0];]
The apply()
method for an elementary pass returns a boolean indicating whether or not the pass had any effect on the circuit. For a SequencePass
, the return value indicates whether any of the constituent passes had some effect.
A RepeatPass
repeatedly calls apply()
on a pass until it returns False
, indicating that there was no effect:
from pytket.passes import CommuteThroughMultis, RemoveRedundancies, RepeatPass
seqpass = SequencePass([CommuteThroughMultis(), RemoveRedundancies()])
reppass = RepeatPass(seqpass)
This pass will repeatedly apply CommuteThroughMultis
(which commutes single-qubit operations through multi-qubit operations where possible towards the start of the circuit) and RemoveRedundancies
(which cancels inverse pairs, merges coaxial rotations and removes redundant gates before measurement) until neither pass has any effect on the circuit.
Let’s use pytket
’s built-in visualizer to see the effect on a circuit:
from pytket.circuit.display import render_circuit_jupyter
circ = Circuit(3)
circ.X(0).Y(1).CX(0, 1).Z(0).Rx(1.3, 1).CX(0, 1).Rz(0.4, 0).Ry(0.53, 0).H(1).H(2).Rx(
1.5, 2
).Rx(0.5, 2).H(2)
[X q[0]; Y q[1]; H q[2]; CX q[0], q[1]; Rx(1.5) q[2]; Z q[0]; Rx(1.3) q[1]; Rx(0.5) q[2]; CX q[0], q[1]; H q[2]; Rz(0.4) q[0]; H q[1]; Ry(0.53) q[0]; ]
render_circuit_jupyter(circ)
cu = CompilationUnit(circ)
reppass.apply(cu)
circ1 = cu.circuit
render_circuit_jupyter(circ1)
If we want to repeat a pass until the circuit satisfies some desired property, we first define a boolean function to test for that property, and then pass this function to the constructor of a RepeatUntilSatisfied
pass:
from pytket.passes import RepeatUntilSatisfiedPass
def no_CX(circ):
return circ.n_gates_of_type(OpType.CX) == 0
circ = (
Circuit(2)
.CX(0, 1)
.X(1)
.CX(0, 1)
.X(1)
.CX(0, 1)
.X(1)
.CX(0, 1)
.Z(1)
.CX(1, 0)
.Z(1)
.CX(1, 0)
)
custom_pass = RepeatUntilSatisfiedPass(seqpass, no_CX)
cu = CompilationUnit(circ)
custom_pass.apply(cu)
circ1 = cu.circuit
render_circuit_jupyter(circ1)
The RepeatWithMetricPass
provides another way of generating more sophisticated passes. This is defined in terms of a cost function and another pass type; the pass is applied repeatedly until the cost function stops decreasing.
For example, suppose we wish to associate a cost to each gate in out circuit, with \(n\)-qubit gates having a cost of \(n^2\):
def cost(circ):
return sum(pow(len(x.args), 2) for x in circ)
Let’s construct a new circuit:
circ = Circuit(2)
circ.CX(0, 1).X(1).Y(0).CX(0, 1).X(1).Z(0).CX(0, 1).X(1).Y(0).CX(0, 1).Z(1).CX(1, 0).Z(
1
).X(0).CX(1, 0)
[CX q[0], q[1]; Y q[0]; X q[1]; CX q[0], q[1]; Z q[0]; X q[1]; CX q[0], q[1]; Y q[0]; X q[1]; CX q[0], q[1]; Z q[1]; CX q[1], q[0]; X q[0]; Z q[1]; CX q[1], q[0]; ]
We will repeatedly apply CommuteThroughMultis
, DecomposeMultiQubitsCX
and RemoveRedundancies
until the cost
function stops decreasing:
from pytket.passes import RepeatWithMetricPass
pass1 = SequencePass(
[CommuteThroughMultis(), DecomposeMultiQubitsCX(), RemoveRedundancies()]
)
pass2 = RepeatWithMetricPass(pass1, cost)
cu = CompilationUnit(circ)
pass2.apply(cu)
print(cu.circuit.get_commands())
[X q[1];, CX q[0], q[1];, Y q[0];, Z q[0];, Y q[0];, CX q[0], q[1];, X q[0];]
Targeting architectures¶
If we are given a target architecture, we can generate passes tailored to it.
