How can I use QuantumStateGPU.get_vector() to get cupy.array?

[yuguoshao]

In my experiments , I want to use state vector to do some numeral calculations. So I wish that I can get state vector on GPU directly. Compared to Numpy, Cupy or PyTorch may be better choices. But I can't find out how get_vector() runs. I can only find:

def get_vector(self): # real signature unknown; restored from __doc__
        """
        get_vector(self: qulacs.QuantumState) -> numpy.ndarray[complex128[m, 1]]
        
        Get state vector
        """
        pass

Is there anyway to help me?

[corryvrequan]

Unfortunately, qulacs does not have an interface to get a pointer to state vectors on GPU or directly pass it to other GPU numerical libraries.
If we allow overheads of two copies from GPU->main ram and main ram -> GPU, we can copy the data managed in qulacs to other libraries. For example:

import qulacs
num_qubit = 10
statevec_on_qulacs = qulacs.QuantumStateGpu(10) # allocate 2^10 complex vector on GPU
statevec_on_cpu = statevec_on_qulacs.get_vector() # copy from GPU -> main ram
statevec_on_cupy = cupy.array(statevec_on_cpu) # copy from main ram -> GPU

Qulacs is currently written in C++/CUDA and its functions are exported to python with pybind. So, I think it is hard to provide an interface to directly convert internal state vectors on GPU to Cupy or relevant libraries unless pybind supports it...

[yuguoshao]

Okay , thank you very much ! I hope one day quantum state can convert to vector directly on Gpu.😭😭😭

[leofang]

Hi, a CuPy dev here. May I ask is the difficulty in that the statevector is not stored contiguously in memory? Otherwise there must be a way to get the pointer address and memory size, and CuPy can take from there.

[corryvrequan]

Hi @leofang .
Thanks for your comments, and I'm very sorry for the late reply.

May I ask is the difficulty in that the state vector is not stored contiguously in memory?
No. The state vector is stored contiguously in memory in Qulacs.

Therefore, I agree that there is no technical problem in passing the pointer address of memory, and I didn't know CuPy supports wrapping foreign memory address as a CuPy array object.
cupy/cupy#4644

While I couldn't find a simple way to export a function to return CuPy object with pybind11, if it is useful for users, it is easy to implement a function to return the pointer address of state vector on GPU.

[leofang]

@corryvrequan Yes, I think it'd be useful. At the pybind11 level the pointer address to the start of the state vector (assuming there's no extra requirement in padding / alignment / offset) can be returned as either uintptr_t or a Python int. Then, at the Python level you can use cupy.cuda.UnownedMemory to wrap it as you found out in the linked issue. If you decide to give it a go I'm happy to help review (it should need only few lines of code) 🙂