Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Generic dot kernel: use multiple kernels instead of atomics #1244

Closed
maleadt opened this issue Nov 18, 2021 · 0 comments · Fixed by #1245
Closed

Generic dot kernel: use multiple kernels instead of atomics #1244

maleadt opened this issue Nov 18, 2021 · 0 comments · Fixed by #1245
Labels
enhancement New feature or request good first issue Good for newcomers

Comments

@maleadt
Copy link
Member

maleadt commented Nov 18, 2021

Like mapreduce, dot should reduce into a buffer and use multiple steps to reduce inputs that do not fit in a single block, to avoid non-deterministic results. See #1240 (comment).

Explore whether this doesn't degrade performance too much.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Assignees
No one assigned
Labels
enhancement New feature or request good first issue Good for newcomers
Projects
None yet
Milestone
No milestone
Development

Successfully merging a pull request may close this issue.

Don't use nondeterministic atomics in dot when requested. JuliaGPU/CUDA.jl
1 participant
@maleadt