QubitUnitary is directly passed to cpp layer with adjoint_diff

[chaeyeunpark]

The following code now raises an error from C++ layer:

import sys
import pennylane as qml
from pennylane import numpy as np
import json

dev = qml.device('lightning.qubit', wires=3)

@qml.qnode(dev, diff_method="adjoint")
def circuit(theta):
    arr = np.eye(8)

    qml.QubitUnitary(arr, wires = [0, 1,2])
    qml.RX(theta, wires = [1])
    qml.adjoint(qml.QubitUnitary)(arr, wires = [0, 1,2])
    return qml.expval(qml.PauliX(0))

if __name__ == "__main__":
    params = np.array([0.1])
    grad = qml.grad(circuit)
    print(grad(params))

Here's the full output:

Traceback (most recent call last):
  File "/home/chae-yeun/work/Xanadu/pennylane-tests/test_qubit_unitary.py", line 20, in <module>
    print(grad(params))
  File "/home/chae-yeun/work/Xanadu/pennylane/pennylane/_grad.py", line 115, in __call__
    grad_value, ans = grad_fn(*args, **kwargs)
  File "/home/chae-yeun/work/Xanadu/devel/lib/python3.10/site-packages/autograd/wrap_util.py", line 20, in nary_f
    return unary_operator(unary_f, x, *nary_op_args, **nary_op_kwargs)
  File "/home/chae-yeun/work/Xanadu/pennylane/pennylane/_grad.py", line 133, in _grad_with_forward
    vjp, ans = _make_vjp(fun, x)
  File "/home/chae-yeun/work/Xanadu/devel/lib/python3.10/site-packages/autograd/core.py", line 10, in make_vjp
    end_value, end_node =  trace(start_node, fun, x)
  File "/home/chae-yeun/work/Xanadu/devel/lib/python3.10/site-packages/autograd/tracer.py", line 10, in trace
    end_box = fun(start_box)
  File "/home/chae-yeun/work/Xanadu/devel/lib/python3.10/site-packages/autograd/wrap_util.py", line 15, in unary_f
    return fun(*subargs, **kwargs)
  File "/home/chae-yeun/work/Xanadu/pennylane/pennylane/qnode.py", line 867, in __call__
    res = qml.execute(
  File "/home/chae-yeun/work/Xanadu/pennylane/pennylane/interfaces/execution.py", line 483, in execute
    res = _execute(
  File "/home/chae-yeun/work/Xanadu/pennylane/pennylane/interfaces/autograd.py", line 326, in execute
    return _execute(
  File "/home/chae-yeun/work/Xanadu/devel/lib/python3.10/site-packages/autograd/tracer.py", line 44, in f_wrapped
    ans = f_wrapped(*argvals, **kwargs)
  File "/home/chae-yeun/work/Xanadu/devel/lib/python3.10/site-packages/autograd/tracer.py", line 48, in f_wrapped
    return f_raw(*args, **kwargs)
  File "/home/chae-yeun/work/Xanadu/pennylane/pennylane/interfaces/autograd.py", line 370, in _execute
    res, jacs = execute_fn(unwrapped_tapes, **gradient_kwargs)
  File "/usr/lib/python3.10/contextlib.py", line 79, in inner
    return func(*args, **kwds)
  File "/home/chae-yeun/work/Xanadu/pennylane/pennylane/_device.py", line 571, in execute_and_gradients
    jacs.append(gradient_method(circuit, **kwargs))
  File "/home/chae-yeun/work/Xanadu/pennylane-lightning/pennylane_lightning/lightning_qubit.py", line 603, in adjoint_jacobian
    jac = adjoint_diff.adjoint_jacobian(
IndexError: _Map_base::at

Partially, the problem is that 3-qubit QubitUnitary does not have proper decomposition. Still, it is also possible to support (non-trainable) QubitUnitary in Lightning as we anyway support it in the forward path (by obtaining the matrix and apply it using e.g. applyMultiQubitOp defined in the LM kernel).

Thus I think we need to decide whether we want to support those operations in adjoint-diff path in Lightning or raise a proper error in the Python layer.