llama.cpp Release b11982 Adds Hexagon Hardware Support for UPSCALE
Release b11982 of llama.cpp introduces Hexagon hardware support for the UPSCALE operation, specifically optimized for image-to-text scenarios such as qwen3-vl. The update includes basic upscale operations, optimized nearest and bilinear scaling, and vector instruction improvements using vgather. This update improves the performance and efficiency of running multimodal AI models on mobile and edge devices equipped with Qualcomm Snapdragon processors. It expands the range of hardware capable of handling complex vision-language tasks locally. The pull request adds support for non-contiguous rows and dimension collapsing, alongside specific tests for permuted UPSCALE operations to ensure stability. Kernel selection logic has also been relocated to the host for better maintainability and code cleanliness.
## BACKGROUND
llama.cpp is an open-source software library that performs inference on large language models in C and C++, serving as the core engine for many local AI tools. Qualcomm Hexagon is a specialized digital signal processor and neural processing unit found in Snapdragon chips, designed to accelerate AI and signal processing workloads efficiently.