NVIDIA / NVIDIA/cutlass

[FEA] BFloat16x2 Atomics

Open
#1,616 3 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

Currently, CUTLASS only implements a specialization of atomic_add for half2, but not nv_bfloat162. This in turn limits BlockStripedReduce to specialize in half2 but not nv_bfloat162.

Is there any reason not to provide a specialization for nv_bfloat162? It looks like a very simple change, but maybe I'm missing something. Thanks in advance for the help!

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 in include/cutlass/functional.h around the existing half2 atomic_add specialization, then inspect include/cutlass/block_striped.h where BlockStripedReduce is limited to half2. Determine whether nv_bfloat162 can receive equivalent support and verify that BlockStripedReduce can use it; done means both paths support nv_bfloat162.

Written by the indexing model from the issue text.

Assessment

Tech stack
cpp
Domain
backend, performance
Issue type
Feature
Difficulty
2/5
Estimated time
1-3 hours
Activity status
Stale
Clarity
Clearly specified
Newbie friendliness
35/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.