Can this work for PTX coming from NVHPC OpenACC/stpar codes?

[sumseq]

I was wondering if ZLUDA requires the code to be written in CUDA, or if it can take an intermediate PTX code and compile that to AMD?

I ask because we use Fortran with NVIDIA GPUs using OpenACC/OpenMP-Target/Do concurrent.

Specifically, "do concurrent" offload is supported by both NVIDIA and Intel (although the Intel implementation is early and needs some work). However, there are ZERO compilers that have publicly said they would work on supporting "do concurrent" offload on AMD GPUs.

The ZLUDA project looks like it has promise to do just that.
If one could compile "do concurrent" code into PTX, and then run it through ZLUDA to get rocm code - could that work?

• Ron