The Quil compiler¶

Expectations for program contents¶

Real QPUs have a much more limited natural gate sets than the standard gate set offered by pyQuil because gates are required to act on physically available hardware (for single-qubit gates, this means acting only on live qubits, and for qubit-pair gates, this means acting on neighboring qubits). However, as a programmer, it is often (though not always) desirable to to be able to write programs which don’t take these details into account. This generally leads to more portable code if one isn’t tied to a specific set of gates or QPU architecture. To ameliorate these limitations, the Quil SDK contains an optimizing compiler that translates arbitrary Quil to native Quil and native Quil to executables suitable for Rigetti hardware.

Interacting with the compiler¶

After installing the SDK, the Quil Compiler, quilc is available on your local machine. You can initialize a local quilc server by typing quilc -S into your terminal.

A QuantumComputer object supplied by the function get_qc() comes equipped with a connection to your local quilc server. This can be accessed using the instance method .compile(), as in the following:

from pyquil.quil import Pragma, Program
from pyquil.api import get_qc
from pyquil.gates import CNOT, H

qc = get_qc("9q-square-qvm")

ep = qc.compile(Program(H(0), CNOT(0,1), CNOT(1,2)))

print(ep)

with output

RZ(pi/2) 0
RX(pi/2) 0
RZ(-pi/2) 1
RX(pi/2) 1
CZ 1 0
RX(-pi/2) 1
RZ(-pi/2) 2
RX(pi/2) 2
CZ 2 1
RZ(-pi/2) 0
RZ(-pi/2) 1
RX(-pi/2) 2
RZ(pi/2) 2

The compiler connection is also available directly via the property qc.compiler. The precise class of this object changes based on context (e.g., QPUCompiler, QVMCompiler), but it always conforms to the interface laid out by AbstractCompiler:

compiler.quil_to_native_quil(program, *, protoquil): This method converts a Quil program into native Quil, according to the ISA that the compiler is initialized with. The input parameter is specified as a Program object. The optional protoquil keyword argument instructs the compiler to restrict both its input and output to protoquil (Quil code that can be executed on a QPU). If the server is started with the -P option, or you specify protoquil=True the returned Program object will be equipped with a .metadata property that gives extraneous information about the compilation output (e.g., gate depth, as well as many others). This call blocks until Quil compilation finishes.
compiler.native_quil_to_executable(nq_program): This method converts a native Quil program, which is promised to consist only of native gates for a given ISA, into an executable suitable for submission to one of a QVM or a QPU. This call blocks until the executable is generated.

The instance method qc.compile described above is a combination of these two methods: first the incoming Quil is nativized, and then that is immediately turned into an executable. Accordingly, the previous example snippet is identical to the following:

from pyquil.quil import Pragma, Program
from pyquil.api import get_qc
from pyquil.gates import CNOT, H

qc = get_qc("9q-square-qvm")

p = Program(H(0), CNOT(0,1), CNOT(1,2))

np = qc.compiler.quil_to_native_quil(p, protoquil=True)
print(np)

ep = qc.compiler.native_quil_to_executable(np)
print(ep)

Timeouts¶

If your circuit is sufficiently complex the compiler may require more time than is permitted by default. To change this timeout, use the compiler_timeout option on get_qc:

qc = get_qc("2q-qvm", compiler_timeout=100) # 100 seconds

Legal compiler input¶

The QPU is not able to execute all possible Quil programs. At present, a Quil program qualifies for execution if has the following form:

The program may begin with a RESET instruction. (If provided, the QPU will actively reset the state of the quantum device to the ground state before program execution. If omitted, the QPU will wait for a relaxation period to pass before program execution instead.)
This is then followed by a block of native quantum gates. A gate is native if it is of the form RZ(θ) for any value θ, RX(k*π/2) for an integer k, or CZ q0 q1 for q0, q1 a pair of qubits participating in a qubit-qubit interaction. Some devices provide the XY(θ) q0 q1 two qubit gate.
This is then followed by a block of MEASURE instructions.

To instruct the compiler to produce Quil code that can be executed on a QPU, you can use the protoquil keyword in a call to compiler.quil_to_native_quil(program, protoquil=True) or qc.compile(program, protoquil=True).

Note

If your compiler server is started with the protoquil option -P (as is the case for your JupyterLab notebook’s compiler) then specifying protoquil=False will override the server and forcefully disable protoquil. Specifying protoquil=None defers to the server’s choice.

