updates to handle pLS with OT hits by slava77 · Pull Request #253 · SegmentLinking/cmssw

slava77 · 2026-04-14T08:34:21Z

most updates are in runTripletDefaultAlgoPPBB and runTripletDefaultAlgoPPEE during pLS matching to the segments of T3. Most of the cuts needed updates: positional and directional z in the barrel (r in endcap), delta beta. The implementation is generalized from the case of pLS ending before the first segment (rightmost case) to other possibilities; near-degenerate cases are pass-through.
- additionally, a bugfix affecting the pre-extension LST: in zPointed window definition, where the z error already estimated at a destination was incorrectly additionally scaled by drOutIn/dSDIn is now corrected (an individual test shows minor impact on the old no-OT pLS test zPointed selection fix in PPBB #216)
PixelSegmentsSoALayout::pLSHitsIdxs are now packed OT/IT<<32 + index to have a fast (and disambiguated) overlap checks
skip duplicate hits in the LSTOutputConverter when adding the OT LST TC hits after the pLS hits were added

The initial submission does not change removeOTRechits to first confirm the default behavior is OK with the CI.

slava77 · 2026-04-14T08:39:42Z

run-ci: [all, hlt]
modifiers: gpu

slava77 · 2026-04-14T08:39:53Z

run-ci: [all, hlt]

slava77 · 2026-04-14T08:52:58Z

@ariostas is it from me incorrectly formatting the command or something else; the error sounds like something else

Unhandled error: HttpError: Invalid Argument - failed to parse workflow: error parsing called workflow
""
-> "SegmentLinking/TrackLooper-actions/.github/workflows/hlt.yml@devel"
: failed to fetch workflow: reference to workflow should be either a valid branch, tag, or commit

github-actions · 2026-04-14T09:01:59Z

The PR was built and ran successfully in standalone mode running on GPU. 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     32.4      0.2      0.4      0.5      0.9      0.3      0.6      0.4      0.3      1.3      0.0      37.3       4.6+/-  2.6      37.3   explicit[s=1]
   avg      1.2      0.3      0.5      0.7      1.1      0.3      0.8      0.5      0.4      1.7      0.0       7.6       6.1+/-  2.7       3.9   explicit[s=2]
   avg      2.9      0.6      0.8      1.1      1.5      0.4      1.2      0.7      0.6      2.8      0.0      12.5       9.2+/-  3.3       3.3   explicit[s=4]
   avg      4.8      0.9      1.2      1.7      2.0      0.5      1.7      1.0      0.8      3.8      0.0      18.5      13.2+/-  3.9       3.2   explicit[s=6]
   avg      4.6      1.4      1.7      2.3      2.6      0.6      2.2      1.3      1.0      4.9      0.0      22.7      17.5+/-  4.2       2.9   explicit[s=8] (target branch)
   avg     31.5      0.2      0.4      0.5      0.8      0.3      0.6      0.4      0.3      1.3      0.0      36.4       4.6+/-  2.4      36.4   explicit[s=1]
   avg      1.1      0.3      0.5      0.7      1.0      0.3      0.8      0.6      0.4      1.8      0.0       7.5       6.1+/-  2.8       3.9   explicit[s=2]
   avg      2.0      0.6      0.8      1.1      1.5      0.4      1.2      0.7      0.6      2.8      0.0      11.6       9.2+/-  3.2       3.0   explicit[s=4]
   avg      3.0      0.9      1.3      1.6      2.0      0.5      1.7      1.0      0.8      3.9      0.0      16.8      13.2+/-  3.9       2.9   explicit[s=6]
   avg      3.4      1.3      1.8      2.3      2.6      0.6      2.3      1.2      1.0      4.9      0.0      21.5      17.5+/-  4.2       2.8   explicit[s=8] (this PR)

slava77 · 2026-04-14T09:02:03Z

run-ci: hlt

github-actions · 2026-04-14T09:04:37Z

The PR was built and ran successfully in standalone mode running on CPU. 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     30.0     79.8    114.8    121.7     53.3    682.0     10.9     41.8     73.7    215.9      0.1    1424.0     712.0+/- 186.8     483.3   explicit[s=4] (target branch)
   avg     30.0     80.4    115.6    121.4     52.7    676.7     10.8     41.1     69.1    201.2      0.1    1399.0     692.4+/- 184.5     474.3   explicit[s=4] (this PR)

