Hi,
The 2 main pieces of software that use NVidia cards are PopSift (optional dependency) and DepthMap.
HIP looks like a great option for porting from CUDA to OpenCL for AMD. It deals with the naming differences and it seems to deal with the thread indexing, where indexing dimensions and memory locality are total opposites in CUDA and OpenCL. However, HIP does not help in porting textures because OpenCL doesn’t have any.
Unfortunately, both PopSift and the DepthMap stage of AliceVision use textures everywhere, and a lot of their speed gain comes from the hardware interpolation engine that is provided by NVidia’s texture hardware.
It is possible to write a software layer that emulates CUDA textures for AMD cards, but right now, we have a higher priority to add CUDA streams and get multi-card support for several CUDA cards to work in DepthMap.
Best regards,
Carsten