Compilation metadata¶

When your compiler is started with the protoquil option (-P), the compiler.quil_to_native_quil() method will return both the compiled program and a dictionary of statistics for the compiled program. This dictionary contains the keys

final_rewiring: see section below on rewirings.
gate_depth: the longest subsequence of compiled instructions where adjacent instructions share resources.
multiqubit_gate_depth: like gate_depth but restricted to multi-qubit gates.
gate_volume: total number of gates in the compiled program.
program_duration: program duration with parallel executation of gates (using hard-coded values of individual gate durations).
qpu_runtime_estimation: estimated runtime on a Rigetti QPU (in milliseconds). This is extrapolated from a single shot of a 16Q program with final measurements on all 16 qubits. If you are running a parametric program then you should estimate the total runtime as size of parameter space * estimated runtime of single shot. This should be treated only as an approximation.
program_fidelity: the estimated fidelity of the compiled program.
topological_swaps: the number of topological swaps incurred during compilation of the program.

For example, to inspect the qpu_runtime_estimation you might do the following:

from pyquil import get_qc, Program

# If you have a reserved QPU, use it here
# qc = get_qc("Aspen-X")
# Otherwise use a QVM
qc = get_qc("8q-qvm")

# Likely you will have a more complex program:
p = Program("RX(pi) 0")

native_p = qc.compiler.quil_to_native_quil(p, protoquil=True)

# The program will now have only native gates
print(native_p)
# And also metadata, with the above properties
print(native_p.native_quil_metadata.qpu_runtime_estimation)

Region-specific compiler features through PRAGMA commands¶

The Quil compiler can also be communicated with through PRAGMA commands embedded in the Quil program.

Preserved regions¶

The compiler can be circumvented in user-specified regions. The start of such a region is denoted by PRAGMA PRESERVE_BLOCK, and the end is denoted by PRAGMA END_PRESERVE_BLOCK. The Quil compiler promises not to modify any instructions contained in such a region.

Warning

If a preserved block is not legal QPU input, then it is not guaranteed to execute or it may produced unexpected results.

The following is an example of a program that prepares a Bell state on qubits 0 and 1, then performs a time delay to invite noisy system interaction before measuring the qubits. The time delay region is marked by PRAGMA PRESERVE_BLOCK and PRAGMA END_PRESERVE_BLOCK; without these delimiters, the compiler will remove the identity gates that serve to provide the time delay. However, the regions outside of the PRAGMA region will still be compiled, converting the Bell state preparation to the native gate set.

DECLARE ro BIT[2]

#   prepare a Bell state
H 0
CNOT 0 1

#   wait a while
PRAGMA PRESERVE_BLOCK
I 0
I 1
I 0
I 1
# ...
I 0
I 1
PRAGMA END_PRESERVE_BLOCK

#   and read out the results
MEASURE 0 ro[0]
MEASURE 1 ro[1]

Parallelizable regions¶

The compiler can sometimes arrange gate sequences more cleverly if the user gives it hints about sequences of gates that commute. A region containing commuting sequences is bookended by PRAGMA COMMUTING_BLOCKS and PRAGMA END_COMMUTING_BLOCKS; within such a region, a given commuting sequence is bookended by PRAGMA BLOCK and PRAGMA END_BLOCK.

Warning

Lying to the compiler about what blocks can commute can cause incorrect results.

The following snippet demonstrates this hinting syntax in a context typical of VQE-type algorithms: after a first stage of performing some state preparation on individual qubits, there is a second stage of “mixing operations” that both re-use qubit resources and mutually commute, followed by a final rotation and measurement. The following program is naturally laid out on a ring with vertices (read either clockwise or counterclockwise) as 0, 1, 2, 3. After scheduling the first round of preparation gates, the compiler will use the hinting to schedule the first and third blocks (which utilize qubit pairs 0-1 and 2-3) before the second and fourth blocks (which utilize qubit pairs 1-2 and 0-3), resulting in a reduction in circuit depth by one half. Without hinting, the compiler will instead execute the blocks in their written order.

DECLARE ro BIT[4]

# Stage one
H 0
H 1
H 2
H 3

# Stage two
PRAGMA COMMUTING_BLOCKS
PRAGMA BLOCK
CNOT 0 1
RZ(0.4) 1
CNOT 0 1
PRAGMA END_BLOCK
PRAGMA BLOCK
CNOT 1 2
RZ(0.6) 2
CNOT 1 2
PRAGMA END_BLOCK
PRAGMA BLOCK
CNOT 2 3
RZ(0.8) 3
CNOT 2 3
PRAGMA END_BLOCK
PRAGMA BLOCK
CNOT 0 3
RZ(0.9) 3
CNOT 0 3
PRAGMA END_BLOCK
PRAGMA END_COMMUTING_BLOCKS

