A numpy.ndarray-like interface to internal HOOMD-blue data.


A __cuda_array_interface__ to internal HOOMD-blue data on the GPU.


Implement zero-copy array.

class hoomd.data.array.HOOMDArray(buffer, callback, read_only=None)

A numpy.ndarray-like interface to internal HOOMD-blue data.

HOOMD-blue’s zero copy local snapshot API (hoomd.State.cpu_local_snapshot) returns HOOMDArray objects. HOOMDArray acts like numpy.ndarray through NumPy’s provided interface. Some exceptions are the view, resize, flat and flatiter methods and the data and base properties.

To ensure memory safety, a HOOMDArray object cannot be accessed outside of the context manager in which it was created. Make an explicit copy to use the array elsewhere (e.g. numpy.array(obj, copy=True)).

In general this class should be nearly as fast as a standard NumPy array, but there is some overhead. This is mitigated by returning a numpy.ndarray whenever possible.

Performance Tips

Let a represent a HOOMDArray.

  • Place the HOOMDArray to the left of the expression (e.g. a + b + c is faster than b + a + c). This has to do with the mechanisms HOOMDArray has to do to hook into NumPy’s functionality.

  • Make copies as early as possible. In other words, if you will need access outside the context manager, use numpy.array(a, copy=True) before doing any calculations.

  • If you know that your access of the internal buffer is safe and we cannot detect this (i.e. we return a HOOMDArray), using HOOMDArray._coerce_to_ndarray should help. Note that for large arrays this only gives minimal performance improvements at greater risk of breaking your program.

class hoomd.data.array.HOOMDGPUArray(*args, **kwargs)

A __cuda_array_interface__ to internal HOOMD-blue data on the GPU.

The HOOMDGPUArray object exposes a GPU data buffer using __cuda_array_interface__. This class provides buffer access through a context manager to prevent invalid memory accesses (hoomd.State.gpu_local_snapshot). To avoid errors, use arrays only within the relevant context manager. For example:

with sim.state.gpu_local_snapshot as data:
    pos = cupy.array(data.particles.position, copy=False)
    pos[:, 2] += 1


When CuPy can be imported, then this class wraps much of the cupy.ndarray class’s functionality. Otherwise, this class exposes only the buffer.

HOOMDGPUArray always supports getting (but not setting) the shape, strides, and ndim properties. HOOMDGPUArray never supports standard binary operators like (+, -, *). This is a current limitation on external classes hooking into CuPy.

When CuPy can be imported, slice/element assignment (e.g. array[0] = 1; array[:2] = 4) and compound assignment operators (e.g. +=, -=, *=) are available. In addition, most methods besides view, resize, flat, flatiter are available. The same is true for properties except the data and base properties. See CuPy’s documentation for a list of methods.


Use, cupy.add, cupy.multiply, etc. for binary operations on the GPU.


Packages like Numba and PyTorch can use HOOMDGPUArray without CuPy installed. Any package that supports version 2 of the __cuda_array_interface__ should support the direct use of HOOMDGPUArray objects.