Passing a `qvector` to a multi-qubit custom operation fails

[ikkoham]

Required prerequisites

• Consult the security policy. If reporting a security vulnerability, do not report the bug using this form. Use the process described in the policy to report the issue.
• Make sure you've read the documentation. Your issue may be addressed there.
• Search the issue tracker to verify that this hasn't already been reported. +1 or comment there if it has.
• If possible, make a PR with a failing test to give us a starting point to work on!

Describe the bug

Describe the bug

PR #4173 ("Support qvector arguments in custom operations", fixes #2452) added support for calling a custom operation with a whole qvector:

• custom_op(qvec)
• custom_op(*qvec)

However, C++ front-end (CUDAQ_REGISTER_OPERATION) was not updated.

Steps to reproduce the bug

C++ (fails)

// repro.cpp
#include <cudaq.h>

// 2-qubit custom operation (SWAP), row-major, {re,im}
CUDAQ_REGISTER_OPERATION(MyGate, 2, 0, {
  {1.0,0.0},{0.0,0.0},{0.0,0.0},{0.0,0.0},
  {0.0,0.0},{0.0,0.0},{1.0,0.0},{0.0,0.0},
  {0.0,0.0},{1.0,0.0},{0.0,0.0},{0.0,0.0},
  {0.0,0.0},{0.0,0.0},{0.0,0.0},{1.0,0.0}
});

struct kernel {
  void operator()() __qpu__ {
    cudaq::qvector q(2);
    x(q[0]);
    MyGate(q);            // <-- whole qvector: FAILS to compile
    // MyGate(q[0], q[1]); // <-- explicit qubits: works fine
  }
};

int main() { auto c = cudaq::sample(kernel{}); c.dump(); return 0; }

$ nvq++ repro.cpp --target nvidia -o repro
repro.cpp:15:5: error: 'quake.custom_unitary_constant' op Invalid matrix size, required 2^N * 2^N for N-qubit operation
    MyGate(q);
    ^

Using explicit qubit refs MyGate(q[0], q[1]) compiles and runs correctly
({ 01:1000 }, SWAP on |10> -> |01>).

Python (same operation, whole qvector works):

import cudaq, numpy as np
cudaq.register_operation("mygate", np.array(
    [1,0,0,0, 0,0,1,0, 0,1,0,0, 0,0,0,1], dtype=complex))

@cudaq.kernel
def kernel():
    q = cudaq.qvector(2)
    x(q[0])
    mygate(q)          # whole qvector: works in Python (PR #4173)

print(cudaq.sample(kernel))   # { 01:1000 }

Expected behavior

Work with no error

Is this a regression? If it is, put the last known working version (or commit) here.

Not a regression

Environment

• CUDA-Q version: 0.15.0
• Python version:
• C++ compiler: nvq++
• Operating system:

Suggestions

No response

[ikkoham]

Another issue: C++'s custom operator is slower than Python's one. This is because parse → get-concrete-matrix in cudaq-quake is slow (Python uses memcpy from numpy array).

[arcusbuilds]

I'm working on a fix for this, following the same approach as #4173. Will open a PR soon.

[schweitzpgi]

C++ should not support this feature.

In the C++ CUDA-Q embedding, quantum sequence types can be passed unambiguously in 2 positions.

1. op<cudaq::ctrl>(veq, targ)
2. op(veq) when op takes a single targ

We do not plan to support more ad hoc overloading in C++ as those overloadings don't follow strong typing, make understanding the code difficult from ad hoc differentiation, and invite error-prone code.

[ikkoham]

I understand the reasoning, and I agree that we shouldn't support this feature.
However, I think the significant API difference between Python and C++ should be documented clearly. This issue was raised because a user asked about it.
Also, I'll file a separate issue for the performance problem where custom gates are slower in C++ than in Python.