llama.cpp Release b11970 Adds Conv2D Support for Qualcomm Hexagon DSPs
llama.cpp release b11970 introduces comprehensive Conv2D and depthwise convolution support, alongside performance optimizations designed for Qualcomm Hexagon Digital Signal Processors (DSPs). This update significantly improves the ability to run large language models and multimodal features efficiently on mobile and edge devices utilizing Qualcomm hardware. By leveraging dedicated DSP and neural processing units, it enables faster and more power-efficient execution on smartphones and IoT hardware. The implementation includes double-buffered staging, implicit GEMM techniques, and fully vectorized paths to maximize throughput. It also refactors DMA pipelines and kernel parameter management to seamlessly match other hardware acceleration backends.
## BACKGROUND
Qualcomm Hexagon is a family of digital signal processors and neural processing units designed for high-performance, low-power digital signal processing and machine learning workloads on mobile devices. Convolutional layers (Conv2D) are fundamental building blocks in neural networks, often used in multimodal vision adapters to process images before feeding them into language models.