Pass GPU Data Between MATLAB and Python

R2026b
Since R2026b

This page explains how to use the functions for transferring GPU data between MATLAB® and Python®, and how to choose between them.

After you pass GPU data to Python, you can call Python GPU computing functions directly from MATLAB. For more information about calling Python from MATLAB, see Directly Call Python Functionality from MATLAB. For information about transferring non-GPU data to Python, see Pass Data Between MATLAB and Python from MATLAB.

To avoid slowing down your code, move data between GPU memory and host memory only when necessary. All of the functions listed below can pass GPU data between MATLAB and Python without first moving the data to host memory.

Pass GPU Data from MATLAB to Python

If you have GPU data in MATLAB, represented by a gpuArray, you can use these functions to send the data to Python.

Your Task in PythonTarget Python LibrariesData Transfer Function

Generic GPU computing

CuPy

pycupyarray
Deep learningPyTorch®arrayToTorchTensor (Deep Learning Toolbox)
Deep learning

Any library that implements the DLPack specification, including TensorFlow

pydlpack

Pass GPU Data from Python to MATLAB

If you have GPU data in Python, you can use these functions to send the data to MATLAB.

Python ObjectOrigin Python LibraryData Transfer Function

cupy.ndarray

CuPy

gpuArray
DLPack

  • Any library that implements the DLPack specification and supports creating a DLPack capsule, including TensorFlow.

  • If you create your own custom DLPack capsule, then you do not need any specific Python library.

torch.Tensor

PyTorchtorchTensorToArray (Deep Learning Toolbox)

By using these functions to send and receive GPU data, you can perform GPU computing across MATLAB and Python.

Process Data on GPU in Python

First, convert a gpuArray to a cupy.ndarray.

X = rand(1000,"gpuArray");
pyX = pycupyarray(X);

Operate on the CuPy array, on the GPU, using Python functions. Replace the call to myFunctions.proccessData with your desired Python functions.

pyZ = py.myFunctions.processData(pyY)

Transfer the resulting Python data back to MATLAB.

Z = gpuArray(pyZ);

Use Multiple GPUs

If you have multiple GPUs, consult this table to determine which GPU your data is sent to.

DirectionGPU Data Destination
MATLAB to Python

The GPU device currently selected by MATLAB.

To see the currently selected GPU device, call gpuDevice.

Python to MATLAB

The GPU device currently selected by MATLAB or, if no device is selected, the default GPU device.

To see this device, call gpuDevice.

Note

When you transfer GPU data between MATLAB and Python, using the same GPU in both environments usually gives the best performance.

See Also

|

Topics