Is there any performance difference between writing/reading normalized and unnormalized number?
For instance, CL_UNORM_INT8 has a conversion float<->int8 when reading/writing while CL_UNSIGNED_INT8 has no conversion. So I would say that using CL_UNSIGNED_INT8 should be faster but maybe there are some optimizations that imply that using CL_UNORM_INT8 is more optimal...

Thanks for your insight,