From 99e2a465a201963cca3a09c12538d070a75dd82f Mon Sep 17 00:00:00 2001
From: Vincenzo Innocente <vincenzo.innocente@cern.ch>
Date: Sun, 6 Dec 2020 17:57:31 +0100
Subject: [PATCH] make the PhiBinner the size of the hits

---
 .../interface/TrackingRecHit2DHeterogeneous.h             | 8 +++++++-
 .../TrackingRecHit/interface/TrackingRecHit2DSOAView.h    | 3 ++-
 RecoLocalTracker/SiPixelRecHits/plugins/PixelRecHits.cu   | 2 +-
 .../SiPixelRecHits/plugins/SiPixelRecHitSoAFromLegacy.cc  | 2 +-
 4 files changed, 11 insertions(+), 4 deletions(-)

diff --git a/CUDADataFormats/TrackingRecHit/interface/TrackingRecHit2DHeterogeneous.h b/CUDADataFormats/TrackingRecHit/interface/TrackingRecHit2DHeterogeneous.h
index 73a6daaa4e387..f273294e0e2cb 100644
--- a/CUDADataFormats/TrackingRecHit/interface/TrackingRecHit2DHeterogeneous.h
+++ b/CUDADataFormats/TrackingRecHit/interface/TrackingRecHit2DHeterogeneous.h
@@ -34,6 +34,7 @@ class TrackingRecHit2DHeterogeneous {
   auto hitsModuleStart() const { return m_hitsModuleStart; }
   auto hitsLayerStart() { return m_hitsLayerStart; }
   auto phiBinner() { return m_hist; }
+  auto binnerStorage() {return m_histStorage;}
   auto iphi() { return m_iphi; }
 
   // only the local coord and detector index
@@ -43,7 +44,7 @@ class TrackingRecHit2DHeterogeneous {
 
 private:
   static constexpr uint32_t n16 = 4;
-  static constexpr uint32_t n32 = 9;
+  static constexpr uint32_t n32 = 10;
   static_assert(sizeof(uint32_t) == sizeof(float));  // just stating the obvious
 
   unique_ptr<uint16_t[]> m_store16;  //!
@@ -60,6 +61,7 @@ class TrackingRecHit2DHeterogeneous {
 
   // needed as kernel params...
   Hist* m_hist;
+  Hist::index_type * m_histStorage;
   uint32_t* m_hitsLayerStart;
   int16_t* m_iphi;
 };
@@ -101,11 +103,15 @@ TrackingRecHit2DHeterogeneous<Traits>::TrackingRecHit2DHeterogeneous(uint32_t nH
   m_store32 = Traits::template make_device_unique<float[]>(nHits * n32 + 11, stream);
   m_HistStore = Traits::template make_device_unique<TrackingRecHit2DSOAView::Hist>(stream);
 
+  static_assert(sizeof(TrackingRecHit2DSOAView::hindex_type)==sizeof(float));
+  static_assert(sizeof(TrackingRecHit2DSOAView::hindex_type)==sizeof(TrackingRecHit2DSOAView::Hist::index_type));
+
   auto get16 = [&](int i) { return m_store16.get() + i * nHits; };
   auto get32 = [&](int i) { return m_store32.get() + i * nHits; };
 
   // copy all the pointers
   m_hist = view->m_hist = m_HistStore.get();
+  m_histStorage = view->m_histStorage = reinterpret_cast<TrackingRecHit2DSOAView::Hist::index_type *>(get32(9));
 
   view->m_xl = get32(0);
   view->m_yl = get32(1);
diff --git a/CUDADataFormats/TrackingRecHit/interface/TrackingRecHit2DSOAView.h b/CUDADataFormats/TrackingRecHit/interface/TrackingRecHit2DSOAView.h
index 6a83a66b60fbd..460a8c1e8b7fa 100644
--- a/CUDADataFormats/TrackingRecHit/interface/TrackingRecHit2DSOAView.h
+++ b/CUDADataFormats/TrackingRecHit/interface/TrackingRecHit2DSOAView.h
@@ -18,7 +18,7 @@ class TrackingRecHit2DSOAView {
   using hindex_type = uint32_t;  // if above is <=2^32
 
   using PhiBinner =
-      cms::cuda::HistoContainer<int16_t, 128, gpuClustering::MaxNumClusters, 8 * sizeof(int16_t), hindex_type, 10>;
+      cms::cuda::HistoContainer<int16_t, 128, -1, 8 * sizeof(int16_t), hindex_type, 10>;
 
   using Hist = PhiBinner;  // FIXME
 
@@ -96,6 +96,7 @@ class TrackingRecHit2DSOAView {
   uint32_t* m_hitsLayerStart;
 
   PhiBinner* m_hist;  // FIXME use a more descriptive name consistently
+  PhiBinner::index_type * m_histStorage;
 
   uint32_t m_nHits;
 };
diff --git a/RecoLocalTracker/SiPixelRecHits/plugins/PixelRecHits.cu b/RecoLocalTracker/SiPixelRecHits/plugins/PixelRecHits.cu
index 5dae8f52432d5..f6872bfa5af8c 100644
--- a/RecoLocalTracker/SiPixelRecHits/plugins/PixelRecHits.cu
+++ b/RecoLocalTracker/SiPixelRecHits/plugins/PixelRecHits.cu
@@ -64,7 +64,7 @@ namespace pixelgpudetails {
 
     if (nHits) {
       cms::cuda::fillManyFromVector(
-          hits_d.phiBinner(), 10, hits_d.iphi(), hits_d.hitsLayerStart(), nHits, 256, nullptr, stream);
+          hits_d.phiBinner(), 10, hits_d.iphi(), hits_d.hitsLayerStart(), nHits, 256, hits_d.binnerStorage(), stream);
       cudaCheck(cudaGetLastError());
     }
 
diff --git a/RecoLocalTracker/SiPixelRecHits/plugins/SiPixelRecHitSoAFromLegacy.cc b/RecoLocalTracker/SiPixelRecHits/plugins/SiPixelRecHitSoAFromLegacy.cc
index c7084f325d05b..69f8f74b05ae7 100644
--- a/RecoLocalTracker/SiPixelRecHits/plugins/SiPixelRecHitSoAFromLegacy.cc
+++ b/RecoLocalTracker/SiPixelRecHits/plugins/SiPixelRecHitSoAFromLegacy.cc
@@ -240,7 +240,7 @@ void SiPixelRecHitSoAFromLegacy::produce(edm::StreamID streamID, edm::Event& iEv
     output->hitsLayerStart()[i] = hitsModuleStart[cpeView.layerGeometry().layerStart[i]];
   }
   cms::cuda::fillManyFromVector(
-      output->phiBinner(), 10, output->iphi(), output->hitsLayerStart(), numberOfHits, 256, nullptr);
+      output->phiBinner(), 10, output->iphi(), output->hitsLayerStart(), numberOfHits, 256, output->binnerStorage());
 
   // std::cout << "created HitSoa for " <<  numberOfClusters << " clusters in " << numberOfDetUnits << " Dets" << std::endl;
   iEvent.put(std::move(output));