llama.cpp adds dp4a GEMM kernels to OpenCL backend

A narrow kernel patch, but one more sign mobile GPU inference keeps getting cheaper

Llama.cpp's latest build, tagged b11789, adds dp4a GEMM variants with vector activation loads to the project's OpenCL backend. The change comes from pull request #30753, according to its release notes. Dp4a is a low-level instruction many mobile and integrated GPUs use to speed up 8-bit integer dot products, the kind of arithmetic that dominates quantised model inference. The same patch renames associated kernels and environment variables. That housekeeping change will matter to anyone who has scripted around the old names.

The practical effect is a slightly faster matrix-multiply path for devices running llama.cpp through OpenCL rather than CUDA or Vulkan. That typically means Android phones and ARM-based laptops without dedicated Nvidia or AMD drivers. The release itself ships as a pre-release build, with binaries for macOS, Windows, Linux and Android. It also includes dedicated packages for Snapdragon devices covering their CPU, Adreno GPU and Hexagon NPU, plus a build for openEuler on both x86 and aarch64.

No benchmark numbers accompany the change, so the size of any speed-up is not stated. The commit is one of the small, frequent patches that make up llama.cpp's release cadence, rather than a headline feature.

Written from the primary source above by The Neural Net’s automated desk (claude-sonnet-5, theneuralnet/pipeline@1) without human review. How this works. Corrections: check the source first — if we got it wrong, the source link is the record.