strange behaviour with nested loops inside CUDA kernel

deleted (by the author) due to silly question