Single Instruction, Multiple Data (SIMD) technology is crucial for enhancing computational efficiency in High-Performance Computing (HPC) and Artificial Intelligence (AI). While automatic vectorization methods offer ease of use, they suffer from limitations in hardware utilization due to compilers’ static analysis capabilities. Manual vectorization, on the other hand, allows for fine-grained control and potentially better hardware utilization, but manual approaches using low-level intrinsics specifically introduce challenges in portability and development complexity. Existing C++ SIMD libraries aim to address these issues but introduce new challenges such as performance and usability fragmentation and underutilization of hardware potential due to limited support for variable vector element counts. To overcome these limitations, this paper introduces HybridSIMD, a novel unified and autotunable SIMD library. HybridSIMD is designed to resolve both fragmentation and hardware underutilization by enabling operator-level hybrid collaborative optimization across different SIMD libraries through a unified interface. A built-in auto-tuning mechanism, leveraging static analysis and hierarchical search, automatically optimizes and tunes programs for high performance across diverse hardware platforms without human intervention. Experimental results across six real-world HPC benchmarks on AVX2, AVX512, and NEON architectures demonstrate that HybridSIMD outperforms state-of-the-art SIMD libraries. Notably, the highest speedups achieved by HybridSIMD are 185.34$\times$ on AVX2, 97.80$\times$ on AVX512, and 71.32$\times$ on NEON, showcasing superior computational efficiency and adaptability.