MLIR build with Dialects

MLIR
1

Is there a flag to pass to cmake when building MLIR to specify which public dialects to build along with the MLIR installation?

Running config with:

cmake  -G Ninja  \
     -DMLIR_ENABLE_BINDINGS_PYTHON=ON \
     -DPYTHON3_EXECUTABLE=$PYTHON_EXECUTABLE \
    -DLLVM_ENABLE_PROJECTS=mlir \
    -DLLVM_BUILD_EXAMPLES=OFF \
    -DLLVM_TARGETS_TO_BUILD="Native;NVPTX;AMDGPU" \
    -DCMAKE_BUILD_TYPE=Release

I can see that the build includes .inc files for each of the public dialects under $build_dir/tools/mlir/include/mlir/Dialect.

Reason I ask is that I am using this current MLIR build as an out-of-tree dependency for a downstream project which requires the GPUOps library as a build dependency:

add_mlir_dialect_library(TritonGPUIR
  Dialect.cpp
  Traits.cpp

  DEPENDS
  TritonGPUTableGen
  TritonGPUAttrDefsIncGen

  LINK_LIBS PUBLIC
  **MLIRGPUOps**
  TritonIR
)

Apologies if this question is ill-posed, as I am still learning MLIR and shaky around how dialects are built / linked.

Thanks!

2

No there isn’t: we’ll build everything. You could script a custom distribution, but in general people are statically linking the libraries they need and that will only pull the minimal amount of things in the final binary.

3

Thanks for the response!

When building the downstream project using the out-of-tree MLIR build per above, I get this error:

CMake Error at /notebooks/CUDA-CPP/MLIR/llvm-project/build_mlir/lib/cmake/llvm/AddLLVM.cmake:556 (add_dependencies):
  The dependency target "MLIRGPUOps" of target "obj.TritonGPUIR" does not
  exist.
Call Stack (most recent call first):
  /notebooks/CUDA-CPP/MLIR/llvm-project/build_mlir/lib/cmake/mlir/AddMLIR.cmake:347 (llvm_add_library)
  /notebooks/CUDA-CPP/MLIR/llvm-project/build_mlir/lib/cmake/mlir/AddMLIR.cmake:626 (add_mlir_library)
  lib/Dialect/TritonGPU/IR/CMakeLists.txt:1 (add_mlir_dialect_library)

How would I build MLIRGPUOps? I can see the inc files generated for each of the public dialects in my MLIR build directory – I’m assuming MLIRGPUOps refers to the public gpu dialect…

4

You’re using an old version of MLIR: [mlir][GPU] Rename MLIRGPUOps CMake target to MLIRGPUDialect · llvm/llvm-project@61223c4 · GitHub

5

Thanks!