llama.cpp b10991 Restores Fast-Path Routines for Qualcomm Hexagon DSP
Release b10991 of llama.cpp restores missing contiguous fast-path operations and the hvx_copy_uu routines for the Qualcomm Hexagon DSP backend. This update prevents performance degradation when executing LLM inference on Snapdragon devices powered by Hexagon DSP hardware. It maintains optimized vector memory operations across mobile and edge deployments using Qualcomm chips. The patch specifically resolves PR #28886 by re-integrating HVX (Hexagon Vector eXtensions) memory copy instructions required for each execution run. Pre-built release binaries have been generated for Linux, Windows, macOS, and Android platforms.
## BACKGROUND
Qualcomm Hexagon is a digital signal processor (DSP) architecture built into Snapdragon system-on-chips, featuring Hexagon Vector eXtensions (HVX) for high-performance SIMD processing. llama.cpp is an open-source inference engine designed to execute large language models efficiently across a wide variety of hardware acceleration backends.