[TOPI][ONNX] Fix for trilu and set_matrix_diag ops

[mikepapadim]

This PR addresses and extends issues of PRs #10873 and #9329.

Refactor the implementation of set_matrix_diag to enable support of the trilu operator in ONNX.

@sfvaroglu @bfgoldstein @AndrewZhaoLuo

[AndrewZhaoLuo]

@mikepapadim can you make sure the tests pass?

[mikepapadim]

I am trying to reproduce them locally. Will update when I have them fixed.

[shingjan]

Overall LGTM. Do we have an update on the PR? I am trying to add triu for our pytorch frontend and this PR could be really useful if merged! Thanks

[shingjan]

are these two lines necessary? i am getting a tables not found on my end and it seems like those two libs aren't referenced anyway

[shingjan]

I am hitting the following error:

E             0: tvm::codegen::CodeGenLLVM::CreateBufferPtr(llvm::Value*, tvm::runtime::DataType, llvm::ArrayRef<llvm::Value*>, tvm::runtime::DataType)
E                   at /home/yj/tvm/src/target/llvm/codegen_llvm.cc:737
E             File "/home/yj/tvm/src/target/llvm/codegen_llvm.cc", line 737
E           TVMError: 
E           ---------------------------------------------------------------
E           An error occurred during the execution of TVM.
E           For more information, please see: https://tvm.apache.org/docs/errors.html
E           ---------------------------------------------------------------
E             Check failed: (index->getType()->isIntegerTy()) is false: Expected buffer index to be an integer

I wonder if there are still places in the codegen that assumes k1,k2 as integer

[AndrewZhaoLuo]

I think it's probably a spurious import (e.g. typed Expr and IDE autoimported this when it should be tvms)

[shingjan]

@mikepapadim I took a look and looks like we need to add

    if not isinstance(k_one, tvm.te.Tensor):
        k_one = const_vector(np.asarray([k_one], dtype=np.int64))
    if not isinstance(k_two, tvm.te.Tensor):
        k_two = const_vector(np.asarray([k_two], dtype=np.int64))

to here:

tvm/python/tvm/topi/transform.py

Line 922 in 057dd7c

[shingjan]

Suggested change

from tables import Expr