NVIDIA / NVIDIA/cutlass

[FEA] Structured sparsity for convolution

Open
#2,303 4 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

feature request inactive-30d inactive-90d
Dominant language
C++
Stars
10.5k
Forks
2.1k
Avg merge
3d 11h
Merged PRs (30d)
7

Description

GEMM currently supports 50% structured sparsity on ampere as in example 15_ampere_sparse_tensorop_gemm. Is there a way to use this GEMM to power a conv2d? I have not been able to find any implementations of sparse tensorop conv2d except for those built in to TensorRT. CuDNN does not support sparse weights, having this feature in cutlass would be very useful.

Contributor guide

No contributing guide indexed for this repository

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Start by studying the 15_ampere_sparse_tensorop_gemm example to understand the existing 50% structured-sparsity support. Compare that capability with the requested conv2d use case and the TensorRT implementations mentioned in the issue. Done means determining and implementing a CUTLASS structured-sparse convolution path, or documenting why the GEMM approach cannot support it.

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
Stale
Clarity
Needs clarification
Newbie friendliness
25/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.