llama.cpp Release b10772 Adds F16 ABS Unary Operation to Qualcomm Hexagon Backend
Release b10772 of llama.cpp adds support for half-precision floating-point (F16) absolute value (ABS) unary operations to the Qualcomm Hexagon HTP backend. The update refactors the backend to unify F32 and F16 unary operation dispatching into a single function, passing all on-device verification tests on Hexagon v81 hardware. Expanding F16 operator coverage on Qualcomm Hexagon DSPs enables llama.cpp to execute more neural network layers directly on mobile hardware accelerators. This reduces reliance on CPU fallback, improving both inference performance and energy efficiency on Snapdragon-powered devices. The update consolidates separate F32 and F16 unary execution routines into a single execute_op_unary function, avoiding code duplication in shared VTCM memory and DMA handling. Hardware verification was completed on a QRD8850 device (Hexagon v81), achieving 8 out of 8 test passes for the F16 ABS operator with zero CPU fallback.
## BACKGROUND
llama.cpp is a high-performance C/C++ framework for running Large Language Models (LLMs) locally across diverse hardware architectures. Qualcomm Hexagon is a digital signal processor (DSP) architecture in Snapdragon chipsets featuring Hexagon Tensor Processor (HTP) engines and Hexagon Vector eXtensions (HVX) designed to accelerate machine learning operations. Unary operations, such as absolute value (ABS) or square root (SQRT), take a single input tensor and apply a mathematical function element-wise.