We ported CAFFE from CUDA to HIP – and here’s what happened…
gpuopen.com
gpuopen.com
> When running on CUDA, the ported HIP application uses a compilation, runtime, and library flow that is very similar to the original CUDA application. Thus, we would expect the HIP/CUDA application to deliver very similar performance as the original CUDA/CUDA app.
You'd expect? I don't know if I'd trust benchmarks AMD put out on this subject, but I'd trust them more than not even bothering to benchmark.
But it seems like they were able to increase portability at no loss of performance (and even it looks like a performance increase in some cases if I read correctly), among other things -- which is always commendable.