Second round of comments - part II

Author: AdrianoDee

DataFormats/TrackSoA/interface/TracksSoA.h

@@ -54,13 +54,13 @@ namespace reco {
     auto end = tracks[i].hitOffsets();
     auto hitId = hits[start].id();
     int nl = 1;
-    auto ol = 0;
+    int ol = 0;
     while (hitId >= layerStarts[ol + 1] and ol < maxLayers)
       ++ol;
     ++start;
     for (; start < end; ++start) {
       hitId = hits[start].id();
-      auto il = 0;
+      int il = 0;
       while (hitId >= layerStarts[il + 1] and il < maxLayers)
         ++il;
       if (il != ol)

DataFormats/TrackingRecHitSoA/interface/TrackingRecHitsDevice.h

@@ -53,7 +53,7 @@ namespace reco {
     }
 
     uint32_t nHits() const { return this->template view<TrackingRecHitSoA>().metadata().size(); }
-    uint32_t nModules() const { return this->template view<HitModuleSoA>().metadata().size(); }
+    uint32_t nModules() const { return this->template view<HitModuleSoA>().metadata().size() - 1; }
 
     int32_t offsetBPIX2() const { return offsetBPIX2_; }
 

DataFormats/TrackingRecHitSoA/interface/TrackingRecHitsHost.h

@@ -48,7 +48,7 @@ namespace reco {
     }
 
     uint32_t nHits() const { return this->template view<TrackingRecHitSoA>().metadata().size(); }
-    uint32_t nModules() const { return this->template view<HitModuleSoA>().metadata().size(); }
+    uint32_t nModules() const { return this->template view<HitModuleSoA>().metadata().size() - 1; }
 
     int32_t offsetBPIX2() const { return this->template view<TrackingRecHitSoA>().offsetBPIX2(); }
 

DataFormats/TrackingRecHitSoA/interface/alpaka/TrackingRecHitsSoACollection.h

@@ -60,13 +60,12 @@ namespace cms::alpakatools {
     static auto copyAsync(TQueue& queue, reco::TrackingRecHitHost const& hostData) {
       using TDevice = typename alpaka::trait::DevType<TQueue>::type;
 
-      auto hostHitView = hostData.template view<reco::TrackingRecHitSoA>();
-      auto moduleHitsView = hostData.template view<reco::HitModuleSoA>();
+      auto nHits = hostData.nHits();
 
       reco::TrackingRecHitDevice<TDevice> deviceData(
-          queue, hostHitView.metadata().size(), moduleHitsView.metadata().size());
+          queue, nHits, hostData.nModules());
 
-      if (hostHitView.metadata().size() == 0) {
+      if (nHits == 0) {
         std::memset(
             deviceData.buffer().data(),
             0,

HeterogeneousCore/AlpakaInterface/interface/prefixScan.h

@@ -188,13 +188,13 @@ namespace cms::alpakatools {
         psum[i] = (j < size) ? co[j] : T(0);
       }
       alpaka::syncBlockThreads(acc);
-      if (blocksPerGrid <= 1024)
+      if (blocksPerGrid <= warpSize*warpSize)
         blockPrefixScan(acc, psum, psum, blocksPerGrid, ws);
       else {
         auto off = 0u;
-        while (off + 1024 < blocksPerGrid) {
-          blockPrefixScan(acc, psum + off, psum + off, 1024, ws);
-          off = off + 1024 - 1;
+        while (off + warpSize*warpSize < blocksPerGrid) {
+          blockPrefixScan(acc, psum + off, warpSize*warpSize, ws);
+          off = off + warpSize*warpSize - 1;
           alpaka::syncBlockThreads(acc);
         }
         blockPrefixScan(acc, psum + off, psum + off, blocksPerGrid - off, ws);

RecoLocalTracker/SiPixelRecHits/plugins/alpaka/PixelRecHits.h

@@ -45,7 +45,7 @@ namespace ALPAKA_ACCELERATOR_NAMESPACE {
           // This is necessary only once - consider moving it somewhere else.
           // Copy the average geometry corrected by the beamspot.
           if (0 == module) {
-            // auto& agc = hits.averageGeometry();
+  
             auto const& ag = cpeParams->averageGeometry();
             auto nLadders = TrackerTraits::numberOfLaddersInBarrel;
 

RecoLocalTracker/SiPixelRecHits/src/PixelCPEFastParamsHost.cc

@@ -255,17 +255,16 @@ void PixelCPEFastParamsHost<TrackerTraits>::fillParamsForDevice() {
 
 #ifdef ONLY_TRIPLETS_IN_HOLE
   // compute ladder baricenter (only in global z) for the barrel
-  //
 
   constexpr int numberOfModulesInLadder = TrackerTraits::numberOfModulesInLadder;
   constexpr int numberOfLaddersInBarrel = TrackerTraits::numberOfLaddersInBarrel;
   constexpr int numberOfModulesInBarrel = TrackerTraits::numberOfModulesInBarrel;
 
