Falcon: A Scalable Analytical Cache Model

Explore a groundbreaking 17-minute video presentation from PLDI 2024 introducing Falcon, a scalable analytical cache model for affine programs. Discover how this innovative approach outperforms existing methods by efficiently analyzing large-scale programs, including 46 TorchVision neural networks. Learn about Falcon's ability to split computations into smaller, manageable pieces, avoiding the worst-case asymptotic behavior of Presburger solvers. Understand the model's impressive performance, with a geomean runtime of 44.9 seconds compared to over 32 minutes for state-of-the-art alternatives. Gain insights into Falcon's parallelization capabilities, efficient updating after local modifications, and its potential applications in superoptimizers, worst-case execution time analyses, and manual program optimization. Presented by researchers from the University of Edinburgh, Advanced Micro Devices, and the University of Cambridge, this talk offers valuable knowledge for compiler designers, performance engineers, and anyone interested in advanced cache modeling techniques.