How to disable half storage mode in nlist (Python api)

[zahraAkbari8]

Hi

I need the neighborlist to add a custom force based on the local density of particles in my simulation. I am using
nl = hoomd.md.nlist.cell(buffer=0.4,default_r_cut = 15)
to create the neighbor list.
This is the custom force function which gets the neighbor list and colors the neighbors of a particular type of particle.

class MyCustomForce(hoomd.md.force.Custom):
    def __init__(self):
        super().__init__(aniso=True)

    def set_forces(self, timestep):
        with self.cpu_local_force_arrays as arrays:
            with sim.state.cpu_local_snapshot as data:
                 neicolors = np.zeros(N)
                 cellids = np.where(data.particles.typeid[:] == 0)[0]

                 with nl.NeighborListLocalAccess as nlistarrays:
                    n_list = nlistarrays.nlist
                    n_neigh =nlistarrays.n_neigh
                    head_list = nlistarrays.head_list
                    for iparticle in cellids:
                        neighbors = n_list[head_list[iparticle]:head_list[iparticle]+n_neigh[iparticle]] 
                        neicolors[neighbors] = 1

These are the results or time step 0 and 1.

I believe the reason why only half of the neighbors are considered in the next time step might be related to the half_nlist property being on, But I could not find a way to change this. Could you please help me understand what might be the solution?

Thanks in advance.

[joaander]

The CPU device uses half neighbor lists and the GPU device uses full neighbor lists.

[zahraAkbari8]

Thanks for your response. Are there any workarounds? Can i change something in the source code to run on CPU but access the full neighbor list?

[joaander]

No, there are no workarounds. Your proposed change to the NeighborList code would break pair force computations and many other NeighborList consumers.

[zahraAkbari8]

Thank you for your answer. I'm trying to run the same code on GPU and it works but sometimes I get the following error:

with nl.gpu_local_nlist_arrays as nlistarrays:
         ^^^^^^^^^^^^^^^^^^^^^^^^^
    self._cpp_obj.compute(self._simulation.timestep)
RuntimeError: Cannot acquire access to array in use.

My code:

class MyCustomForce(hoomd.md.force.Custom):
    def __init__(self):
        super().__init__(aniso=True)

    def set_forces(self, timestep):
         
        with self.gpu_local_force_arrays as arrays:

            with sim.state.gpu_local_snapshot as data:
                 ids = cupy.array(data.particles.typeid, copy=False)
                 
                 idlist= cupy.array(data.particles.typeid, copy=False) 
                 cellids = np.where(idlist[:] == 0)[0]
                 Ncells= len(cellids)
                 Positions = cupy.array(data.particles.position, copy=False)
                 with nl.gpu_local_nlist_arrays as nlistarrays:
                    n_list = nlistarrays.nlist._coerce_to_ndarray()
                    n_neigh =nlistarrays.n_neigh._coerce_to_ndarray()
                    head_list =nlistarrays.head_list._coerce_to_ndarray()
                    
                    # Some calculations on neighbors

Thanks for your help in advance.

[joaander]

This error is the result of a bug in HOOMD-blue. I will need a complete script that reproduces the problem in order to find the cause.

[zahraAkbari8]

I changed the code to this and now i don't get the error anymore:

class MyCustomForce(hoomd.md.force.Custom):
    def __init__(self):
        super().__init__(aniso=True)

    def set_forces(self, timestep):
         
        with self.gpu_local_force_arrays as arrays:
            with sim.state.gpu_local_snapshot as data:
                     with nl.gpu_local_nlist_arrays as nlistarrays:
                          ids = cupy.array(data.particles.typeid, copy=False)
                 
                          idlist= cupy.array(data.particles.typeid, copy=False) 
                          cellids = np.where(idlist[:] == 0)[0]
                          Ncells= len(cellids)
                          Positions = cupy.array(data.particles.position, copy=False)
                          n_list = nlistarrays.nlist._coerce_to_ndarray()
                          n_neigh =nlistarrays.n_neigh._coerce_to_ndarray()
                          head_list =nlistarrays.head_list._coerce_to_ndarray()
                    
                         # Some calculations on neighbors