# Stage three
H 0
H 1
H 2
H 3

MEASURE 0 ro[0]
MEASURE 1 ro[1]
MEASURE 2 ro[2]
MEASURE 3 ro[3]

Rewirings¶

When a Quil program contains multi-qubit instructions that do not name qubit-qubit links present on a target device, the compiler will rearrange the qubits so that execution becomes possible. In order to help the user understand what rearrangement may have been done, the compiler emits comments at various points in the raw Quil code (which is not currently visible from a pyQuil Program object’s .out() method): # Entering rewiring and # Exiting rewiring. From the perspective of the user, both comments serve the same purpose: # Entering rewiring: #(n0 n1 ... nk) indicates that the logical qubit labeled j in the program has been assigned to lie on the physical qubit labeled nj on the device. This is strictly for human-readability: these comments are discarded and have no effect.

SWAPs¶

When the compiler needs to move an instruction’s qubits closer it will insert SWAP gates which can be costly. If, however, the swaps are inserted at the very beginning of the program, the compiler can treat them as virtual swaps which do not appear in the resulting program but instead affect the initial rewiring of the program.

For example, consider running a CZ on non-neighboring qubits on a linear device:

import networkx as nx
from pyquil import Program, get_qc
from pyquil.api import QCSClient
from pyquil.api._quantum_computer import _get_qvm_with_topology
from pyquil.gates import CZ

graph = nx.from_edgelist([(0, 1), (1, 2)])
qc = _get_qvm_with_topology(
    client_configuration=QCSClient(),
    name="line",
    topology=graph,
    noisy=False,
    qvm_type="qvm",
    compiler_timeout=30.0,
    execution_timeout=30.0
)

p = Program(CZ(0, 2))
print(qc.compile(p))

CZ 2 1
HALT

We see that the resulting program has only a single CZ even though the original program would usually require the insertion of a SWAP gate. The compiler instead opted to just relabel (or rewire) the qubits, thus not inflating the number of gates in the result.

For larger and more complex programs (with more entanglement) it may not always be possible to avoid inserting swaps. For example, the following program requires a SWAP that increases its gate depth:

import networkx as nx
from pyquil import Program, get_qc
from pyquil.api import QCSClient
from pyquil.api._quantum_computer import _get_qvm_with_topology
from pyquil.gates import H, CZ

graph = nx.from_edgelist([(0, 1), (1, 2)])
qc = _get_qvm_with_topology(
    client_configuration=QCSClient(),
    name="line",
    topology=graph,
    noisy=False,
    qvm_type="qvm",
    compiler_timeout=30.0,
    execution_timeout=30.0
)

p = Program(CZ(0, 1), H(0), CZ(1, 2), CZ(0, 2))
print(qc.compile(p))

CZ 2 1
RX(-pi/2) 2
RX(-pi/2) 2
CZ 2 1
CZ 1 0
RZ(pi/2) 1
RX(-pi/2) 1
RX(-pi/2) 1
RX(-pi/2) 2
RX(pi/2) 2
XY(pi) 2 1
RX(-pi/2) 1
CZ 1 0
RZ(pi/2) 1
RX(pi/2) 2
RX(pi/2) 2
HALT

Note

SWAP gates generally cost three CZ gates or three XY gates. However, if your device has both CZ and XY gates available, then the compiler can produce a SWAP gate that uses only two two-qubit gates (one CZ and one XY).

Initial rewiring¶

In addition, you have some control over how the compiler constructs its rewiring, which is controlled by PRAGMA INITIAL_REWIRING. The syntax is as follows.

# <type> can be NAIVE, RANDOM, PARTIAL, or GREEDY
#
# The double quotes are required.
PRAGMA INITIAL_REWIRING "<type>"

Including this before any non-pragmas will allow the compiler to alter its rewiring behavior.

The default initial rewiring strategy¶

Note

Each initial rewiring strategy is described in more detail after the discussion about defaults.

When no INITIAL_REWIRING pragma is provided the compiler will choose one of two options depending on the program:

NAIVE: The qubits used in all instructions in the program satisfy the topological constraints of the device.
PARTIAL: Otherwise.

For example, if your program consists of two-qubit instructions where the qubits in each instruction are nearest neighbors on the device, the compiler will employ the native strategy:

from pyquil import Program, get_qc
from pyquil.gates import CZ

qc = get_qc("Aspen-X", as_qvm=True)
p = Program(CZ(3, 4))

print(qc.compile(p))

CZ 3 4

In the above example, CZ 3 4 touches qubits that are already nearest neighbors (and support a CZ instruction) and so the compiler employs the naive strategy (and thus does not rewire those qubits to use better ones).

