I have read related papers on TVM, and have repeatedly mentioned Halide’s idea of separation of scheduling and calculation. May I ask which layer or stage of the TVM compilation process occurs in TVM for the separation of scheduling and calculation? Is there any specific document explaining how TVM uses this idea? Hope to answer, thank you.
Currently in TVM, we use
te.compute to define computation, and
te.schedule for scheduling. After scheduling is done, we lower it to TIR, and then do code generation
Excuse me, in the process of writing tvm code, which file is this part of the separation scheduling code in? Let’s study how his code is implemented at the bottom