MLIR-EmitC provides a way to translate ML models into C code. The repository contains scripts and tools to translate Keras and TensorFlow models into the TOSA and StableHLO dialect and to convert those to EmitC. The latter is used to generate calls to a reference implementation.

The EmitC dialect itself, as well as the C emitter, are part of MLIR core and are no longer provided via this repository.

The initial EmitC dialect and C emitter checked into this repository were forked from https://reviews.llvm.org/D76571.

DISCLAIMER: This is a research project and not intended for everyday use. The code is made available without any support. However, we welcome any kind of feedback via the issue tracker.

git clone https://github.com/iml130/mlir-emitc.git
cd mlir-emitc
git submodule update --init

There are two variants to build EmitC: As part of an LLVM/MLIR build (via the LLVM external projects mechanism) and against a pre-built LLVM/MLIR.

The setup assumes that you have built LLVM and MLIR in $BUILD_DIR and installed them to $PREFIX. You can use the build_tools/build_mlir.sh shell script to configure, build and install LLVM and MLIR.

Note: The hash of the latest tested LLVM version is given in build_tools/llvm_version.txt. Since MLIR evolves fast, it is possible that EmitC fails to build with a newer LLVM.

To build and launch the tests, run

mkdir build && cd build
cmake -G Ninja -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang   .. -DMLIR_DIR=$PREFIX/lib/cmake/mlir -DLLVM_EXTERNAL_LIT=$BUILD_DIR/bin/llvm-lit
cmake --build . --target check-emitc

Note: If you don't use build_tools/build_mlir.sh, make sure to pass -DLLVM_INSTALL_UTILS=ON when building LLVM with CMake in order to install FileCheck to the chosen installation prefix.

To additionally build and execute the unittests, run

cmake --build . --target MLIREmitCTests
./reference-implementation/unittests/MLIREmitCTests

MLIR-EmitC can also be built as part of an LLVM/MLIR build, using the LLVM_EXTERNAL_PROJECTS mechanism (see https://llvm.org/docs/CMake.html).

To build and launch the tests, run

mkdir build && cd build
cmake -G Ninja -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang   -DCMAKE_BUILD_TYPE=Release -DEMITC_ENABLE_HLO=OFF -DLLVM_ENABLE_PROJECTS=mlir -DLLVM_EXTERNAL_PROJECTS="mlir-emitc" -DLLVM_EXTERNAL_MLIR_EMITC_SOURCE_DIR=`realpath ../` -DLLVM_TARGETS_TO_BUILD=host ${ROOT_PATH_TO_llvm-project}/llvm
cmake --build . --target check-emitc

Conversions are supported for StableHLO ops and some ops of the arith and Tensor dialect. In addition, support for converting Tensor Operator Set Architecture (TOSA) dialect to EmitC is emerging. The emitc-opt tool supports the following options:

The currently supported StableHLO ops are listed in the docs/stablehlo-op-coverage.md document. Supported TOSA ops are listed in the docs/tosa-op-coverage.md document.

After converting to EmitC dialect, C code can be emitted using emitc-translate --mlir-to-cpp. Furthermore, emitc-translate has specific support to emit code with variables declared at top using --mlir-to-cpp --declare-variables-at-top.