Portable Data-Parallel Python Extensions with oneAPI

Nikita Grigorian and Oleksandr Pavlyk

Sources and details for the Portable Data-Parallel Python Extensions with oneAPI poster at the SciPy 2024 conference. This poster presents ongoing work to enable writing, building, and implementing portable extensions for data-parallel computation in Python.

About the authors

This poster was made by and is being presented on behalf of the Python team at Intel.

dpctl

A package with Python bindings for DPC++ runtime classes and an Array API Specification conformant tensor submodule built on pure SYCL for portability.

numba-dpex

An extension to the Numba JIT compiler which provides a SYCL-like API for kernel programming and an extension to Numba's parallelizer to generate and offload kernels to data-parallel devices.

dpnp

An oneAPI- and oneMKL-powered array library built as a drop-in replacement for Numpy that can be executed on any data-parallel device.