AMD officially released open-source “GPUFORT”

by ·

Nvidia’s dominance in the field of high-performance computing relies heavily on CUDA-based solutions, recently, AMD provided the code through a project called GPUFORT. According to Phoronix, the project belongs to the Radeon Open eCosystem (ROCm) to help large CUDA code bases work outside the Nvidia ecosystem.

Nvidia’s solution greatly restricts developers from looking for alternatives and affects migration work, AMD has been working hard for a long time to help developers migrate as much CUDA-specific code as possible to the interface supported by the Radeon open source computing stack. AMD used both C and C++ code in previous projects. GPUFORT is different in that it converts OpenACC-based CUDA Fortran and Fortran code to OpenMP 4.5+ for execution on GPU or Fortran + HIP C++ code.

According to reports, GPUFORT is not a compiler itself, but a Python code library that performs a source-to-source conversion. At the same time, GPUFORT is not a complete automation solution, it needs to review and manually correct the content generated based on CUDA coding. In the current state, it is only used to convert high-performance computing (HPC) applications into compatible code formats supported by the AMD ROCm ecosystem. AMD engineers admitted that they are still analyzing the composition of the code to improve the accuracy of the compilation and that the complete OpenACC standard has not yet been implemented.
AMD released GPUFORT on GitHub, showing code examples and how to handle them, as well as user installation guides.

Tags: