cuPoly is not free of bugs (as expected). Following we mention the most relevant ones we are aware of:
- Method simple_scaling_tDivQ(), at src/cuda/cudaengine.cu, is not overflow-safe when t is not a power of 2. The multiplication between v and d_RNStomega_int may result in an overflow, which requires careful handling not done at the current version. Until this is fixed, we inserted a constraint at CUDAEngine::init().
- There is a floating-point division at CUDAEngine::polynomial_basis_ext_Q_compute_v() that may fail when we work close to the 64-bit boundaries of double precision. It's not clear that this will not break correctness in some cases.
- This implementation was conceived as a proof of concept, so there are parts of it that are obviously insecure and make it entirely unfit for production. For instance, at sampler.h we hardcode a SEED. This was necessary during the early days of cuPoly for debugging, but the SEED generation needs to be refactored ASAP.
- There is something wrong in CMakeLists.txt. Sometimes the compiler won't link correctly with the CUDA toolkit, but it works if you call one more time
make. - poly_t::coefs is not needed anymore. It can be removed to reduce memory consumption and initialization time.