AIDE-QC

Advanced Integrated Development Environment for Quantum Computing

Pass Manager

The QCOR infrastructure has a wide variety of quantum circuit transformation plugins, which are often referred to as transpilers in the literature.

There are two categories of transpilers: circuit optimization and placement.

Circuit optimizers are plugins that can rewrite the input circuits into a more optimal form (fewer gates) while reserving the intended transformation (e.g. in terms of the overall unitary matrix).

Placement plugins act as the final router to map or embed the circuit into the backend hardware topology. This placement action is not relevant for simulation but is very important for NISQ hardware execution due to limited connectivity.

These reentrant transpiler plugins are executed as passes over the circuit IR tree. The sequence of these passes is determined by the QCOR’s pass manager. The pass manager will execute circuit optimization passes depending on the optimization level or manual command-line settings, then a single placement pass (selectable) taking into account the target qpu topology.

Basic usage 

There are 3 levels of optimization that have been pre-defined in QCOR:

  • Level 0: no optimization.

  • Level 1: rotation-folding, single-qubit-gate-merging, and circuit-optimizer

  • Level 2: level 1 passes + two-qubit-block-merging

The descriptions of those optimization passes are shown below.

Pass nameDescription
circuit-optimizerA collection of simple pattern-matching-based circuit optimization routines.
single-qubit-gate-mergingCombines adjacent single-qubit gates and finds a shorter equivalent sequence if possible.
two-qubit-block-mergingCombines a sequence of adjacent one and two-qubit gates operating on a pair of qubits and tries to find a more optimal gate sequence via Cartan decomposition if possible.
rotation-foldingA wrapper of the Staq ’s RotationOptimizer which implemented the rotation gate merging algorithm.
voqcA wrapper of the VOQC (Verified Optimizer for Quantum Circuits) OCaml library, which implements generic gate propagation and cancellation optimization strategy.

The optimization level can be specified by -opt [0,1,2] the QCOR command-line option.

The pass manager also collected detailed statistics about each optimization pass that it executed. This information can be printed out by providing the -print-opt-stats option to the QCOR compiler when compiling your source code.

For backend placement, by default, QCOR will perform the swap-shortest-path placement if the backend has topology (connectivity) information.
Different placement strategies can be specified using the -placement command-line option.

The list of available placement services is shown in the following.

Placement nameDescription
swap-shortest-pathWrapper of the Staq ’s SwapMapper
triQNoise adaptive layout based on TriQ library.
enfieldWrapper of enfield allocators. This needs to be installed manually from the XACC fork

Advanced usage 

Custom sequences of optimization passes can be specified using the following compile option:

qcor -qpu qpp -opt-pass pass1[,pass2] source.cpp

in which, users provide an ordered (comma-separated) list of passes to be executed using the -opt-pass key.

This can be used in conjunction with the -print-opt-stats option to collect the statistics if required.

Manual qubit mapping can be used instead of topology-based placement by providing the -qubit-map option followed by the list of qubit indices specifying the mapping.

For example, -qubit-map 2,1,4,3,0 dictates the following mapping 0->2, 1->1, 2->4, 3->3, 4->0.
It is important to note that no further placement is performed after the -qubit-map based mapping. Hence, users need to make sure that the mapping is appropriate for the target hardware backend.

Examples 

Circuit Optimization 

In this example, we just compile a simple circuit using the QCOR compiler.
The circuit contains a pair of mergeable rotation gates.

C++Python
__qpu__ void test_circ(qreg q) {
  H(q[0]);
  Rx(q[0], 0.123);
  Rx(q[0], 0.456);
}

int main() {
  // Create a qubit register
  auto q = qalloc(1);
  // Run the quantum kernel
  test_circ(q);
  q.print();
}
@qjit
def test_circ(q : qreg):
  H(q[0])
  Rx(q[0], 0.123)
  Rx(q[0], 0.456)


# Create a qubit register
q = qalloc(1)
# Run the quantum kernel
test_circ(q)
q.print()
qcor -opt 1 -print-opt-stats simple_circuit.cpp 
./a.out
python3 simple_circuit.py -opt 1 -print-opt-stats

We compiled the above QCOR kernel source using optimization level 1 and request optimization statistics.

When the executable is executed, we can see the optimization data printed out, which confirm that the two Rx rotation gates
have been merged into one.

 - Number of Gates Before: 3
 - Number of Gates After: 2
--------------------------------
| GATE     |BEFORE   |AFTER    |
--------------------------------
| Rx       |2        |1        |
| H        |1        |1        |
--------------------------------

Placement 

Let’s take a look at an example where qubit placement is required for target hardware backends.

// Create a GHZ state
__qpu__ void create_ghz(qreg q) {
  H(q[0]);
  CX(q[0], q[1]);
  CX(q[0], q[2]);
  for (int i = 0; i < q.size(); i++) {
    Measure(q[i]);
  }
}

int main() {
  // Print out the kernel for submission
  auto q = qalloc(3);
  create_ghz::print_kernel(std::cout, q);
}

It’s a simple experiment to create a 3-qubit entangle state. Albeit its simplicity, not all quantum devices can execute this circuit as is since it requires triangular connectivity between qubit 0, 1, and 2.

Let’s compile and execute the above source code targeting two different hardware backends:

ibmq_ourenseibmqx2
0 -- 1 -- 2 
     |      
     3       
     |       
     4  
      1
    / | 
   /  |
  /   |
 /    |
0 --  2  --  3
      |     /
      |    /
      |   / 
      |  /
      4
qcor -qpu ibm:ibmq_ourense simplePlacement.cpp 
./a.out
qcor -qpu ibm:ibmqx2 simplePlacement.cpp 
./a.out
H q0
CNOT q1,q0
CNOT q0,q1
CNOT q1,q2
Measure q1
Measure q0
Measure q2
H q0
CNOT q0,q1
CNOT q0,q2
Measure q0
Measure q1
Measure q2

As we can see, due to the limited connectivity of the ourense backend, an additional CNOT gate is required (plus the remapping of qubits) after placement. The qx2 backend has triangular connectivity groups (0-1-2 and 2-3-4), and thus it can map the GHZ state-preparation circuit to its qubits directly.