GPU execution models and programming solutions

As already mentioned, GPUs serve as accelerators to CPUs, i.e., computationally intensive tasks are off-loaded from CPUs to GPUs. Standard programming languages such as Fortran and C/C++ do not permit such off-loading because they lack addressing distinct memory spaces and don’t know the GPU architecture. For that purpose, we need special language extensions that allow GPU programming.

Many solutions exist for programming GPUs and we will talk about the two most used, i.e., CUDA and OpenCL. CUDA (Compute Unified Device Architecture) is a set of extensions for higher-level programming languages (C, C++ and Fortran) developed by NVIDIA for its GPUs. CUDA comes with a developer toolkit for compiling, debugging and profiling programs. It’s the first solution for GPU programming (the latest version is 11.6), but unfortunately only GPUs manufactured by NVIDIA support it.

Another solution is OpenCL (Open Computing Language) which is a standard open-source programming model initially developed by major manufacturers (Apple, Intel, ATI/AMD, NVIDIA) and is now maintained by Khronos. It also provides extensions to C, while C++ is supported in SYCL (a similar but independent solution by Khronos). Although its programming/execution model is similar to CUDA, it is more low-level. It can also come with a developer toolkit, depending on the hardware, but its main advantage over CUDA is that it’s supported by many types of Processing Units (CPUs, GPUs, FPGAs, MICs…) and in reality oriented to heterogeneous computing. In principle, that means an OpenCL program can run either on a GPU (not depending on the manufacturer) or on a CPU (or any other PU). OpenCL’s latest standard is currently at 3.0. Unfortunately, the NVIDIA GPUs does not support it (contrary to Intel and AMD GPUs), the support is still offered only for OpenCL 1.2.