CNN and RVV Co-design for Efficient Model Serving

Paper i proceeding, 2025

Convolutional algorithm performance depends on layer dimensions, with SIMD demands and cache sharing influencing runtime selection. To identify the best settings, we perform a co-design exploration of convolutional layer parameters and three algorithms: Direct, im2col+GEMM, and Winograd, jointly with hardware parameters for RISC-V vector architectures. Our results show that incorporating hardware parameters with layer dimensions boosts execution time and efficiency, emphasizing the need for co-design.