CuPy can also be used in conjunction with other frameworks.
__array_ufunc__ interface (see NEP 13 — A Mechanism for Overriding Ufuncs for details).
This enables NumPy ufuncs to be directly operated on CuPy arrays.
__array_ufunc__ feature requires NumPy 1.13 or later.
import cupy import numpy arr = cupy.random.randn(1, 2, 3, 4).astype(cupy.float32) result = numpy.sum(arr) print(type(result)) # => <class 'cupy.core.core.ndarray'>
cupy.ndarray also implements
__array_function__ interface (see NEP 18 — A dispatch mechanism for NumPy’s high level array functions for details).
This enables code using NumPy to be directly operated on CuPy arrays.
__array_function__ feature requires NumPy 1.16 or later; note that this is currently defined as an experimental feature of NumPy and you need to specify the environment variable (
NUMPY_EXPERIMENTAL_ARRAY_FUNCTION=1) to enable it.
Numba is a Python JIT compiler with NumPy support.
__cuda_array_interface__, which is the CUDA array interchange interface compatible with Numba v0.39.0 or later (see CUDA Array Interface for details).
It means you can pass CuPy arrays to kernels JITed with Numba.
The following is a simple example code borrowed from numba/numba#2860:
import cupy from numba import cuda @cuda.jit def add(x, y, out): start = cuda.grid(1) stride = cuda.gridsize(1) for i in range(start, x.shape, stride): out[i] = x[i] + y[i] a = cupy.arange(10) b = a * 2 out = cupy.zeros_like(a) print(out) # => [0 0 0 0 0 0 0 0 0 0] add[1, 32](a, b, out) print(out) # => [ 0 3 6 9 12 15 18 21 24 27]
cupy.asarray() supports zero-copy conversion from Numba CUDA array to CuPy array.
import numpy import numba import cupy x = numpy.arange(10) # type: numpy.ndarray x_numba = numba.cuda.to_device(x) # type: numba.cuda.cudadrv.devicearray.DeviceNDArray x_cupy = cupy.asarray(x_numba) # type: cupy.ndarray
MPI for Python (mpi4py) is a Python wrapper for the Message Passing Interface (MPI) libraries.
MPI is the most widely used standard for high-performance inter-process communications. Recently several MPI vendors, including Open MPI and MVAPICH, have extended their support beyond the v3.1 standard to enable “CUDA-awareness”; that is, passing CUDA device pointers directly to MPI calls to avoid explicit data movement between the host and the device.
With the aforementioned
__cuda_array_interface__ standard implemented in CuPy, mpi4py now provides (experimental) support for passing CuPy arrays to MPI calls, provided that mpi4py is built against a CUDA-aware MPI implementation. The following is a simple example code borrowed from mpi4py Tutorial:
# To run this script with N MPI processes, do # mpiexec -n N python this_script.py import cupy from mpi4py import MPI comm = MPI.COMM_WORLD size = comm.Get_size() # Allreduce sendbuf = cupy.arange(10, dtype='i') recvbuf = cupy.empty_like(sendbuf) comm.Allreduce(sendbuf, recvbuf) assert cupy.allclose(recvbuf, sendbuf*size)
This new feature will be officially released in mpi4py 3.1.0. To try it out, please build mpi4py from source for the time being. See the mpi4py website for more information.
DLPack is a specification of tensor structure to share tensors among frameworks.
Zero-copy conversion from a DLPack tensor to a
Here is a simple example:
import cupy # Create a CuPy array. cx1 = cupy.random.randn(1, 2, 3, 4).astype(cupy.float32) # Convert it into a DLPack tensor. dx = cx1.toDlpack() # Convert it back to a CuPy array. cx2 = cupy.fromDlpack(dx)
Here is an example of converting PyTorch tensor into
import cupy import torch from torch.utils.dlpack import to_dlpack from torch.utils.dlpack import from_dlpack # Create a PyTorch tensor. tx1 = torch.randn(1, 2, 3, 4).cuda() # Convert it into a DLPack tensor. dx = to_dlpack(tx1) # Convert it into a CuPy array. cx = cupy.fromDlpack(dx) # Convert it back to a PyTorch tensor. tx2 = from_dlpack(cx.toDlpack())
Note that as of DLPack v0.3 for correctness it (implicitly) requires users to ensure that such conversion (both importing and exporting a CuPy array) must happen on the same CUDA/HIP stream. If in doubt, the current CuPy stream in use can be fetched by, for example, calling
cupy.cuda.get_current_stream(). Please consult the other framework’s documentation for how to access and control the streams. This requirement might be relaxed/changed in a future DLPack version.