github-actions · 2026-04-14T10:02:48Z

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

OOTB All Tracks

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

github-actions · 2026-04-14T10:13:09Z

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

OOTB All Tracks

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

slava77 · 2026-04-14T12:42:16Z

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

dups by category: this PR

vs "reference" from #203

CMSSW plots have no differences.
I will try to figure out the origin

slava77 · 2026-04-14T13:34:43Z

      bool pmatched = false;
-      if (phits1[i] == -1)
+      // short pLSs are padded with the last hit idx
+      if (i + 1 == Params_pLS::kHits && phits1[i] == phits1[i - 1])


note to self:
here the hits are not in the same order. So, 3-hit pLS will have the same hit at i=1 and i=3.
OTOH, hit indices are always filled with something non-default. So, this check here if corrected is only to save a cycle of checks with j

slava77 · 2026-04-16T16:27:29Z

run-ci: [all, hlt]

github-actions · 2026-04-16T16:35:19Z

There was a problem while building and running in standalone mode. The logs can be found here.

github-actions · 2026-04-16T16:36:32Z

There was a problem while building and running with CMSSW. The logs can be found here.

slava77 · 2026-04-16T16:47:13Z

run-ci: [all, hlt]

github-actions · 2026-04-16T17:09:20Z

The PR was built and ran successfully in standalone mode running on CPU. 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     21.3     67.9    100.6     94.0     35.3    499.3      8.0     35.6     65.2    171.8      0.1    1099.1     578.5+/- 148.2     370.7   explicit[s=4] (target branch)
   avg     21.4     67.5    100.2     94.1     33.9    450.4      8.1     32.2     60.9    168.3      0.6    1037.7     565.8+/- 144.9     379.9   explicit[s=4] (this PR)

github-actions · 2026-04-16T18:05:27Z

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

OOTB All Tracks

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

github-actions · 2026-04-16T18:16:34Z

The PR was built and ran successfully with HLT setup running on CPU (procModifiers = ). Here are some plots.

HLT General Plots

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

slava77 · 2026-04-16T21:26:00Z

run-ci: [all, hlt]
modifiers: gpu

github-actions · 2026-04-16T21:47:31Z

The PR was built and ran successfully in standalone mode running on GPU. 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     31.1      0.2      0.4      0.5      0.8      0.3      0.6      0.4      0.3      1.3      0.0      36.0       4.6+/-  2.5      36.0   explicit[s=1]
   avg      1.1      0.3      0.5      0.7      1.0      0.3      0.8      0.5      0.4      1.7      0.0       7.4       6.0+/-  2.8       3.8   explicit[s=2]
   avg      1.8      0.6      0.8      1.1      1.4      0.4      1.2      0.7      0.6      2.7      0.0      11.4       9.2+/-  3.3       3.0   explicit[s=4]
   avg      2.6      0.9      1.3      1.6      2.0      0.5      1.7      1.0      0.8      3.8      0.0      16.2      13.1+/-  4.1       2.8   explicit[s=6]
   avg      3.6      1.4      1.7      2.3      2.7      0.6      2.2      1.2      1.1      4.8      0.0      21.6      17.4+/-  4.3       2.8   explicit[s=8] (target branch)
   avg     31.3      0.2      0.4      0.5      0.9      0.3      0.6      0.4      0.3      1.3      0.0      36.1       4.6+/-  2.5      36.2   explicit[s=1]
   avg      1.1      0.3      0.5      0.7      1.0      0.3      0.8      0.6      0.4      1.7      0.0       7.4       6.0+/-  2.7       3.8   explicit[s=2]
   avg      2.2      0.6      0.8      1.1      1.5      0.4      1.3      0.8      0.6      2.7      0.0      11.9       9.3+/-  3.3       3.1   explicit[s=4]
   avg      2.8      0.9      1.3      1.6      2.0      0.5      1.6      1.0      0.8      3.8      0.0      16.4      13.1+/-  3.9       2.9   explicit[s=6]
   avg      3.9      1.4      1.8      2.3      2.7      0.7      2.2      1.2      1.1      4.9      0.0      22.3      17.7+/-  4.3       2.9   explicit[s=8] (this PR)

slava77 · 2026-04-16T21:59:12Z

run-ci: [all, hlt]

slava77 · 2026-04-16T22:01:13Z

run-ci: [all, hlt]

