Merge pull request #44774 from AdrianoDee/runtime_nhits

[14_0_X] Allow Runtime Number of Hits for Alpaka Pixel Reconstruction
cms-sw · Apr 18, 2024 · 653fed5 · 653fed5
1 parent 8134154
commit 653fed5
Show file tree

Hide file tree

Showing 6 changed files with 23 additions and 24 deletions.
diff --git a/Geometry/CommonTopologies/interface/SimplePixelTopology.h b/Geometry/CommonTopologies/interface/SimplePixelTopology.h
@@ -323,7 +323,6 @@ namespace pixelTopology {
     using tindex_type = uint32_t;  // for tuples
     using cindex_type = uint32_t;  // for cells
 
-    static constexpr uint32_t maxNumberOfHits = 256 * 1024;
     static constexpr uint32_t maxCellNeighbors = 64;
     static constexpr uint32_t maxCellTracks = 302;
     static constexpr uint32_t maxHitsOnTrack = 15;
@@ -417,7 +416,6 @@ namespace pixelTopology {
     using tindex_type = uint16_t;  // for tuples
     using cindex_type = uint32_t;  // for cells
 
-    static constexpr uint32_t maxNumberOfHits = 48 * 1024;
     static constexpr uint32_t maxCellNeighbors = 36;
     static constexpr uint32_t maxCellTracks = 48;
     static constexpr uint32_t maxHitsOnTrack = 10;

diff --git a/RecoLocalTracker/SiPixelClusterizer/plugins/alpaka/SiPixelRawToClusterKernel.dev.cc b/RecoLocalTracker/SiPixelClusterizer/plugins/alpaka/SiPixelRawToClusterKernel.dev.cc
@@ -488,9 +488,9 @@ namespace ALPAKA_ACCELERATOR_NAMESPACE {
           alpaka::syncBlockThreads(acc);
         }
 #ifdef GPU_DEBUG
-        ALPAKA_ASSERT_ACC(0 == clus_view[0].moduleStart());
-        auto c0 = std::min(maxHitsInModule, clus_view[1].clusModuleStart());
-        ALPAKA_ASSERT_ACC(c0 == clus_view[1].moduleStart());
+        ALPAKA_ASSERT_ACC(0 == clus_view[1].moduleStart());
+        auto c0 = std::min(maxHitsInModule, clus_view[2].clusModuleStart());
+        ALPAKA_ASSERT_ACC(c0 == clus_view[2].moduleStart());
         ALPAKA_ASSERT_ACC(clus_view[1024].moduleStart() >= clus_view[1023].moduleStart());
         ALPAKA_ASSERT_ACC(clus_view[1025].moduleStart() >= clus_view[1024].moduleStart());
         ALPAKA_ASSERT_ACC(clus_view[numberOfModules].moduleStart() >= clus_view[1025].moduleStart());
@@ -504,13 +504,8 @@ namespace ALPAKA_ACCELERATOR_NAMESPACE {
           if (i == bpix2 || i == fpix1)
             printf("moduleStart %d %d\n", i, clus_view[i].moduleStart());
         }
+
 #endif
-        // avoid overflow
-        constexpr auto MAX_HITS = TrackerTraits::maxNumberOfHits;
-        for (uint32_t i : cms::alpakatools::independent_group_elements(acc, numberOfModules + 1)) {
-          if (clus_view[i].clusModuleStart() > MAX_HITS)
-            clus_view[i].clusModuleStart() = MAX_HITS;
-        }
 
       }  // end of FillHitsModuleStart kernel operator()
     };   // end of FillHitsModuleStart struct

diff --git a/RecoLocalTracker/SiPixelRecHits/plugins/alpaka/PixelRecHitKernels.dev.cc b/RecoLocalTracker/SiPixelRecHits/plugins/alpaka/PixelRecHitKernels.dev.cc
@@ -113,15 +113,6 @@ namespace ALPAKA_ACCELERATOR_NAMESPACE {
           hrv_d.contentSize = nHits;
           hrv_d.contentStorage = hits_d.view().phiBinnerStorage();
 
