Use of Instruction instead of LLVM intrinsic in Tensorize schedule #13850

Development
#1

|Hi TVM team,

I am writing a custom operator in TVM and adding a Tensorize schedule for its execution. I See that we can use the LLVM intrinsics(for example: llvm.aarch64.neon.uaddlp, llvm.aarch64.neon.addp) through tvm.tir.call_llvm_pure_intrin() call. I just wanted to know if we can use directly the instruction instead of Intrinsic(for example: %3 = mul <8 x i16> %1, %0 for vmulq_u16).

Look forward to your reply.

Thanks

#2

It’s ugly but possible. See the change to arm_cpu/tensor_intrin.py in https://github.com/apache/tvm/pull/5754. Note that we no longer have this code in main.