CuPBoP is a framework which support executing unmodified CUDA source code on non-NVIDIA devices. Currently, CuPBoP support serveral CPU backends, including x86, AArch64, and RISC-V. Supporting Vortex (a RISC-V GPU) is working in progress.

• Linux system
• LLVM 14.0.1
• CUDA Toolkit

Although CuPBoP does not require NVIDIA GPUs, it needs CUDA to compile the source programs to NVVM/LLVM IRs. CUDA toolkit can be built on machines without NVIDIA GPUs. For building CUDA toolkit, please refer to https://developer.nvidia.com/cuda-downloads.

1. Clone from github

   git clone --recursive https://github.com/cupbop/CuPBoP
   cd CuPBoP
   export CuPBoP_PATH=`pwd`
   export LD_LIBRARY_PATH=$CuPBoP_PATH/build/runtime:$CuPBoP_PATH/build/runtime/threadPool:$LD_LIBRARY_PATH
   export CUDA_PATH=/usr/local/cuda-11.7 # set to your own location

2. Build CuPBoP

   mkdir build && cd build
   #set -DDEBUG=ON for debugging
   cmake .. \
      -DLLVM_CONFIG_PATH=`which llvm-config` \
      -DCUDA_PATH=$CUDA_PATH
   make

3. (Optional) Use CuPBoP to execute Hetero-mark benchmark for verification

   make test

In this section, we provide an example of how to use CuPBoP to execute a CUDA program.

cd examples/vecadd
# Compile CUDA source code (both host and kernel) to bitcode files
clang++ -std=c++11 vecadd.cu \
      -I../.. --cuda-path=$CUDA_PATH \
      --cuda-gpu-arch=sm_50 -L$CUDA_PATH/lib64 \
      -lcudart_static -ldl -lrt -pthread -save-temps -v  || true
# Apply compilation transformations on the kernel bitcode file
$CuPBoP_PATH/build/compilation/kernelTranslator \
      vecadd-cuda-nvptx64-nvidia-cuda-sm_50.bc kernel.bc
# Apply compilation transformations on the host bitcode file
$CuPBoP_PATH/build/compilation/hostTranslator \
      vecadd-host-x86_64-unknown-linux-gnu.bc host.bc
# Generate object files
llc --relocation-model=pic --filetype=obj  kernel.bc
llc --relocation-model=pic --filetype=obj  host.bc
# Link with runtime libraries and generate the executable file
g++ -o vecadd -fPIC -no-pie \
      -I$CuPBoP_PATH/runtime/threadPool/include \
      -L$CuPBoP_PATH/build/runtime  \
      -L$CuPBoP_PATH/build/runtime/threadPool \
      host.o kernel.o \
      -I../.. -lc -lCPUruntime -lthreadPool -lpthread
# Execute
./vecadd

Any kinds of contributions are welcome. Please refer to Contribution.md for more detail.

If you want to refer CuPBoP in your projects, please cite the related papers:

• COX: Exposing CUDA Warp-Level Functions to CPUs
• CuPBoP: CUDA for Parallelized and Broad-range Processors

• Ruobing Han
• Jun Chen
• Bhanu Garg
• Xule Zhou
• John Lu
• Chihyo Ahn
• Haotian Sheng
• Blaise Tine
• Hyesoon Kim

• POCL is an open-source OpenCL implementations that based on LLVM. We reuse some code from it (e.g., apply optimizations, load/store LLVM IRs).
• Hetero-Mark and Rodinia Benchmark are two benchmark suites for heterogeneous system computation. CuPBoP uses them as integrated test to verify the correctness.
• moodycamel::ConcurrentQueue is a fast multi-producer, multi-consumer lock-free concurrent queue for C++11. CuPBoP uses it as the task queue for launching and executing kernels.