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

Reduction To Tridiagonal "1-stage" (local) #1099

Draft
wants to merge 53 commits into
base: master
Choose a base branch
from
Draft

Reduction To Tridiagonal "1-stage" (local) #1099

wants to merge 53 commits into from

Conversation

albestro
Copy link
Collaborator

This effort aims at evaluating a possible improvement for the GPU backend by using a 1-stage reduction instead of the 2-stage currently implemented which seems to fit better on a MC backend.

At the moment of writing:

  • a basic local MC implementation (not refined, just "make it work")
  • a basic test
  • a trivial swap of the algorithm in eigensolver from 2-stage to 1-stage, since there no miniapp has been implemented yet

Next steps:

  • making it work on GPU backend
  • profiling and analysis for identification of main problems for better defining optimization strategies

@albestro albestro added this to the Optimizations milestone Feb 13, 2024
@albestro albestro self-assigned this Feb 13, 2024
@albestro albestro force-pushed the alby/1stage-local branch 2 times, most recently from 51a23e9 to a7e75c7 Compare February 15, 2024 11:34
@albestro
introduce file structure for new algorithm reduction_to_trid
9a4b742
@albestro
extract some utility function in a common place
ffaf34b 
it's a wip: some workaorunds for generalizing them have been introduced
@albestro
WIP: first iteration for panel implementation
5780443
@albestro
move her2k helpers to misc
2716038
@albestro
implement also trailing matrix update
ad9432e
@albestro
fixing and changes for the panel part
d1bf4a7
@albestro
bug fixes
2aae99a
@albestro
minor chage in test
3b72574
@albestro
copy last diag value in tridiagonal matrix
b124d2c 
for the moment treat real and complex cases the same way
@albestro
bug fix last element + add early exit
d71e88c 
left to lapack management of the last tile
@albestro
start refactoring reduction test: factor out common parts
3ed78dd
@albestro
first working test
4324c9f
@albestro
use lapack wrapper for hetrd
daa321e
@albestro
switch from GEMV to GEMM for both real and complex values
d36fd73
@albestro
use transform instead of then (apart from a couple of trailing tasks)
223ec7f
@albestro
remove continuation for last tile (and drop superfluous last element …
ecc2a7f 
…copy)
@albestro
WIP: introduce 1-stage reduction in eigensolver, just for quick testing
77c76a5
@albestro
bug fix: manage also the case of last tile with single element
c64c1b9
@albestro
group w = At * v in big chunks (temporarily) with all tasks writing t…
2d2137a 
…o the same output tile "by-row"
@albestro
extrapolate three main blocks of panel work
0a207af
@albestro
remove superfluous splitTile in trailing matrix update
32c3444
@albestro
start factoring out kernelHEMV
6d13e7b
@albestro
Revert "remove superfluous splitTile in trailing matrix update"
3448de7 
This reverts commit 36f26ab.

It introduced a bug (probably tile lifetime related)
@albestro
add missing include guard
4cadda2
albestro and others added 18 commits March 20, 2024 16:52
@albestro
fix MC build
fd0cab1
@albestro
bug fix
9008dbb
@albestro
basic implementation of last two pieces missing (not yet working on GPU)
e996c60
@albestro
bug fix in compute last tile
11ef707
@albestro
reduce copy volume between cpu and gpu
cf9f342 
just transfer the updated columns
@albestro
start splitting panel update in two tasks
6800a46
@albestro
start bulkerifying updatePanel
161e081 
+ minor optimization skipping update without left panel
@albestro
split updatePanel in updateV and computeReflector
71274e7
@albestro
minor changes
e4c5b92
@albestro
setupW on GPU
5af8486
@albestro
Add cublas gemv batched
b504192
@albestro
trivially migrate to batched for one of the for-loop gemvs
471c735
@albestro
move gpu specialization to a separate file
80676c7
@albestro
add GEMV batched from cublas
a379b30
@albestro
move also updateV to GPU, now just computeReflector is on CPU
cfbea12
@albestro
introduce fixed vectors with pointers for cublas batched
ec8da98
@albestro
switch the other similar gemv in setupW to batched
cbaef73
@albestro
fix format
370ae59
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
Priority:on hold Type:Optimization
Projects
DLA-F Planning
Status: In Progress
Milestone
Optimizations
Development

Successfully merging this pull request may close these issues.

None yet
1 participant
@albestro