8000
Skip to content

Phase 1 — SpMM gather-matmul-scatter NKI kernel validated on hardware #14

@scttfrdmn

Description

@scttfrdmn

Roadmap phase tracker

This issue tracks trnsparse's work on Phase 1 of the trnsci roadmap.

See the suite-level roadmap
for the full phase matrix and cross-project dependencies. A reader-friendly
version lives at trnsci.dev/roadmap/.

What this phase means for trnsparse:
Implement the gather-matmul-scatter SpMM kernel: DMA engine gathers
non-zero columns into dense SBUF tiles, Tensor Engine multiplies,
DMA scatters back.

Done means:

  • trnsparse.spmm on NKI validates against the PyTorch reference on uniform-nnz sparse matrices up to 10k × 10k at 1% density.
  • trnsparse.spmv lands as a special case of SpMM with a single-column RHS.
  • set_backend actually routes through nki/dispatch.py (currently all ops force PyTorch regardless — see the existing v0.1.2 caveat in the changelog).
  • Non-uniform-nnz matrices are documented as a Phase 3 item (nnz-bucketing is perf, not correctness).
  • scripts/run_neuron_tests.sh green.

Coordination

  • Label: phase-1-correctness — matches the same label in every sub-project.
  • Cross-project dependencies, if any, are called out in the "Done means"
    list above. Link any PRs / child issues here for tracking.

Close this issue when all "Done means" items are satisfied.

Metadata

Metadata

Assignees

No one assigned

    Labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions

      0