Why pure CUDA kernel implementation is slower than Thrust implementation?

Question

I\'m new in CUDA programming and I\'m probably something wrong. My aim is to transform point cloud. I implemented this transformation both CUDA kernel and Thrust. Thrust imp