MLIR interpreter

MLIR interpreter is a development tool, mainly intended for debugging. As the name implies, it executes MLIR programs. What makes it special is that it implements dialects at all levels of abstraction and all standard data types (tensor, memref, vector, scalar). One use case is to track down compiler bugs by executing the IR after each one and looking for changed results.

Features:

• Tensors, memrefs, vectors, scalars
• Many dialects supported: from mhlo via linalg and scf all the way to arith.
• Easy to extend (look at dialects/affine.cc for an example).

Creating a new dialect

To create a new dialect, follow the following steps:

1. Create a .cc implementation file in dialects/. The name of the file should match your dialect name.
2. Add a dependency to your dialect to the dialects:dialects target.
3. Each op is implemented in a function.
   • Op results:
     • If your op has no result, the function should be void.
     • If your op has a tensor, memref or vector result, the return type should be InterpreterValue.
     • Scalars can be returned as their corresponding C++ type (e.g. int64_t, float)
     • You can also return a SmallVector if your op has multiple results.
   • Op arguments:
     • Your function should take an InterpreterState&, an instance of your op and any arguments your op takes.
     • Like above, arguments can either be taken as InterpreterValues, as scalar C++ types or as SmallVector thereof (for variadic arguments).
   • Error handling:
     • If some of the inputs are invalid, use InterpreterState::AddFailure to report this.
4. Register the ops: REGISTER_MLIR_INTERPRETER_OP(YourOp).
5. Add test cases in dialects/tests.