Aer function to check the memory requirement upfront when using simulators

[patdenim]

What is the expected behavior?

It would be nice to know what is the memory required before executing a circuit using a simulator.
It is impossible to exactly know the memory needed when using the extended_simulator. Looking at the code extended_stabilizer_state.hpp. The formula required_chi() used by the extended simulator does not seem to be only a function of the exact gate counts in the circuit

size_t State::required_memory_mb(uint_t num_qubits,
                                 const std::vectoroperations::Op &ops) const {
  size_t required_chi = compute_chi(ops.cbegin(), ops.cend());
  // 5 vectors of num_qubits*8byte words
  // Plus 2*CHSimulator::scalar_t which has 3 4 byte words

Here is an example that shows this would help if knowing the exact memory required per circuit.
Consider 2 different circuits, with the following gates (Circuit A is having less gates than Circuit B) :

Circuit A
OrderedDict([
('u1', 272),
('sx', 133),
('cx', 114),
('h', 36),
('measure', 36)])

Circuit B
OrderedDict([
('u1', 480),
('cx', 228),
('sx', 216),
('h', 36),
('measure', 36)])

Then execute those 2 circuits on the extended_stabilizer

extended_stabilizer_simulator = AerSimulator(method='extended_stabilizer')
# Transpile circuit for backend
tcirc = transpile(circuit, extended_stabilizer_simulator)
extended_stabilizer_result = extended_stabilizer_simulator.run(tcirc, shots=1).result()
print('This succeeded?: {}'.format(extended_stabilizer_result.success))

Circuit A is failing with message

Simulation failed and returned the following error message:
ERROR: a circuit requires more memory than max_memory_mb.

while Circuit B works. (even though Circuit B is much bigger than Circuit A)

[doichanj]

I found issue in calling required_memory_mb function that State::set_config is not called before calculating required memory. In extended stabilizer, extended_stabilizer_approximation_error was not set from config parameters. This parameter can control size requirement when setting lerger error acceptance

[doichanj]

By the way, extended stabilizer allocates state buffer whose size is calculated by the same logic as compute_chi function at beginning. I think setting larger extended_stabilizer_approximation_error is the only way to decrease memory usage

[doichanj]

I think this should be fixed by PR #1933 with option commented above