Remove Redundant Functions + 60% MD Speedup

[GNiendorf]

Below is run on the L40.

This PR Timing:

Master Timing:

MD creation time goes from 1.0ms to 0.4ms, meaning that the redundant matrix creation mentioned below represented the majority of the MD creation time on GPU... 😱 @slava77

[GNiendorf]

/run standalone

[GNiendorf]

This matrix creation code was being run many times, representing a significant portion of the MD creation time it seems like.

[github-actions]

The PR was built and ran successfully in standalone mode. Here are some of the comparison plots.

The full set of validation and comparison plots can be found here.

Here is a timing comparison:

   Evt    Hits       MD       LS      T3       T5       pLS       pT5      pT3      TC       Reset    Event     Short             Rate
   avg     44.5    326.0    116.8     47.7     93.5    500.2    113.4    150.7    101.4      2.3    1496.7     951.9+/- 247.8     410.9   explicit_cache[s=4] (target branch)
   avg     48.1    324.0    116.0     46.6     92.5    509.4    114.3    150.4    101.5      2.5    1505.4     947.8+/- 246.7     409.2   explicit_cache[s=4] (this PR)

[slava77]

MD creation time goes from 1.0ms to 0.4ms, meaning that the redundant matrix creation mentioned below represented the majority of the MD creation time on GPU

nice find/fix.
Thank you.

[slava77]

apparently this is addressing SegmentLinking/TrackLooper#303

[VourMa]

Indeed, nice catch and clean up. I would even considering adding this to the CMSSW PR (especially if the renaming to adhere to CMSSW naming is real and not an artefact of the rebases). What do you think?

@ariostas Is the _devel branch fully up to date with the branch for the CMSSW PR?

[ariostas]

The _devel branch is not up to date since there are still some big changes coming. I actually just noticed that the _devel PR is a bit of a mess, so we'll have to figure that out at some point. But I agree that this should go into the _realfiles branch. This is great!

I think there's other PRs by Gavin that should also be included in _realfiles since there have been review comments about them. In particular, #39 and #54.

[VourMa]

I think there's other PRs by Gavin that should also be included in _realfiles since there have been review comments about them. In particular, #39 and #54.

I would be less eager to also push these ones, because:

1. They introduce a new (physics) feature (lower pT cut) which is not strictly needed at this stage.
2. They require new data files, with all the complication this brings to the cms-sw way of running things for PRs.
3. They define a clear-cut development, separated from the rest (up to a good degree).

So I think they would be great for a standalone PR.

[ariostas]

I would be less eager to also push these ones, because...

That's true. But Andrea has asked for occupancy tables instead of a bunch of if-else statements, and if we want to fix SDL_INF we'll have to update the data files anyway, so maybe we could just get that out of the way

[slava77]

/run all

[github-actions]

The PR was built and ran successfully in standalone mode. Here are some of the comparison plots.

The full set of validation and comparison plots can be found here.

Here is a timing comparison:

   Evt    Hits       MD       LS      T3       T5       pLS       pT5      pT3      TC       Reset    Event     Short             Rate
   avg     44.1    325.7    115.6     45.1     91.0    502.0    113.6    149.4    100.7      2.2    1489.5     943.4+/- 247.6     406.8   explicit_cache[s=4] (target branch)
   avg     47.6    324.3    116.0     45.6     91.8    511.2    112.9    150.0    100.8      2.4    1502.5     943.8+/- 245.0     411.4   explicit_cache[s=4] (this PR)

[slava77]

linter check isn't happy here either

[github-actions]

The PR was built and ran successfully with CMSSW. Here are some plots.

OOTB All Tracks

The full set of validation and comparison plots can be found here.

[GNiendorf]

@ariostas When I run the code format I see a bunch of suggestions for files not related to this PR:

[slava77]

@ariostas When I run the code format I see a bunch of suggestions for files not related to this PR:

did you update to the new release or still reused the old release area with a new cmsrel?

[GNiendorf]

Oh no, I haven't updated. Let me do that and try again.

[slava77]

/run all

[github-actions]

The PR was built and ran successfully in standalone mode. Here are some of the comparison plots.

The full set of validation and comparison plots can be found here.

Here is a timing comparison:

   Evt    Hits       MD       LS      T3       T5       pLS       pT5      pT3      TC       Reset    Event     Short             Rate
   avg     45.0    326.4    116.0     47.9     92.9    501.9    113.2    150.3    100.6      2.5    1496.7     949.8+/- 244.3     410.6   explicit_cache[s=4] (target branch)
   avg     47.2    323.1    115.9     46.9     91.5    506.0    113.4    148.9    101.7      2.3    1496.9     943.6+/- 249.2     410.0   explicit_cache[s=4] (this PR)

[ariostas]

When I run the code format I see a bunch of suggestions for files not related to this PR:

I have the same issue. I'm not sure how to get around it. I think they didn't format all files after clang-format got updated

[github-actions]

The PR was built and ran successfully with CMSSW. Here are some plots.

OOTB All Tracks

The full set of validation and comparison plots can be found here.

[slava77]

the linter complaints are for parts of the files not touched by this PR

[slava77]

Here is a timing comparison:

there is practically no change in the CPU backend: MD changes from 326.4 to 323.1; the other run was faster by around 1 s.
I guess the optimization works out differently (or something else is very slow)