I expect still no changes in the offline setup but some minor increase in the fake rate in the HLT setup (based on the local tests, mostly localized in the shorter tracks)

slava77 · 2026-04-16T22:05:18Z

@slava77 slava77 force-pushed the CMSSW_17_0_X-lst015-caExt-PPBBEE branch from 565d599 to 0c2097e 5 hours ago

I had a silly typo (1 vs 2) mixup. I tried to subscribe to the pro account, but with that I was suggested to prove a negative in line with "upload an image explaining why you are not on campus"

slava77 · 2026-04-16T22:08:24Z

earlier I thought I was just making a random comment

@ariostas
do you know why this apparently went through a parsing hook? (with the thumbs down) ... since no action was actually expected it's probably OK to ignore the negative reaction

github-actions · 2026-04-16T22:18:38Z

The PR was built and ran successfully in standalone mode running on CPU. 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     29.4     79.1    114.0    117.3     51.5    687.8     10.7     40.2     71.0    215.0      0.1    1416.2     699.0+/- 185.6     473.5   explicit[s=4] (target branch)
   avg     29.5     78.4    113.3    117.1     51.6    688.0     10.9     39.9     66.3    196.9      0.1    1391.9     674.4+/- 177.8     473.1   explicit[s=4] (this PR)

github-actions · 2026-04-16T22:52:43Z

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

OOTB All Tracks

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

github-actions · 2026-04-16T23:20:08Z

The PR was built and ran successfully with HLT setup running on CPU (procModifiers = ). Here are some plots.

HLT General Plots

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

github-actions · 2026-04-16T23:21:56Z

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

OOTB All Tracks

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

github-actions · 2026-04-16T23:47:25Z

The PR was built and ran successfully with HLT setup running on GPU (procModifiers = ). Here are some plots.

HLT General Plots

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

VourMa

The changes look good to me, I have some comments mostly for my understanding.

More in general, I am not sure what we want to do with disabling removeOTRechits. The performance looks to me worse in more or less all fronts. Is there any obvious advantage? Should this be introduced as a means to study how to improve but not enable it by default?

slava77 · 2026-04-22T18:03:56Z

More in general, I am not sure what we want to do with disabling removeOTRechits. The performance looks to me worse in more or less all fronts. Is there any obvious advantage? Should this be introduced as a means to study how to improve but not enable it by default?

https://indico.cern.ch/event/1640803/#16-contribution p7, 8 may help to tilt the discussion in the direction of disabling removeOTRechits: the LST track candidates are longer (3-hit entries are gone and the 4-hit component is reduced quite a bit), the duplicate rate is x2 lower.

For the fakes it looks like the first order effect for the final tracks is the hltInitialStepTrackCandidates.candMinNHitsCut is not doing anything anymore. Additionally, at the LST candidate/track level even though the fake rate seems roughly flat vs pt and eta; it still looks from the nHits distributions that there is an increase in shorter tracks (unlinked pLS) and even some reduction for the longer tracks (pLS+LST). So, it may be a bit awkward that the solution to reduce fakes in the CA extension is to truncate the OT part and redo it with LST or mkFit; this part should be done/cleaned in CA ext or later in the full track quality selection. [and now we apparently weigh this above getting cleaner LST tracks]

…xtended pLS stopping in barrel and endcap OT: - fix error computations and dBetas - account for radius uncertainties in lumi-based and pointed z selection, most important for connections with pLS on the same tilted layer - use the tl-axis length in dBetaRes term, it becomes large in the near-3-point cases - rename a few more variables for clarity - fixup to the dzDrtScale use in PPBB - PPEE: fix dLum and dzDrtScale in rt{Lo,Hi}, fix dz sign in rt{Lo,Hi}_point

…to have a fast (and disambiguated) overlap checks; simplify pLS overlap check

…nge in pLS-LS matching

…tBits (kMaxPLSHitBitsInHitsSoA = 8) to open room for more hits

…n OT and was not extended by LST (at most 3 IT hits)

slava77 · 2026-06-10T14:24:15Z

run-ci: [all, hlt]
release: CMSSW_17_0_X_2026-06-07-0000

github-actions · 2026-06-10T14:50:11Z

The PR was built and ran successfully in standalone mode running on CPU. 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     26.9    165.2    127.3    113.5     66.3    642.8     10.4     47.2     71.4    212.2      0.1    1483.2     813.5+/- 201.5     489.4   explicit[s=4] (target branch)
   avg     26.6    164.5    125.3    117.3     63.7    651.2     10.7     44.7     67.0    198.6      0.1    1469.9     792.1+/- 194.7     483.4   explicit[s=4] (this PR)

github-actions · 2026-06-10T18:19:15Z

The PR was built and ran successfully with HLT setup running on CPU (procModifiers = ). Here are some plots.

