Pushing the Limits of Online Auto-Tuning: Machine Code Optimization in Short-Running Kernels

This paper proposes an online auto-tuning approach for computing kernels. Differently from existing online auto-tuners, which regenerate code with long compilation chains from the source to the binary code, our approach consists on deploying auto-tuning directly at the level of machine code generation. This allows auto-tuning to pay off in very short-running applications. As a proof of concept, our approach is demonstrated in two benchmarks, which execute during hundreds of milliseconds to a few seconds only. In a CPU-bound kernel, the average speedups achieved are 1.10 to 1.58 depending on the target micro-architecture, up to 2.53 in the most favourable conditions (all run-time overheads included). In a memory-bound kernel, less favourable to our runtime auto-tuning optimizations, the average speedups are 1.04 to 1.10, up to 1.30 in the best configuration. Despite the short execution times of our benchmarks, the overhead of our runtime auto-tuning is between 0.2 and 4.2 % only of the total application execution times. By simulating the CPU-bound application in 11 different CPUs, we showed that, despite the clear hardware disadvantage of In-Order (IO) cores vs. Out-of-Order (OOO) equivalent cores, online auto-tuning in IO CPUs obtained an average speedup of 1.03 and an energy efficiency improvement of 39 % over the SIMD reference in OOO CPUs.