Are there any te schedule examples that exceed the cublas using the tensor core?

Questions
#1

Hi, are there any te schedule examples that exceed cublas using the tensor core? Such as dense op? Thanks!

#2

I’m afraid not. Please try TensorIR with meta-schedule :wink:

#3

May I ask if it is because of the limitation of te schedule? :sob:

#4

Could you tell me why? Thanks! :kissing_heart:

#5

There are serval reasons:

  1. On technical, TE schedule is based on scalar computing and the schedule tree, which makes it hard (or even impossible) to generate well-performed tensorized programs.
  2. For the community reason. As we are pushing TensorIR and Meta-schedule, which has stronger expression ability and performance in most cases, there is little motivation to reproduce it using legacy technics with great difficulty.
#6

OK, thanks! I’m trying the tensorir, it’s very good!