Do concurrent selective accelerator? host, device execution

The blog post discusses the limitations of using 'do concurrent' for GPU acceleration in Fortran and proposes a new keyword 'RESIDENT' that would allow developers to specify that loop iterations should only access data from its original location. This would avoid unnecessary data copying between CPU and GPU, optimizing performance for certain tasks that do not benefit from offloading to the GPU. The proposal aims to facilitate parallel execution on both CPU and GPU while maintaining programmer control over data locality.