The purpose of this directory is to demonstrate C++ metaprogramming features that are available in MLIR core. At a high-level, metaprogramming can be interpreted as “programming with a level of indirection”.
This experimental mlir::ModelBuilder is used to generate operations using the Linalg Dialect.
This exposes the mlir::edsc::ValueHandle and mlir::edsc::StructuredIndexed classes and other functionality into an mlir::ModelBuilder. to build an MLIR function for a whole model (in the case of the TestMNISTJit example, a 3-MLP).
The MLIR function can be compiled using MLIR passes and progressively lowered to LLVMIR. The mlir::ModelRunner then kicks in and produces an executable version in memory that can be called directly, by name, from the main function.
This allows writing a 3-MLP test in C++, allocate some buffers, JIT-compile and run it end-to-end, as the test demonstrates.
Note that for the moment mlir::ModelBuilder does not perform any type of advanced transformations and basically only lowers LLVM, where LLVM tools (opt and llc) pick it up and apply their optimizations.
At this time this is only intended for prototyping and experimenting on vanilla Linux CPU targets.