Overview

OrchesPy is a Python library for numerical programs working on the top of NumPy or NumPy-like libraries for accelerators in a heterogeneous system. OrchesPy facilitates porting a NumPy program to a heterogeneous system and reduces overheads of data transfer between accelerators.

Heterogeneous computing, using different types of computational units in a system, has been popular for performance and power efficiency. There are various accelerators for computing: GPU (NVIDIA, AMD, etc), Vector Engine (NEC), FPGA, etc. Accelerators have different characteristics in performance. For better performance, a system can have multiple types of accelerators with different performance characteristics.

It requires many modifications to port a program on such a heterogeneous system with multiple types of accelerators. A programmer needs to use different libraries for different accelerators. For example, to port a NumPy program to a system with NVIDIA CUDA GPU and NEC SX-Aurora TSUBASA, a program uses NumPy on host part, CuPy on GPU part and NLCPy on VE part. A programmer also needs to add data transfer code between parts running on different devices.

We developed OrchesPy, another NumPy-like library for heterogeneous computing to make it easier to port and optimize a NumPy program on a heterogeneous system with multiple types of accelerators. The name “OrchesPy” comes from that the library orchestrates NumPy-like Python libraries for various devices.

OrchesPy provides following three features:

• “device” directive-like decorator converting N-dimension arrays (array_like) in arguments transparently in order to execute code on a specified device,

• “transfer_array” function to transfer an array_like from a device to another device explicitly in portable way, and

• “transfer_array_content” function to transfer array_like to specified array_like.