whisper.cpp Vulkan fails on Adreno 702 (Turnip); OpenCL needs subgroups
On an Adreno 702 with Mesa Turnip 25.2 (Debian 13, kernel 6.16), whisper.cpp Vulkan builds and loads the model, but every decode fails. v1.7.6 fails with Mesa 'workgroup barrier ... impossible to have enough concurrent waves'. Setting GGML_VK_DISABLE_F16=1 instead fails with 'Shared memory size too small for matrix multiplication' (16 KiB per workgroup). v1.8.7 fails with vk::Device::waitForFences ErrorDeviceLost plus a UCHE READ TRANSLATION GPU fault in the kernel log, even with async, graph optimization and fusion disabled. The ggml OpenCL backend rejects Mesa Rusticl on this GPU because it reports no cl_khr_subgroups, so don't expect a GPU speedup there; budget for CPU (tiny.en-q5_1 at 3 threads is about 0.67x real time on 4 Cortex-A53-class cores). Also, compiling the Vulkan shader translation unit on a 4 GB board ran out of memory; cross-build in an arm64 Docker container instead.