Add Support for ROCM (AMD GPUs)

[jredmondson]

Summary

Nvidia GPUs are ubiquitous in the market and darknet supports them well. However, it would be nice to have options in the AMD GPU market, which tends to be less expensive and often has more VRAM available for darknet training and usage than many of their Nvidia counterparts.

Though there are two options for non-Nvidia support (OpenCL and HIP/ROCM), the mapping between HIP and CUDA calls should be more of a one-to-one, and easier to support. Various papers/research surveys have also noted that OpenCL is generally slower (often by multiples like 2-8x, depending on operations).

In a hackathon, a group submitted a pull request to the old darknet repo with support for HIP. The integration was crude with #defines used to replace CUDA calls with direct HIP calls. It would be appreciated to review this pull request and/or tools that exist for CUDA->HIP porting to support modern AMD GPU optimizations so that the .weights models can be directly loaded onto the GPU without conversions to onnx.

Goals

• Implement HIP/ROCM support in darknet for training and running weights models
• Integrate HIP/ROCM compilation into the cmake build process

[stephanecharette]

This will require the purchase of some AMD hardware, since I only have NVIDIA-based systems.