clang - Optimizing cuda to using LLVM

Question

I want to write an LLVM Pass for Cuda codes. I stuck on creating executable from bitcodes.

I used

clang-10 -O0 -S  -emit-llvm --cuda-gpu-arch=sm_61 --cuda-path=/usr/local/cuda-10.0  SAXPY.cu -L/usr/local/cuda-10.0/lib64 -lcudart_static -ldl -lrt -pthread

The comamand generated two files named: SAXPY.ll and SAXPY-cuda-nvptx64-nvidia-cuda-sm_61.ll

Then I used:

opt -load LLVMCUDA.so -CUDA < SAXPY.ll SAXPY-cuda-nvptx64-nvidia-cuda-sm_61.ll >  a.out

After tring to execute a.out, It gave me following error:

/usr/bin/lli-10: error: error creating EE: No available targets are compatible with triple "nvptx64-nvidia-cuda"

My question is, how can I convert from bitcode to executable?

Answer 1

等待大神答复