This project was called LLIS at the very beginning, and so this name is used in the codebase.
Please refer to the instructions in the sosp23_artifact/
directory.
- Linux (tested on Ubuntu 22.04)
- NVIDIA driver (tested on 535.54.03)
- CUDA runtime (tested on 12.2.0)
- GCC (tested on 11.3.0)
- CMake (tested on 3.22.1)
- Boost (tested on 1.82.0)
- LLVM / Clang (tested on 14)
- spdlog (tested on 1.11.0; 1.12.0 is known to not work)
- tvm-llis (Custom version of TVM modified to work with Paella)
mkdir build
cd build
cmake -DCMAKE_BUILD_TYPE=<release|debug> -DCMAKE_CUDA_ARCHITECTURES=<cuda_arch> .. # cuda_arch is 60 for 6.0, 75 for 7.5, etc
make -j$(nproc) install
Custom TVM depends on the libraries of Paella/LLIS. So, it can only be built after doing the previous step.
Please refer to README-llis.md of tvm-llis for instructions.
Applications and job adapters depend on the custom TVM. So, they can only be built after doing the previous step.
cmake .. -Utvm_FOUND # Find TVM again after we have installed it
make -j$(nproc) install