Efficiency of LLVM Compilers on High-Performance Architectures

The development of LLVM-based compilers introduces new possibilities for performance portability and modern standards in high-performance computing (HPC). This work evaluates their effectiveness specifically on AMD's GPU/CPU architectures, focusing on AMD Instinct accelerators. We assess compiler performance in terms of runtime efficiency, compilation overhead, and support for contemporary language features. The benchmarking suite includes both compute-bound and memorybound applications representative of real-world CPU/GPU-enabled HPC workloads. We compare several LLVM-driven compilers, such as Clang (from ROCm), AMD's AOCC and Intel's icx, as well as GNU's GCC and NVIDIA's NVCC, analyzing their ability to generate efficient code for AMD CPU/GPUs. This study offers practical insights for developers and compiler researchers aiming to optimize C applications on GNU, AMD, Intel CPU/GPU platforms.