numba-dpex 0.22.1

An extension for Numba to add data-parallel offload capability

Data-parallel Extension for Numba* (numba-dpex) is a standalone extension for the Numba Python JIT compiler. Numba-dpex provides a generic kernel programming API and an offload feature that extends Numba's auto-parallelizer to generate data-parallel kernels for parfor nodes.

Numba-dpex's kernel API has a design and API similar to Numba's cuda.jit module, but is based on the SYCL language. The code-generation for the kernel API currently supports SPIR-V-based OpenCL and oneAPI Level Zero devices that are supported by Intel® DPC++ SYCL compiler runtime. Supported devices include Intel® CPUs, integrated GPUs and discrete GPUs.

The offload functionality in numba-dpex is based on Numba's parfor loop-parallelizer. Our compiler extends Numba's parfor feature to generate kernels and offload them to devices supported by DPC++ SYCL compiler runtime. The offload functionality is supported via a new NumPy drop-in replacement library: dpnp. Note that dpnp and NumPy-based expressions can be used together in the same function, with dpnp expressions getting offloaded by numba-dpex and NumPy expressions getting parallelized by Numba.