Copyright 2021 Fluid Numerics LLC
This repository is used to define HIP kernels for 3-D array transpose operations.
This repository is motivated by the following needs :
- GPU developers are increasingly interested in using open-source and portable ("code once, run anywhere") software for GPU acceleration.
- 3-D Array Transposes are a critical operation for multi-dimensional Fourier Spectral methods and machine learning algorithms.
Currently, porTTgpu implements "naive" tensor transpose operations that achieve suboptimal performance. You can review the current benchmarks across Nvidia and AMD hardware here
Our goal is to improve performance of the 3-D array transpose kernels by exploiting coallesced global memory access patterns (Contributions welcome).