This is a project for me to learn numerical linear algebra (solvers, multigrid, sparse matrix, spmv, etc.) and high performance computing. Currently, I am focusing on OpenMP based Shared-Memory Parallelism (SMP). May start implementing GPU based algorithm once I finish OMSCS CS 8803 O21: GPU Hardware and Software.

Algorithms I implemented are mainly from published articles or from other open source libraries. The performance of these algorithms will be tested by using Google Benchmark.

CUDA H100 Deployment

For H100 systems with an older driver than the CUDA toolkit, avoid relying on PTX JIT. For example, CUDA 12.8-generated PTX is not JIT-compatible with an R535 driver even though many CUDA 12.x binaries can run under minor-version compatibility. Build native H100 cubins instead:

cmake --preset release-cuda-h100
cmake --build --preset release-cuda-h100

The preset uses CMAKE_CUDA_ARCHITECTURES=90-real, which emits sm_90 code without PTX fallback. To verify that a build is not relying on PTX JIT on an H100 node, enable tests for a separate build and run the CUDA test subset with PTX JIT disabled:

cmake -S . -B release_h100_tests \
  -DUSE_CUDA=ON \
  -DCMAKE_CUDA_ARCHITECTURES=90-real \
  -DCMAKE_BUILD_TYPE=Release \
  -DENABLE_TESTS=ON \
  -DENABLE_DATA_DOWNLOADS=OFF
cmake --build release_h100_tests -j
CUDA_DISABLE_PTX_JIT=1 ctest --test-dir release_h100_tests -R cuda --output-on-failure

If CUDA_FORCE_PTX_JIT=1 fails on a cluster with an older driver, but CUDA_DISABLE_PTX_JIT=1 passes, the deployment is using native cubins as intended.