HLT General Plots

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

slava77 · 2026-06-11T11:36:25Z

this is kind of low

Here's what I had in 1610pre3 (blue v black):

slava77 · 2026-06-11T15:06:25Z

I was able to reproduce in 17_0_0_pre2

Fake and dups in 16_1_0_pre3:

vs 17_0_0_pre2

the change is most visible in fakes with 4 layers (3 pixel layers)

There is as much as 40% fractional decrease in the fake rate with about 1% drop in the efficiency around $$|\eta| \sim 1$$
I'm not sure it's a good tradeoff

slava77 · 2026-06-11T15:26:31Z

the change is most visible in fakes with 4 layers (3 pixel layers)

These 3IT+1OT hit fakes are also possibly migrating to 3IT+3OT.
I can guess it's the 3-IT-hit pLS. But then the logic for these is supposedly unchanged: they previously didn't have OT hits on output, now they have OT hits in the LST input but they are dropped in the output.
So, it looks like I'm doing this drop of OT hits differently in 16_1_0_pre3 vs 17_1_0_pre2
@VourMa please have a look, in case I missed something in the LSTOutputProducer perhaps in part connected with #252

VourMa

Quick review of the logic for removing OT hits from seeds.

VourMa · 2026-06-12T10:00:14Z

+    const auto iSeed = lstOutput_view.pixelSeedIndex()[i];
+    const bool dropHitsOTpL =
+        iType == lst::LSTObjType::pLS && dropOTHitsPurePLS_ &&
+        std::prev(pixelSeeds[iSeed].recHits().end())->geographicalId().subdetId() > PixelSubdetector::PixelEndcap &&


Are the recHits ordered from IT -> OT? Is this condition even needed?

VourMa · 2026-06-12T10:10:48Z

@@ -110,6 +116,8 @@ void LSTOutputConverter::fillDescriptions(edm::ConfigurationDescriptions& descri
  desc.add<edm::InputTag>("lstPixelSeeds", edm::InputTag("lstInputProducer"));
  desc.add<bool>("includeT5s", true);
  desc.add<bool>("includeNonpLSTSs", false);
+  desc.add<bool>("dropOTHitsPurePLS", false);
+  desc.add<int>("maxITHitsToDropOTHitsPurePLS", 3);


How was this decided?

VourMa · 2026-06-12T10:19:02Z

+          for (auto const& hit : recHits)
+            if (hit.sharesInput(OTHits[hitIdx], TrackingRecHit::all)) {
+              hitOK = false;
+              break;
+            }


Could you remind me what cases is this supposed to catch?

VourMa · 2026-06-12T11:11:36Z

+      if (dropHitsOTpL) {        // true if need to drop OT hits
+        using Hit = SeedingHitSet::ConstRecHitPointer;
+        std::vector<Hit> hitsForSeed;
+        hitsForSeed.reserve(std::ranges::size(pixelSeeds[iSeed].recHits()));
+        for (auto const& hit : pixelSeeds[iSeed].recHits()) {
+          if (hit.geographicalId().subdetId() > PixelSubdetector::PixelEndcap)
+            continue;
+          hitsForSeed.emplace_back(dynamic_cast<Hit>(&hit));
+          recHits.push_back(hit.clone());
+        }
+        GlobalTrackingRegion region;
+        seedCreator_->init(region, iSetup, nullptr);
+        seedCreator_->makeSeed(seeds, hitsForSeed);
+        if (seeds.empty())
+          edm::LogInfo("LSTOutputConverter") << "failed to convert a pLS object to a seed" << i << iSeed;
+        seed = seeds[0];
+        seedRef = edm::RefToBase<TrajectorySeed>(edm::Ref(outputTSRP, outputTS.size()));
      }


This is done only when produceSeeds_. Don't we want it also when we produce tracks?

slava77 mentioned this pull request Apr 14, 2026

place CI results in less ambiguously named directories SegmentLinking/TrackLooper-actions#33

