Build on Fedora

[stumpOS]

I have been trying to run the python tests in the triton-mlir branch. I followed the instructions from the script here but the build fails on 'TRITON_USE_ASSERT_ENABLED_LLVM=TRUE pip3 install -e '.[tests]''.

There are several errors, all of the type:

/usr/bin/ld: /home/stumpos/.triton/llvm/llvm+mlir-14.0.0-x86_64-linux-gnu-ubuntu-18.04-assert/lib/libMLIRArithmetic.a(ArithmeticOps.cpp.o): relocation R_X86_64_PC32 against undefined hidden symbol `_ZZN4mlir6m_ZeroEvENUlRKN4llvm5APIntEE_4_FUNES3_' can not be used when making a shared object

I noticed that the release of LLVM was for Ubuntu and I'm on Fedora so I downloaded the source for LLVM 14.0.6 and built on my machine. When I redirect the version of llvm to the version I compiled from source, cmake fails on generation:

CMake Error at /home/stumpos/compilers/llvm-project-14.0.6.src/build/lib/cmake/llvm/AddLLVM.cmake:895 (add_executable):
      add_executable cannot create target "FileCheck" because an imported target
      with the same name already exists.
    Call Stack (most recent call first):
      bin/FileCheck/CMakeLists.txt:1 (add_llvm_executable)

Has anyone seen this before?

[stumpOS]

The FileCheck executable is added in bin/FileCheck. I'm not sure where the conflicting executable of the same name is, but maybe LLVM? When I rename the FileCheck to TritonFileCheck the error goes away (but a new error surfaces)

[ptillet]

Yeah, I think building LLVM from source may have caused your LLVM to already have a FileCheck, which conflicts with our copy of it (it isn't included in the package we download). How are you building LLVM from source? I think the options provided should be CMAKE_CONFIGS="-DLLVM_ENABLE_PROJECTS=mlir -DLLVM_TARGETS_TO_BUILD=X86;NVPTX;AMDGPU".

[stumpOS]

The following worked for me:

(compiled with clang 15.0.2)
Step 1: Build llvm 14.0.6 from source. Set the environment variables LLVM_INCLUDE_DIRS, LLVM_LIBRARY_DIR to point to the include and lib. For example, I did the following:

inside /home/stumpos/compilers/llvm-project-14.0.6.src/build:
cmake ../llvm -GNinja -DCMAKE_BUILD_TYPE=Release -DLLVM_ENABLE_PROJECTS="mlir" -DLLVM_TARGETS_TO_BUILD="X86;NVPTX;AMDGPU" -DLLVM_INCLUDE_TESTS=OFF

inside triton's CMake (this is so that when I call setup the environment variables are not overridden by paths that point to an incompatible compiler; not the recommended way to do this obviously)
set(LLVM_INCLUDE_DIRS /home/stumpos/compilers/llvm-project-14.0.6.src/build/include)
set(LLVM_LIBRARY_DIR /home/stumpos/compilers/llvm-project-14.0.6.src/build/lib)

Step 2: (not sure what to turn off to stop the llvm file check from also being added)
rename FileCheck in triton/bin/FileCheck/CMakeLists.txt to TritonFileCheck

Step 3:
pip3 install -e '.[tests]'

[stumpOS]

(repeating this here so I can mark it as answer)
The following worked for me:

(compiled with clang 15.0.2)
Step 1: Build llvm 14.0.6 from source. Set the environment variables LLVM_INCLUDE_DIRS, LLVM_LIBRARY_DIR to point to the include and lib. For example, I did the following:

inside /home/stumpos/compilers/llvm-project-14.0.6.src/build:
cmake ../llvm -GNinja -DCMAKE_BUILD_TYPE=Release -DLLVM_ENABLE_PROJECTS="mlir" -DLLVM_TARGETS_TO_BUILD="X86;NVPTX;AMDGPU" -DLLVM_INCLUDE_TESTS=OFF

inside triton's CMake (this is so that when I call setup the environment variables are not overridden by paths that point to an incompatible compiler; not the recommended way to do this obviously)
set(LLVM_INCLUDE_DIRS /home/stumpos/compilers/llvm-project-14.0.6.src/build/include)
set(LLVM_LIBRARY_DIR /home/stumpos/compilers/llvm-project-14.0.6.src/build/lib)

Step 2: (not sure what to turn off to stop the llvm file check from also being added)
rename FileCheck in triton/bin/FileCheck/CMakeLists.txt to TritonFileCheck

Step 3:
pip3 install -e '.[tests]'