144 kernel definitions with workloads, reference-scalar starting points and expert baselines from ncnn, llama.cpp, Arm SIMD Loops and KleidiAI, plus end-to-end GEMMs of Qwen3.5-4B and Llama-3.1-8B-Instruct.
Co-authored-by: Shuaiwei Huang <112870732+ArtificialRay@users.noreply.github.com>
Co-authored-by: Allen Wu <110312634+wuallenwu@users.noreply.github.com>
Co-authored-by: Weihua Du <73473861+StigLidu@users.noreply.github.com>
Co-authored-by: shuaiwei huang <shuaiwei huang@users.noreply.huggingface.co>
Co-authored-by: allen wu <allen wu@users.noreply.huggingface.co>
Co-authored-by: weihua du <weihua du@users.noreply.huggingface.co>