Heterogeneous ClusterTPAssociation (#105)

Implement a heterogeneous Cluster-to-TrackingParticle associator running on the GPU.
cms-patatrack · Oct 8, 2020 · ef64ba6 · ef64ba6
1 parent 345855d
commit ef64ba6
Show file tree

Hide file tree

Showing 2 changed files with 174 additions and 2 deletions.
diff --git a/RecoPixelVertexing/PixelTriplets/plugins/gpuPixelDoublets.h b/RecoPixelVertexing/PixelTriplets/plugins/gpuPixelDoublets.h
@@ -0,0 +1,161 @@
+#ifndef RecoLocalTracker_SiPixelRecHits_plugins_gpuPixelDoublets_h
+#define RecoLocalTracker_SiPixelRecHits_plugins_gpuPixelDouplets_h
+
+#include <algorithm>
+#include <cassert>
+#include <cmath>
+#include <cstdint>
+#include <cstdio>
+#include <limits>
+
+#include "DataFormats/Math/interface/approx_atan2.h"
+#include "RecoLocalTracker/SiPixelRecHits/plugins/siPixelRecHitsHeterogeneousProduct.h"
+
+namespace gpuPixelDoublets {
+
+  __device__
+  std::pair<int,int>
+  findPhiLimits(int16_t phiMe, int16_t * iphi, uint16_t * index, uint16_t size, int16_t iphicut) {
+
+  assert(iphicut>0);
+
+  // find extreemes in top
+  int16_t minPhi = phiMe-iphicut;
+  int16_t maxPhi = phiMe+iphicut;
+
+  // std::cout << "\n phi min/max " << phiMe << ' ' << minPhi << ' ' << maxPhi << std::endl;
+
+  // guess and adjust
+  auto findLimit = [&](int16_t mPhi) {
+    int jm = float(0.5f*size)*(1.f+float(mPhi)/float(std::numeric_limits<short>::max()));
+    // std::cout << "jm for " << mPhi << ' ' << jm << std::endl;
+    jm = std::min(size-1,std::max(0,jm));
+    bool notDone=true;
+    while(jm>0 && mPhi<iphi[index[--jm]]){notDone=false;}
+    if (notDone) while(jm<size && mPhi>iphi[index[++jm]]){}
+    jm = std::min(size-1,std::max(0,jm));
+    return jm;
+  };
+
+  auto jmin = findLimit(minPhi);
+  auto jmax = findLimit(maxPhi);
+
+
+  /*
+  std::cout << "j min/max " << jmin << ' ' << jmax << std::endl;
+  std::cout << "found min/max " << iphi[index[jmin]] << ' ' << iphi[index[jmax]] << std::endl;
+  std::cout << "found min/max +1 " << iphi[index[jmin+1]] << ' ' << iphi[index[jmax+1]] << std::endl;
+  std::cout << "found min/max -1 " << iphi[index[jmin-1]] << ' ' << iphi[index[jmax-1]] << std::endl;
+  */
+
+  return std::make_pair(jmin,jmax);
+  }
+
+
+  __global__
+  void getDoubletsFromSorted(int16_t * iphi, uint16_t * index, uint32_t * offsets, float phiCut) {
+    auto iphicut = phi2short(phiCut);
+    auto i = blockIdx.x*blockDim.x + threadIdx.x;
+    if (i>=offsets[9]) {
+      // get rid of last layer
+      return;
+    }
+
+    assert(0==offsets[0]);
+    int top = (i>offsets[5]) ? 5: 0;
+    while (i>=offsets[++top]){};
+    assert(top<10);
+    auto bottom = top-1;
+    if (bottom == 3 or bottom == 6) {
+      // do not have UP... (9 we got rid already)
+      return;
+    }
+    assert(i >= offsets[bottom]);
+    assert(i < offsets[top]);
+
+    if (index[i]>= (offsets[top]-offsets[bottom])) {
+      printf("index problem: %d %d %d %d %d\n",i, offsets[top], offsets[bottom], offsets[top]-offsets[bottom], index[i]);
+      return;
+    }
+
+    assert(index[i]<offsets[top]-offsets[bottom]);
+
+    int16_t phiMe = iphi[offsets[bottom]+index[i]];
+
+    auto size = offsets[top+1] - offsets[top];
+    assert(size<std::numeric_limits<uint16_t>::max());
+
+    auto jLimits = findPhiLimits(phiMe, iphi+offsets[top],index+offsets[top],size,iphicut);
+
+    auto slidingWindow = [&](uint16_t mysize, uint16_t mymin,uint16_t mymax) {
+      auto topPhi = iphi+offsets[top];
+      uint16_t imax =  std::numeric_limits<uint16_t>::max();
+      uint16_t offset = (mymin>mymax) ? imax-(mysize-1) : 0;
+      int n=0;
+      for (uint16_t i = mymin+offset; i!=mymax; i++) {
+        assert(i<=imax);
+        uint16_t k = (i>mymax) ? i-offset : i;
+        assert(k<mysize);
+        assert(k>=mymin || k<mymax);
+        if (int16_t(topPhi[k]-phiMe)>2*iphicut && int16_t(phiMe-topPhi[k])>2*iphicut)
+          printf("deltaPhi problem: %d %d %d %d, deltas %d:%d cut %d\n",i,k,phiMe,topPhi[k],int16_t(topPhi[k]-phiMe),int16_t(phiMe-topPhi[k]),iphicut);
+        n++;
+      }
+      int tot = (mymin>mymax) ? (mysize-mymin)+mymax : mymax-mymin;
+      assert(n==tot);
+    };
+
+    slidingWindow(size,jLimits.first,jLimits.second);
+  }
+
+  template<typename Hist>
+  __device__
+  void doubletsFromHisto(int16_t const * iphi, Hist const * hist, uint32_t const * offsets, float phiCut) {
+    auto iphicut = phi2short(phiCut);
+    auto i = blockIdx.x*blockDim.x + threadIdx.x;
+    if (i>=offsets[9]) {
+      // get rid of last layer
+      return;
+    }
+
+    assert(0==offsets[0]);
+    int top = (i>offsets[5]) ? 5: 0;
+    while (i>=offsets[++top]){};
+    assert(top<10);
+    auto bottom = top-1;
+    if (bottom==3 || bottom==6) {
+      // do not have UP... (9 we got rid already)
+      return;
+    }
+    assert(i>=offsets[bottom]);
+    assert(i<offsets[top]);
+
+    auto mep = iphi[i];
+    auto kl = hist[top].bin(mep-iphicut);
+    auto kh = hist[top].bin(mep+iphicut);
+    auto incr = [](auto & k) { return k = (k+1)%Hist::nbins();};
+    int tot  = 0;
+    int nmin = 0;
+    auto khh = kh;
+    incr(khh);
+    for (auto kk=kl; kk!=khh; incr(kk)) {
+      if (kk!=kl && kk!=kh) nmin+=hist[top].size(kk);
+      for (auto p=hist[top].begin(kk); p<hist[top].end(kk); ++p) {
+        if (std::min(std::abs(int16_t(iphi[*p]-mep)), std::abs(int16_t(mep-iphi[*p]))) > iphicut)
+          continue;
+        ++tot;
+      }
+    }
+    if (0==hist[top].nspills) assert(tot>=nmin);
+    // look in spill bin as well....
+  }
+
+  __global__
+  void getDoubletsFromHisto(siPixelRecHitsHeterogeneousProduct::HitsOnGPU const * hhp, float phiCut) {
+    auto const & hh = *hhp;
+    doubletsFromHisto(hh.iphi_d,hh.hist_d,hh.hitsLayerStart_d,phiCut);
+  }
+
+} // namespace end
+
+#endif // RecoLocalTracker_SiPixelRecHits_plugins_gpuPixelDouplets_h
diff --git a/Validation/RecoTrack/python/TrackValidation_cff.py b/Validation/RecoTrack/python/TrackValidation_cff.py
@@ -709,9 +709,14 @@ def _uniqueFirstLayers(layerList):
 
