SamKG opened a new issue #8263: URL: https://github.com/apache/tvm/issues/8263
It is well-established that performance of PRNG algorithms differs greatly on different hardware [1]. Currently, TVM seems to only support threefry random number generation. However, a user who is deploying to a GPU may benefit greatly from Philox instead (the paper linked below demonstrates a 3x performance difference!). In addition, decreasing the precision of the PRNG algorithm can also improve performance where suitable. Ideally, TVM would allow for the choice of PRNG to be user-selectable or tunable in some way. I'd be happy to give this issue a stab myself if someone could offer some pointers. [1] http://www.thesalmons.org/john/random123/papers/random123sc11.pdf -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected]
