Stream-K Optimization and Exploration

Abstract

We explore optimization options for the Stream-K algorithm, a work-centric parallelization of matrix multiplication (GEMM). In our study, we investigated differences between the theoretical and practical implementations, particularly noting the impact of padding. Our debugging efforts revealed a persistent bug related to block mapping, which we could not fully resolve, but we managed to implement some optimizations. Setting the padding to zero for the M, N, and K dimensions resulted in an average 0.6% improvement in performance, achieving 1.44 ms, 89.37 TFlops, and 66.91 GB/s. However, adjusting the block size and parameters led to the process getting stuck, indicating a need for further tuning. Additionally, exploring the potential of Block2Time highlighted its promise in enhancing runtime predictions and optimizing load balancing.