Hi, all!
I find both tutorial and paper are using float32 as the data type. And the tensorcore experimental repo seems not update for a long time. So does Ansor support TensorCore INT8 now? Thanks a lot!
2 Likes
Sorry for that. Though we have a strong will to support TensorCore in Ansor, currently I don’t have extra bandwidth to work on this topic.
As far as I know, some guys are working on the new TensorIR, and based on which, TVM will get a new infrastructure to combine the current AutoTVM and AutoScheduler together. The auto schedule search with TensorCore support will be fully supported then.
p.s. The repo you got is a good example to write extra sketch rules, and it provides an TensorCore implementation which should work well. Check the GitDiff, these codes should be easy to understand.
4 Likes