TensorComprehensions icon indicating copy to clipboard operation
TensorComprehensions copied to clipboard

Support for user-provided cache key to reuse autotuning results on different machines

Open seongwook-ham opened this issue 7 years ago • 3 comments

@ttheodor i autotune the function in server with 4gpu and get cache file. and i try to reuse generated cache file in different server with 4gpu. but it throws error 2 server has diffenet gpu. @nicolasvasilache said it is key related problem. could you give me some instrunction to reuse cache files?

seongwook-ham avatar Mar 27 '18 22:03 seongwook-ham

@ttheodor for more context I am thinking we could override the automatic HW key by some user specified string for load / store. The user is then responsible for ensuring things are proper.

@seongwook-ham note that performance on the machine you did not autotune for will be not great or even bad. If you want good results, autotune on the target HW (for the foreseeable future).

nicolasvasilache avatar Mar 27 '18 22:03 nicolasvasilache

@ttheodor when will it be possible to use #233 ? and how could i use it?

seongwook-ham avatar Mar 29 '18 20:03 seongwook-ham

@ttheodor is there some progress?

seongwook-ham avatar Apr 04 '18 03:04 seongwook-ham