PR #19054 opencl: add flattened q6_K mv

opencl: flatten `q6_K` and add `kernel_mul_mv_q6_K_f32_flat`

lhez committed 187 days ago

opencl: clean up

lhez committed 187 days ago

opencl: refactor q6_K mv - put loop body in `block_q_6_K_dot_y_flat`

lhez committed 187 days ago

opencl: tweak the workgroup size a bit

lhez committed 187 days ago

opencl: output 4 values per subgroup for `kernel_mul_mv_q6_K_f32_flat`

lhez committed 187 days ago

opencl: proper alignment for q6_K

lhez committed 187 days ago

opencl: boundary handling for flattened q6_K mv

lhez committed 187 days ago

opencl: rename q6_K mv kernel file

lhez committed 187 days ago

opencl: put flattened q6_K mv in its own file

lhez committed 187 days ago

opencl: use lower k in file name

lhez committed 187 days ago

opencl: use K in variable names

lhez committed 187 days ago

llama.cpp opencl: add flattened q6_K mv #19054 Merged