How to leverage tvm.tir.transform in TVM 0.7

Questions
1

Hi everyone, I’ve been working for a couple of months on the TVM stack and I really like it :slight_smile:

I have a question related to the use of TVM 0.7 APIs, in particular tvm.tir.transform.Simplify.

In TVM 0.6 I could simply call tvm.tir.ir_pass.Simplify(stmt) in any of my custom IR passes to leverage the benefits of simplified arithmetics. I would then be able to apply further optimizations on top of the simplified IR. I read the documentation (here) and I’ve tried several ways to obtain the same result in TVM 0.7 but I have not found a way to leverage tvm.tir.transform.Simplify. I think I might be missing something.

What is the correct way to use this API? Could you show me an example in which you pass a tir.PrimFunc or a stmt object to the new API?

Thank you!

2

We have moved to having all passes use the pass infrastructure now, you should be able to just invoke Simplify on your entire IRModule before running your pass. i.e Simplify()(mod).

3

If you want to just simplify a single expression/stmt you can use the internal API to do so.

Here is how simplify pass is defined:

Pass Simplify() {
  auto pass_func = [](PrimFunc f, IRModule m, PassContext ctx) {
    auto* n = f.CopyOnWrite();
    arith::Analyzer analyzer;
    n->body = arith::StmtSimplifier(&analyzer).Simplify(std::move(n->body));
    return f;
  };
  return CreatePrimFuncPass(pass_func, 0, "tir.Simplify", {});
}

 arith::Analyzer analyzer;
 arith::StmtSimplifier(&analyzer).Simplify(stmt)
4

Thank you @jroesch, that definitely brings in more clarity.

The way I was using this pass in its previous version was to call it on stmt within my own pass as a step after some IR transformations, and it worked flawlessly.

Looking at your answer, it seems that what I am looking for is the Python API equivalent of

 arith::Analyzer analyzer;
 arith::StmtSimplifier(&analyzer).Simplify(stmt)

Basically what I was looking for is being able to run the Simplify pass in my own pass at a function or statement granularity rather than the whole module.

5

I think someone just needs to expose this in Python you could probably do so with

TVM_REGISTER_GLOBAL("SimplifyStmt").set_body_typed([](tir::Stmt stmt) {
    arith::Analyzer analyzer;
    return arith::StmtSimplifier(&analyzer).Simplify(stmt);
});

and in Python

tvm.get_global_func("SimplifyStmt")(my_stmt)
6

Thank you @jroesch, this is very helpful!