high performance computing on graphics processing units: hgpu.org

GPUs are well-known for their remarkable ability to accelerate computations through massive parallelism. However, offloading computations to GPUs necessitates manual identification of code regions that should be executed on the device, memory that needs to be transferred, and synchronization to be handled. Recent work has leveraged the portable target offloading interface provided by LLVM/OpenMP, taking GPU acceleration to a new level. This approach, known as the direct GPU compilation scheme, involves compiling the entire host application for the GPU and executing it there, thereby eliminating the need for explicit offloading directives. Nonetheless, due to limitations of the current GPU compiler toolchain and execution, seamlessly executing CPU code on GPUs with certain features remains a significant challenge. In this paper, we examine the limits of CPU code execution on GPUs by applying the direct GPU compilation scheme to LLVM’s test-suite, analyze the encountered errors, and discuss potential solutions for enabling more code to execute on GPUs without any changes if feasible. By studying these issues, we shed light on how to improve GPU acceleration and make it more accessible to developers.