We are developing a new hardware accelerator for Deep learning for that, we have the backend compiler
it just parses the layers information to the hardware (codgen). There is no concept of tiled data format in that and scheduling also so we want to use TVM stack and add our python based backend compiler.
So how can I integrate my hardware backend to TVM stack?
you can take a look at src/codegen, it converts TVM tensor IR to hardware related code. e.g. cuda/llvm/c/…