For any instance where we don't need/want blue noise we use curand's XORWOW random number generator. Now while curand likes to claim that they are fast, a XORWOW state struct is 44 bytes in size. This is very impractical for memory performance. In turn I have now encountered multiple occasions where it was beneficial to replace calls to that one with xorshift. At the end of the day it will probably be best to introduce a custom random number generator that takes up 4 or 8 bytes at most so that the state fits into register space without any issues.
For any instance where we don't need/want blue noise we use curand's XORWOW random number generator. Now while curand likes to claim that they are fast, a XORWOW state struct is 44 bytes in size. This is very impractical for memory performance. In turn I have now encountered multiple occasions where it was beneficial to replace calls to that one with xorshift. At the end of the day it will probably be best to introduce a custom random number generator that takes up 4 or 8 bytes at most so that the state fits into register space without any issues.