ACTA: Automatic Configuration of the Tensor Memory Accelerator for High-End GPUs

Achieving peak GPU performance requires optimizing data locality and asynchronous execution to minimize memory access costs and overlap computation with transfers.While features like the Tensor Memory Accelerator (TMA) and warp specialization address these challenges, their complexity often limits programmers.In this work, we present ACTA (Automatic Configuration of the Tensor Memory Accelerator), a software library that simplifies and optimizes TMA usage.By leveraging the GPU Specification Table (GST), ACTA dynamically determines the optimal tile sizes and queue configurations for each kernel and architecture.Its algorithm ensures efficient overlap between memory and computation, drastically reducing programming complexity and eliminating the need for exhaustive design space exploration.Our evaluation across a diverse set of GPU kernels demonstrates that ACTA achieves performance within 2.78% of exhaustive tuning while requiring only a single configuration pass.This makes ACTA a practical and efficient solution for optimizing modern GPU workloads, combining near-optimal performance with significantly reduced programming effort.