AMD launches GPUFORT, an open-source attempt against NVIDIA's CUDA
Red team developed GPUFORT with the intention of providing developers with an open-source alternative to NVIDIA’s walled garden. To help developers mitigate with the confines of the CUDA proprietary software, AMD has begun to transfer CUDA code to be compatible with Radeon’s parallel computing stack.
GPUFORT has been identified as a research project built under the Radeon Open Ecosystem (ROCm) as a “source-to-source” translation system. At this stage, AMD has concentrated on the C and C ++ programming portion of CUDA Fortran code. The large CUDA code base is translated with Python code in a non-automated process via GPUFORT.
While GPUFORT is somewhat in its early stages, it is capable of completing syntax checks. The project can successfully translate CUDA code and High Performance Computing (HPC) applications to be compatible with ROCm using the Python code base. However, GPUFORT has pitfalls in some areas such as utilizing the complete OpenACC standard, identifying the portions of code that can be downloaded as well as reorganizing loops. AMD released the source code for the project publicly on GitHub under the MIT license.
Source(s)
