Noise in final result

[lyd405121]

I am using cuda 10.2 to compile the code.
But I found noise block in the image.

------------------The only code I have modified -------------------------
Some function are deprecated， like __ballot __shfl_up __shfl，etc...
Then I modify all of them from __ballot(*) to __ballot_sync(0xfffffff, *)
So did I do it right？