-  constexpr int firstEndcapPos = TrackerTraits::firstEndcapPos;
-  constexpr int firstEndcapNeg = TrackerTraits::firstEndcapNeg;
-
   constexpr float ladderFactor = 1.f / float(numberOfModulesInLadder);
 
+  constexpr int firstEndcapPos = TrackerTraits::firstEndcapPos;
+  constexpr int firstEndcapNeg = TrackerTraits::firstEndcapNeg;
+ 
   auto& aveGeom = buffer_->averageGeometry();
   int il = 0;
   for (int im = 0, nm = numberOfModulesInBarrel; im < nm; ++im) {

RecoTracker/PixelSeeding/plugins/alpaka/CAHitNtupletGenerator.cc

@@ -43,7 +43,7 @@ namespace ALPAKA_ACCELERATOR_NAMESPACE {
       //// Pixel Cluster Cuts (@cell level)
       desc.add<double>("dzdrFact", 8.0f * 0.0285f / 0.015f);
       desc.add<unsigned int>("minYsizeB1", 1)
-          ->setComment("Cut on inner hit cluster size (in Y) for barrel-forward cells. Barrel 1 cut.");
+->setComment("Cut on inner hit cluster size (in Y) for barrel-forward cells. Barrel 1 cut.");
       desc.add<unsigned int>("minYsizeB2", 1)
           ->setComment("Cut on inner hit cluster size (in Y) for barrel-forward cells. Barrel 2 cut.");
       desc.add<unsigned int>("maxDYsize12", 28)
@@ -54,15 +54,35 @@ namespace ALPAKA_ACCELERATOR_NAMESPACE {
           ->setComment("Cut on cluster size differences (in Y) for barrel-forward cells. Barrel-forward cells.");
 
       // Container sizes
-      desc.add<std::string>("maxNumberOfDoublets", std::to_string(pixelTopology::Phase1::maxNumberOfDoublets));
-      desc.add<std::string>("maxNumberOfTuples", std::to_string(pixelTopology::Phase1::maxNumberOfTuples));
-      desc.add<double>("avgHitsPerTrack", 5.0f);
-      desc.add<double>("avgCellsPerHit", 25.0f);
-      desc.add<double>("avgCellsPerCell", 2.0f);
-      desc.add<double>("avgTracksPerCell", 1.0f);
+      // 
+      // maxNumberOfDoublets and maxNumberOfTuples may be defined at runtime depending on the number of hits.
+      // This is done via a TFormula expecting 'x' as nHits.
+      // e.g. : maxNumberOfDoublets = cms.string( '0.00022*pow(x,2)  + 0.53*x + 10000' )
+      // will compute maxNumberOfDoublets for each event as
+      //  
+      //  	maxNumberOfDoublets = 2.2e-4 * nHits^2 + 0.53 * nHits + 10000
+      // 
+      // this may also be simply a constant (as for the default parameters)
+      //  
+      // 	 maxNumberOfDoublets = cms.string(str(512*1024))
+      //
+
+      desc.add<std::string>("maxNumberOfDoublets", std::to_string(pixelTopology::Phase1::maxNumberOfDoublets))
+	->setComment("Max nummber of doublets (cells) as a string. The string will be parsed to a TFormula, depending on nHits (labeled 'x'), \n and evaluated for each event. May also be a constant.");
+      desc.add<std::string>("maxNumberOfTuples", std::to_string(pixelTopology::Phase1::maxNumberOfTuples))
+	 ->setComment("Max nummber of tuples as a string. Same behavior as maxNumberOfDoublets.");
+      desc.add<double>("avgHitsPerTrack", 5.0f)
+          ->setComment("Number of hits per track. Average per track.");
+      desc.add<double>("avgCellsPerHit", 25.0f)
+          ->setComment("Number of cells for which an hit is the outer hit. Average per hit.");
+      desc.add<double>("avgCellsPerCell", 2.0f)
+         ->setComment("Number of cells connected to another cell. Average per cell.");
+      desc.add<double>("avgTracksPerCell", 1.0f)
+          ->setComment("Number of tracks to which a cell belongs. Average per cell.");
 
       // nTuplet Cuts and Params
       desc.add<double>("ptmin", 0.9f)->setComment("Cut on minimum pt");
+      //// p [GeV/c] = B [T] * R [m] * 0.3 (factor from conversion from J to GeV and q = e = 1.6 * 10e-19 C)
       //// 87 cm/GeV = 1/(3.8T * 0.3)
       //// take less than radius given by the hardPtCut and reject everything below
       desc.add<double>("hardCurvCut", 1.f / (0.35 * 87.f))