Recently, I have rewritten a 3D nemerical model into GPU offload code, both with ifx and nvfortran. With ifx, it uses OPENMP directives. With nvfortran, it uses "DO CONCURRENT" standard Fortran. Running under Intel B580 GPU, the model is about 10x faster than serial code under Intel Core i5-12400F CPU, while under Nvidia A100 GPU it's about 25x.
I didn't use "DO CONCURRENT" in ifx because it's extra slow, transfering data between host and device all the time. To solve this issue, nvfortran gives 2 useful options: A. "-stdpar=gpu"; B."-stdpar=gpu -acc=gpu gpu=nomanaged". Both of the options are fast. Option A uses unified memory, automatically managing data transfer on need. Option B avoids most of the data transfer.
So does ifx support data transfer management for "DO CONCURRENT" now? Options like each of the 2 options above would be great. It saves a lot of coding time and makes ifx version and nvfortran version similar. I think option B is not hard for a compiler to achieve.
Wish it soon. Thanks!
