Porting CUDA Applications to Run on AMD GPUs
Giving developers the ability to write code once and use it on different platforms is important. Organizations are increasingly moving to open source and open standard solutions which can aid in code portability. AMD developed a porting solution that allows developers to port proprietary NVIDIA® CUDA® code to run on AMD graphic processing units (GPUs).
This paper describes the AMD ROCm™ open software platform which provides porting tools to convert NVIDIA CUDA code to AMD native open-source Heterogeneous Computing Interface for Portability (HIP) that can run on AMD Instinct™ accelerator hardware. The AMD solution addresses performance and portability needs of artificial intelligence (AI), machine learning (ML) and high performance computing (HPC) for application developers. Using the AMD ROCm platform, developers can port their GPU applications to run on AMD Instinct accelerators with very minimal changes to be able to run their code in both NVIDIA and AMD environments.
For example, TempoQuest (TQI) used AMD’s HIP tools on their AceCAST™ WRF weather prediction software to convert OpenACC-Fortran and CUDA-C code to run on AMD Instinct™ MI200 series GPUs. TQI developers indicate that converting the code using the HIP conversion tools was trivial with only a few minor changes required for performance tuning and to accommodate for some minor CUDA-to-HIP incompatibilities.
