CUDA program to LLVM

Nobody has claimed this yet.

Hi, I was trying to convert the official CUDA Sample [vectorAdd program](https://github.com/NVIDIA/cuda-samples/blob/master/Samples/0_Introduction/vectorAdd/vectorAdd.cu) to mlir and llvm using Polygeist. The Polygeist was built using these commands:

**LLVM-PROJECT BUILD**
`cmake -G Ninja ../llvm \
-DLLVM_ENABLE_PROJECTS="clang;mlir;lld" \
-DLLVM_ENABLE_RUNTIMES="compiler-rt;libcxx;libcxxabi" \
-DLLVM_TARGETS_TO_BUILD="X86;NVPTX" \
-DMLIR_ENABLE_CUDA_RUNNER=ON \
-DMLIR_ENABLE_EXECUTION_ENGINE=ON \
-DLLVM_ENABLE_RTTI=ON \
-DLLVM_ENABLE_ASSERTIONS=ON \
-DCMAKE_BUILD_TYPE=Debug \
-DCMAKE_CUDA_COMPILER=/usr/local/cuda/bin/nvcc \
-DCMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES=/usr/local/cuda/include \
-DCLANG_LINKER_WRAPPER_ENABLED=OFF`

`ninja`

**POLYGEIST BUILD**
`cmake -G Ninja .. \
-DMLIR_DIR=$PWD/../llvm-project/build/lib/cmake/mlir \
-DCLANG_DIR=$PWD/../llvm-project/build/lib/cmake/clang \
-DPOLYGEIST_ENABLE_CUDA=1 \
-DCMAKE_CUDA_COMPILER=/usr/local/cuda/bin/nvcc \
-DCUDA_TOOLKIT_ROOT_DIR=/usr/local/cuda \
-DLLVM_ENABLE_ASSERTIONS=ON \
-DLLVM_ENABLE_RTTI=ON \
-DCMAKE_BUILD_TYPE=Debug \
-DCMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES=/usr/local/cuda/include`

`ninja`

The program converts to mlir without any errors using:
`./cgeist --immediate -cuda-path=/usr/local/cuda -I/usr/local/cuda/include -I/usr/lib/gcc/x86_64-linux-gnu/11/include --resource-dir=/local/Polygeist/llvm-project/build/lib/clang/18 -I/local/cuda-samples/Common --cuda-gpu-arch=sm_86 vectorAdd.cu`
The ouput is [MLIR_trace.txt](https://github.com/user-attachments/files/20614908/MLIR_trace.txt)

But when I directly try to convert the using:
`./cgeist -cuda-path=/usr/local/cuda -I/usr/local/cuda/include -I/usr/lib/gcc/x86_64-linux-gnu/11/include --resource-dir=/local/Polygeist/llvm-project/build/lib/clang/18 -I/local/cuda-samples/Common --cuda-gpu-arch=sm_86 vectorAdd.cu`
The ouput is [LLVM_trace.txt](https://github.com/user-attachments/files/20614894/LLVM_trace.txt)

Are there any steps that I am missing? What can be the issue? Thanks!

Start with the cgeist commands in the issue and compare the attached MLIR_trace.txt and LLVM_trace.txt outputs. Read the CUDA vectorAdd.cu sample and Polygeist's direct conversion entry point to determine why immediate conversion succeeds while direct conversion does not. Done means identifying the missing step or reproducible cause.

Written by the indexing model from the issue text.