 
 ### Pixel tracking only mode (placeholder for now)
-tpClusterProducerPixelTrackingOnly = tpClusterProducer.clone(
-    pixelClusterSrc = "siPixelClustersPreSplitting"
+
+tpClusterProducerHeterogeneousPixelTrackingOnly = tpClusterProducerHeterogeneous.clone(
+   pixelClusterSrc = "siPixelClustersPreSplitting"
 )
+tpClusterProducerPixelTrackingOnly = tpClusterProducer.clone()
+# Need to use the modifier to customize because the exact EDProducer type depends on the modifier
+gpu.toModify(tpClusterProducerPixelTrackingOnly, src = "tpClusterProducerHeterogeneousPixelTrackingOnly")
+
 quickTrackAssociatorByHitsPixelTrackingOnly = quickTrackAssociatorByHits.clone(
     cluster2TPSrc = "tpClusterProducerPixelTrackingOnly"
 )
@@ -739,12 +744,18 @@ def _uniqueFirstLayers(layerList):
 tracksValidationTruthPixelTrackingOnly.replace(quickTrackAssociatorByHits, quickTrackAssociatorByHitsPixelTrackingOnly)
 tracksValidationTruthPixelTrackingOnly.replace(trackingParticleRecoTrackAsssociation, trackingParticlePixelTrackAsssociation)
 tracksValidationTruthPixelTrackingOnly.replace(VertexAssociatorByPositionAndTracks, PixelVertexAssociatorByPositionAndTracks)
+
+_tracksValidationTruthPixelTrackingOnlyGPU = tracksValidationTruthPixelTrackingOnly.copy()
+_tracksValidationTruthPixelTrackingOnlyGPU.insert(0, tpClusterProducerHeterogeneousPixelTrackingOnly)
+gpu.toReplaceWith(tracksValidationTruthPixelTrackingOnly, _tracksValidationTruthPixelTrackingOnlyGPU)
+
 tracksValidationPixelTrackingOnly = cms.Sequence(
     tracksValidationTruthPixelTrackingOnly +
     trackValidatorPixelTrackingOnly
 )
 
 
+
 ### Lite mode (only generalTracks and HP)
 trackValidatorLite = trackValidator.clone(
     label = ["generalTracks", "cutsRecoTracksHp"]