Open

slava77 commented Apr 14, 2026

View reviewed changes

slava77 force-pushed the CMSSW_17_0_X-lst015-caExt-PPBBEE branch from 55057df to 565d599 Compare April 16, 2026 16:26

slava77 force-pushed the CMSSW_17_0_X-lst015-caExt-PPBBEE branch from 565d599 to 0c2097e Compare April 16, 2026 16:47

slava77 requested a review from VourMa April 17, 2026 06:35

VourMa reviewed Apr 21, 2026

View reviewed changes

Comment thread RecoTracker/LSTCore/src/alpaka/PixelTriplet.h Outdated

Comment thread RecoTracker/LSTCore/src/alpaka/PixelTriplet.h

Comment thread RecoTracker/LSTCore/src/alpaka/PixelTriplet.h

Comment thread RecoTracker/LSTCore/src/alpaka/PixelTriplet.h

slava77devel added 8 commits June 10, 2026 05:44

fix z luminous region cut for general rIn/rOut relationship

78bc043

PixelSegmentsSoALayout::pLSHitsIdxs are now packed OT/IT<<32 + index …

0d234e8

…to have a fast (and disambiguated) overlap checks; simplify pLS overlap check

skip duplicate hits

de615eb

load the last pLS hit to the 4th entry; use it in the hit position ra…

5c7738a

…nge in pLS-LS matching

moved to pixelSeeds firstHit, nHits (kMaxPLSHitsInHitsSoA = 4), hitDe…

12314fe

…tBits (kMaxPLSHitBitsInHitsSoA = 8) to open room for more hits

second-guess CA extension and use dropOTHitsPurePLS if the pLS ends i…

5e6353c

…n OT and was not extended by LST (at most 3 IT hits)

now do not removeOTRechits

2d7eb5f

slava77 force-pushed the CMSSW_17_0_X-lst015-caExt-PPBBEE branch from 6895f88 to 2d7eb5f Compare June 10, 2026 14:19

VourMa reviewed Jun 12, 2026

View reviewed changes

Conversation

slava77 commented Apr 14, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

slava77 commented Apr 14, 2026

Uh oh!

slava77 commented Apr 14, 2026

Uh oh!

slava77 commented Apr 14, 2026

Uh oh!

github-actions Bot commented Apr 14, 2026

Uh oh!

slava77 commented Apr 14, 2026

Uh oh!

github-actions Bot commented Apr 14, 2026

Uh oh!

github-actions Bot commented Apr 14, 2026

Uh oh!

github-actions Bot commented Apr 14, 2026

Uh oh!

slava77 commented Apr 14, 2026

Uh oh!

slava77 Apr 14, 2026

Choose a reason for hiding this comment

Uh oh!

slava77 commented Apr 16, 2026

Uh oh!

github-actions Bot commented Apr 16, 2026

Uh oh!

github-actions Bot commented Apr 16, 2026

Uh oh!

slava77 commented Apr 16, 2026

Uh oh!

github-actions Bot commented Apr 16, 2026

Uh oh!

github-actions Bot commented Apr 16, 2026

Uh oh!

github-actions Bot commented Apr 16, 2026

Uh oh!

slava77 commented Apr 16, 2026

Uh oh!

github-actions Bot commented Apr 16, 2026

Uh oh!

slava77 commented Apr 16, 2026

Uh oh!

slava77 commented Apr 16, 2026

Uh oh!

slava77 commented Apr 16, 2026

Uh oh!

slava77 commented Apr 16, 2026

Uh oh!

github-actions Bot commented Apr 16, 2026

Uh oh!

github-actions Bot commented Apr 16, 2026

Uh oh!

github-actions Bot commented Apr 16, 2026

Uh oh!

github-actions Bot commented Apr 16, 2026

Uh oh!

github-actions Bot commented Apr 16, 2026

Uh oh!

VourMa left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

slava77 commented Apr 22, 2026

Uh oh!

slava77 commented Jun 10, 2026

Uh oh!

github-actions Bot commented Jun 10, 2026

Uh oh!

github-actions Bot commented Jun 10, 2026

Uh oh!

slava77 commented Jun 11, 2026

Uh oh!

slava77 commented Jun 11, 2026

Uh oh!

slava77 commented Apr 14, 2026 •

edited

Loading

slava77 commented Jun 11, 2026 •

edited

Loading