If however, the program uses qubits that must be rewired, then the compiler defaults to the partial strategy:

from pyquil import Program, get_qc
from pyquil.gates import CZ

qc = get_qc("Aspen-X", as_qvm=True)
p = Program(CZ(3, 4))

print(qc.compile(p))

RZ(-pi/2) 0
RX(pi/2) 0
RZ(-pi/2) 0
RZ(pi/2) 1
XY(pi) 1 0
RZ(pi/2) 1
RX(pi/2) 1
RZ(-pi/2) 1
XY(pi) 1 0
RZ(-pi/2) 0
RX(-pi/2) 0

NAIVE¶

In this mode, the compiler chooses the naive mapping between logical qubits and physical qubits, where logical qubit i is assigned to physical qubit i. With this initial rewiring, the compiler will generally not move an instruction’s qubits around even if it results in a poor execution fidelity. For example assume that Aspen-X has a low-fidelity CZ 0 1, then compiling this program with naive rewiring will not move the CZ to a better qubit pair:

from pyquil import Program, get_qc
from pyquil.gates import CZ

qc = get_qc("Aspen-X", as_qvm=True)
p = Program('PRAGMA INITIAL_REWIRING "NAIVE"', CZ(0, 1))

print(qc.compile(p))

PRAGMA INITIAL_REWIRING "NAIVE"
CZ 0 1

If, however, your program includes an instruction that does not use neighboring qubits the compiler will be required to insert swaps (virtual or real, see swaps) that might affect the logical-physical qubit mapping. For example,

from pyquil import Program, get_qc
from pyquil.gates import CZ

qc = get_qc("Aspen-X", as_qvm=True)
p = Program('PRAGMA INITIAL_REWIRING "NAIVE"', CZ(0, 2))

print(qc.compile(p))

PRAGMA INITIAL_REWIRING "NAIVE"
CZ 6 5

In the above program CZ 0 2 is not a native instruction (meaning it cannot be directly executed on the target device) and so the compiler must insert a swap (virtual, in this case) into the program. When rewiring must occur in this mode it is not guaranteed that the resulting program will have optimal fidelity.

PARTIAL¶

In this mode, the compiler begins with an empty mapping from logical qubits to physical qubits. During the progression of compilation this mapping will be filled-in, and thus at any point the mapping is said to be partial. Generally this gives the compiler the opportunity to assign a logical-to-physical qubit mapping that optimizes the fidelity of the resulting program by incorporating fidelity information about any qubit in the device ISA.

For example, if the instruction CZ 0 1 has poor fidelity, under the partial rewiring strategy the compiler can find an alternative that improves the program fidelity:

from pyquil import Program, get_qc
from pyquil.gates import CZ

qc = get_qc("Aspen-X", as_qvm=True)
p = Program('PRAGMA INITIAL_REWIRING "PARTIAL"', CZ(0, 1))

print(qc.compile(p))

PRAGMA INITIAL_REWIRING "PARTIAL"
CZ 20 27

Here the compiler sees that the instruction CZ 20 27 will produce a program with better fidelity and so opts to reassign qubits in the original program.

GREEDY¶

In this mode, the compiler chooses an initial mapping between logical and physical qubits based upon a greedy optimization of the distances between qubits used in the program and those available on the device. When compared to the PARTIAL strategy it is generally more efficient because it uses a simple heuristic; however, it will also produce a program with worse overall fidelity. If compilation feels too slow and you’re willing to trade fidelity for compilation speed, then you may see success with this strategy.

Which strategy should I use?¶

Generally, as quantum software engineers, we want to maximize the execution fidelity of our programs. In other cases, however, for example in QCVV, we want to have more control about where instructions are placed.

Choosing an initial rewiring strategy¶
Desired effect	Recommended initial rewiring strategy
Maximize program execution fidelity	`PARTIAL`
Preserve, where possible, the qubits used in the input program	`NAIVE`
Faster qubit allocation at expense of fidelity	`GREEDY`

Note that each of these have drawbacks described in the sections above.

Common Error Messages¶

The compiler itself is subject to some limitations, and some of the more commonly observed errors follow:

! ! ! Error: Matrices do not lie in the same projective class. The compiler attempted to decompose an operator as native Quil instructions, and the resulting instructions do not match the original operator. This can happen when the original operator is not a unitary matrix, and could indicate an invalid DEFGATE block. In some rare circumstances, it can also happen due to floating point precision issues. In the latter case, the issue is resolved simply by recompiling the program. If you issue cannot be solved, please contact support@rigetti.com or post an issue to the github project page.