Follow up to using cuTile in KMeans
Open
Nobody has claimed this yet.
- Dominant language
- Cuda
- Stars
- 854
- Forks
- 236
- Avg merge
- 3d 3h
- Merged PRs (30d)
- 62
Description
This issue captures follow-up work created in https://github.com/NVIDIA/cuvs/pull/2249.
Immediate follow ups:
- Add benchmarking scripts that aim to identify regressions in newer
tileirasversions and sweep tile shapes, as well as exercise the sweep for new architectures
CUDA 13.5 tasks:
- Update tile shapes to known improvements:
FP32:
| SM | Strict | Relaxed |
|---|---|---|
| 80 | 128x128x16, occ=2 | 128x128x16 |
| 86 | 64x256x16, occ=2 | 128x128x16 |
| 90 | 64x256x32 | 128x128x64 |
| 100 | 128x128x128 | 128x128x128 |
| 120 | 64x128x32, occ=2 | 64x256x32, occ=2 |
FP16:
| SM | Strict | Relaxed |
|---|---|---|
| 80 | 128x128x128, occ=2 | 128x64x128, occ=2 |
| 86 | 128x128x32, occ=2 | 128x128x32, occ=2 |
| 90 | 128x128x64 | 64x128x128 |
| 100 | 128x128x128, occ=2 | 128x256x16, occ=2 |
| 120 | 64x256x64, occ=2 | 64x256x128, occ=2 |
- Check if we can move to C++ Tile
Contributor guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Research direction
Start with the cuTile KMeans work introduced in PR 2249 and identify the existing tile-shape configuration and benchmarking entry points. Add the requested regression and tile-shape sweeps, update shapes for the listed CUDA 13.5 architectures, and document whether moving to C++ Tile is feasible.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- cpp
- Domain
- machine-learning, performance
- Issue type
- Feature
- Difficulty
- 5/5
- Estimated time
- Over a week
- Activity status
- Quiet
- Clarity
- Mostly clear
- Newbie friendliness
- 35/100