llama.cpp b9222 adds Hexagon TRI op for faster AI inference

The latest release of llama.cpp (b9222) introduces support for the TRI (Tensor Reduction and Interleaving) operation on Qualcomm Hexagon HTP (Hexagon Tensor Processor) cores. This addition, contributed by Todor Boinovski and Max Krasnyansky from Qualcomm, enables more efficient neural network inference on Hexagon-based hardware – commonly found in smartphones and edge devices. The TRI op is critical for optimizing certain tensor operations used in large language models, allowing better utilization of Hexagon's vector processing capabilities.

Beyond the Hexagon enhancement, this release includes a broad set of platform builds: macOS (Apple Silicon, Intel, iOS), Linux (x64, arm64, s390x, with Vulkan, ROCm 7.2, OpenVINO, SYCL FP32/FP16), Windows (x64, arm64, CUDA 12/13, Vulkan, SYCL, HIP), Android (arm64 CPU), and openEuler (x86 and aarch64 with various backends). The release also cleans up merge conflict artifacts and editor configuration errors in the Hexagon unary and ggml op code. This comprehensive packaging makes llama.cpp more accessible for developers targeting diverse hardware accelerators.