OrchesPy: Device-independent library for NumPy program on heterogeneous system

OrchesPy is a library for NumPy programs to execute part of the program on accelerators by decorating functions.

Prerequisites

Using OrchesPy requires the following packages.

• Python 3: tested with Python 3.8.
• NumPy 1.23.2: since CuPy and NLCPy require numpy>=1.17 and NLCPy requires numpy<=1.23.2.
• To run programs on VE, install NLCPy = 2.2.0 and its dependencies such as veoffload.
• To run programs on CUDA GPU, install CuPy = 11.0.0 and its dependencies such as CUDA toolkit working with your GPU.
• Install Inter-Device Copy Library = 0.1.0b1 for GPU-VE transfer.

To build OrchesPy, see also the section "Install from source".

Installation

You can install OrchesPy from PyPI or from source.

Install from PyPI

Execute the following command.

$ pip install orchespy

Install from source

To build OrchesPy, install CUDA toolkit and veoffload (VEO) for CuPy and NLCPy. Download the source tree from GitHub.

$ git clone https://github.com/SX-Aurora/orchespy.git

Execute the following command.

$ cd orchespy
$ pip install .

PIP will install dependencies automatically, and build and install OrchesPy on your environment.

Documentation

• OrchesPy User's Guide

License

The BSD-3-Clause license (see LICENSE file).