llama.cpp

b3d75875 - vulkan: optimize im2col (#21713)

vulkan: optimize im2col (#21713) * vulkan: improve im2col memory write layout * cap workgroups * minimal device tuning * use vendor_id instead of subgroup size