Skip to content

CA-Red2Band: Communication-Avoiding Reduction To Band #1177

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
Draft
wants to merge 30 commits into
base: master
Choose a base branch
from
Draft

CA-Red2Band: Communication-Avoiding Reduction To Band #1177

wants to merge 30 commits into from

Conversation

albestro
Copy link
Collaborator

@albestro albestro commented Aug 16, 2024 •
edited
Loading

First basic implementation without all required bells and whistles.

  • It is not fully optimised
    • it does more communications than needed
    • it performs superfluous computation, e.g. it applies HH even when they are identities by construction
  • It does not manage edge-cases (e.g. incomplete tiles)
  • It does not work sub-band
  • It does not have a GPU backend

Changelog:

  • Some old reduction to band functions have been moved to a common place
  • Some old reduction to band functions (and related, e.g. t-factor) have been fixed to deal with distributed matrices where we want to work just on the local part

@albestro albestro added this to the Optimizations milestone Aug 16, 2024
@albestro albestro self-assigned this Aug 16, 2024
@albestro
Copy link
Collaborator Author

cscs-ci run

1 similar comment
@albestro
Copy link
Collaborator Author

cscs-ci run

@albestro
file and folder structure for new algorithm
3725b84
@albestro
factor out red2band functions so they can be reused in new implementa…
df2f1de 
…tion variant
@albestro
basic test structure
b9ce996
@albestro
WIP: qr for both 1st and 2nd passes
52971b2
@albestro
move some basic building block that can be shared between
7be4199 
implementations
@albestro
first implementation of trailing matrix update (1st pass)
12015bd
@albestro
fixes for qr 2nd pass
255b5e1
@albestro
implementation 2nd pass update trailing + fixes
382bee6
@albestro
switch to local tfactor for 2nd pass + minor changes
e4647f0
@albestro
change mat_hh_2nd distribution to an equivalent one
9c6a417 
it is easier to manage on the test side and it does not change much on
the algorithmic side
@albestro
adapt computePanelReflectors to deal with a tile for taus
a76d9eb
@albestro
fix distribution for both taus_1st and taus_2nd
a2d5638
@albestro
update test structure
5c146f8
@albestro
minor changes
ad6ade0
@albestro
CHECK: broadcast panel now communicate also last tile
b1b6185
@albestro
hh2 fix panel size problem for reflectors
8f8b199
@albestro
hh2 fix W1 panel, which has a different size wrt others
3883430
@albestro
fix deadlock: potentially unused tile
7839564
@albestro
hh1: fix her2k transposed indices usage
e26812b
@albestro
hh2: fix her2k computation LR + R
9d8d648
@albestro
t-factor minor changes
155fd59
@albestro
fix t-factor for local usage on distributed matrix
7c3dc60
@albestro
fix her2k HH1 + remove console output
2b81a93
@albestro
basic test implementation
41fc7ab
@msimberg msimberg moved this from In Progress to Todo in DLA-F Planning Sep 11, 2025
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

@albestro albestro

Labels

Type:Optimization

Projects

DLA-F Planning
Status: Todo

Milestone

Optimizations

Development

Successfully merging this pull request may close these issues.

1 participant

@albestro