Research CommonsResearch Commons
cpptensor/Roadmap

Roadmap

Where cpptensor is headed — deeper CUDA support, broader op coverage, and platform reach.

cpptensor is beta. The CPU path is the validated, primary workflow; other areas are progressing. For the live state, follow the repository.

Solid today

  • N-dimensional tensors with views and contiguous materialization.
  • Runtime ISA dispatch (generic / AVX2 / AVX-512) with safe fallback.
  • OpenBLAS-backed matmul / dot, plus svd / eig.
  • Numerically careful reductions and dot.
  • Versioned checkpoint save/load.
  • Catch2 tests and Google Benchmark targets.

In progress / exploratory

  • CUDA backend. Currently early and off by default (BUILD_CUDA=OFF); deepening GPU coverage. See Backends & ISA.
  • std::simd backend. Declared in CMake but not yet wired as an active toggle.
  • Broader platform support. Windows is experimental and outside the documented conda workflow today.
  • Expanding the op catalog. Operator status and semantics are tracked in docs/Ops_Status.md.

Contributing

The repo documents how to add new ops and kernels in docs/HowToAddNewOpsWithKernel.md. Builds run a strict warning/sanitizer gate in CI — see Build options.