Apparently it should only be useful for people with a strong gpu but weak cpu since it moves texture decoding to the gpu.
yea but the main problem of emulation is not in texture's complexity, unfortunately

Did the new commit fix whatever was killing OpenCL performance (something about having to transfer textures back and forth)?
Regarding OpenCL, if I enable this option I get full speed in NSMB with EFB -> Ram Enabled complete with spinning coins. However only for the DX9 backend as there is no speedup with DX11, haven't tested OpenGL yet.