In pytket
an architecture is defined by a connectivity graph, i.e. a list of pairs of qubits capable of executing two-qubit operations. For example, we can represent a 5-qubit linear architecture, with qubits labelled n[i]
, as follows:
from pytket.architecture import Architecture
from pytket.circuit import Node
n = [Node("n", i) for i in range(5)]
arc = Architecture([[n[0], n[1]], [n[1], n[2]], [n[2], n[3]], [n[3], n[4]]])
Suppose we have a circuit that we wish to run on this architecture:
circ = Circuit(5)
circ.CX(0, 1)
circ.H(0)
circ.Z(1)
circ.CX(0, 3)
circ.Rx(1.5, 3)
circ.CX(2, 4)
circ.X(2)
circ.CX(1, 4)
circ.CX(0, 4)
[CX q[0], q[1]; CX q[2], q[4]; H q[0]; Z q[1]; X q[2]; CX q[0], q[3]; CX q[1], q[4]; CX q[0], q[4]; Rx(1.5) q[3]; ]
render_circuit_jupyter(circ)
A mapping pass lets us rewrite this circuit for our architecture:
from pytket.passes import DefaultMappingPass
mapper = DefaultMappingPass(arc)
cu = CompilationUnit(circ)
mapper.apply(cu)
circ1 = cu.circuit
render_circuit_jupyter(circ1)
If we want to decompose all SWAP and BRIDGE gates to CX gates in the final circuit, we can use another pass:
from pytket.passes import DecomposeSwapsToCXs
pass1 = DecomposeSwapsToCXs(arc)
pass1.apply(cu)
circ2 = cu.circuit
render_circuit_jupyter(circ2)
Note that the pass we just ran also performed some clean-up: the SWAP gate was decomposed into three CX gates, one of which was cancelled by a preceding CX gate; the cancelling gates were removed from the circuit.
Every compilation pass has associated sets of preconditions and postconditions on the circuit. If all preconditions are satisfied before the pass, all postconditions are guaranteed to be satisfied afterwards. When we apply a pass to a circuit, we can optionally pass SafetyMode.Audit
as the second parameter; this will tell the pass to check all preconditions explicitly. By default, there is only limited checking of preconditions and pytket
relies on the programmer assuring these.
For example, the NoClassicalControl
predicate is a precondition of the PauliSimp
pass. Let’s add a classically controlled gate to our circuit:
from pytket.passes import PauliSimp, SafetyMode
from pytket.circuit import Qubit, Bit
q = [Qubit("q", i) for i in range(5)]
c = Bit("c")
circ.add_bit(c)
circ.Measure(q[3], c)
circ.CY(q[0], q[1], condition_bits=[c], condition_value=1)
cu = CompilationUnit(circ)
try:
PauliSimp().apply(cu, safety_mode=SafetyMode.Audit)
except RuntimeError as e:
print("Error:", str(e))
Error: Predicate requirements are not satisfied: NoMidMeasurePredicate
The preconditions and postconditions of all the elementary predicates are documented in their string representations:
PauliSimp()
***PassType: SequencePass***
Preconditions:
GateSetPredicate:{ PhasedX Measure PhaseGadget ZZMax Tdg T Ry Rx Z X PauliExpBox Rz Y S Sdg V Vdg SWAP H YYPhase CY XXPhase CX ZZPhase CZ }
NoMidMeasurePredicate
NoClassicalControlPredicate
Specific Postconditions:
Generic Postconditions:
GateSetPredicate Clear
NoWireSwapsPredicate Clear
ConnectivityPredicate Clear
Default Postcondition: Preserve
Backends and default passes¶
A pytket
Backend
may have a default compilation pass, which will guarantee that the circuit can run on it. This is given by the default_compilation_pass
property. For example, the default pass for Qiskit’s AerBackend
just converts all gates to U1, U2, U3 and CX:
from pytket.extensions.qiskit import AerBackend
b = AerBackend()
b.default_compilation_pass
<bound method _AerBaseBackend.default_compilation_pass of <pytket.extensions.qiskit.backends.aer.AerBackend object at 0x7f9e43cfb760>>
To compile a circuit using the default pass of a Backend
we can simply use the get_compiled_circuit()
method:
circ = Circuit(2).X(0).Y(1).CRz(0.5, 1, 0)
circ1 = b.get_compiled_circuit(circ)
render_circuit_jupyter(circ1)
Every Backend
will have a certain set of requirements that must be met by any circuit in order to run. These are exposed via the required_predicates
property:
b.required_predicates
[NoSymbolsPredicate,
GateSetPredicate:{ Unitary2qBox CnZ ZZPhase Reset Measure noop Conditional CSWAP SWAP CCX StatePreparationBox CU3 CU1 CRy CRx CSX CZ CY Unitary3qBox CX Unitary1qBox TK1 CnX U1 U2 U3 Rz Ry Rx H SXdg SX YYPhase Tdg XXPhase T Sdg S PhasedX Y X ECR Z RangePredicate CRz Barrier }]
We can test whether a given circuit satisfies these requirements using the valid_circuit()
method:
b.valid_circuit(circ)
True
b.valid_circuit(circ1)
True