-          // fillManyFromVector<Acc1D>(h_d.data(), nParts, v_d.data(), offsets_d.data(), offsets[10], 256, queue);
-          /*          cms::alpakatools::fillManyFromVector<Acc1D>(&(hits_d.view().phiBinner()),
-                                                      nLayers,
-                                                      hits_d.view().iphi(),
-                                                      hits_d.view().hitsLayerStart().data(),
-                                                      nHits,
-                                                      (uint32_t)256,
-                                                      queue);
-*/
           cms::alpakatools::fillManyFromVector<Acc1D>(&(hits_d.view().phiBinner()),
                                                       hrv_d,
                                                       nLayers,

diff --git a/RecoTracker/PixelSeeding/plugins/alpaka/CAHitNtupletGeneratorKernels.dev.cc b/RecoTracker/PixelSeeding/plugins/alpaka/CAHitNtupletGeneratorKernels.dev.cc
@@ -37,6 +37,8 @@ namespace ALPAKA_ACCELERATOR_NAMESPACE {
 
         // workspace
         device_hitToTuple_{cms::alpakatools::make_device_buffer<HitToTuple>(queue)},
+        device_hitToTupleStorage_{
+            cms::alpakatools::make_device_buffer<typename HitToTuple::Counter[]>(queue, nhits + 1)},
         device_tupleMultiplicity_{cms::alpakatools::make_device_buffer<TupleMultiplicity>(queue)},
 
         // NB: In legacy, device_theCells_ and device_isOuterHitOfCell_ were allocated inside buildDoublets
@@ -66,6 +68,10 @@ namespace ALPAKA_ACCELERATOR_NAMESPACE {
         device_hitToTuple_apc_{reinterpret_cast<cms::alpakatools::AtomicPairCounter *>(device_storage_.data() + 1)},
         device_nCells_{cms::alpakatools::make_device_view(alpaka::getDev(queue),
                                                           *reinterpret_cast<uint32_t *>(device_storage_.data() + 2))} {
+#ifdef GPU_DEBUG
+    std::cout << "Allocation for tuple building. N hits " << nhits << std::endl;
+#endif
+
     alpaka::memset(queue, counters_, 0);
     alpaka::memset(queue, device_nCells_, 0);
     alpaka::memset(queue, cellStorage_, 0);
@@ -74,14 +80,21 @@ namespace ALPAKA_ACCELERATOR_NAMESPACE {
     alpaka::memcpy(queue, device_cellCuts_, cellCuts_h);
 
     [[maybe_unused]] TupleMultiplicity *tupleMultiplicityDeviceData = device_tupleMultiplicity_.data();
-    [[maybe_unused]] HitToTuple *hitToTupleDeviceData = device_hitToTuple_.data();
     using TM = cms::alpakatools::OneToManyAssocRandomAccess<typename TrackerTraits::tindex_type,
                                                             TrackerTraits::maxHitsOnTrack + 1,
                                                             TrackerTraits::maxNumberOfTuples>;
     TM *tm = device_tupleMultiplicity_.data();
     TM::template launchZero<Acc1D>(tm, queue);
     TupleMultiplicity::template launchZero<Acc1D>(tupleMultiplicityDeviceData, queue);
-    HitToTuple::template launchZero<Acc1D>(hitToTupleDeviceData, queue);
+
+    device_hitToTupleView_.assoc = device_hitToTuple_.data();
+    device_hitToTupleView_.offStorage = device_hitToTupleStorage_.data();
+    device_hitToTupleView_.offSize = nhits + 1;
+
+    HitToTuple::template launchZero<Acc1D>(device_hitToTupleView_, queue);
+#ifdef GPU_DEBUG
+    std::cout << "Allocations for CAHitNtupletGeneratorKernels: done!" << std::endl;
+#endif
   }
 
   template <typename TrackerTraits>
@@ -401,7 +414,7 @@ namespace ALPAKA_ACCELERATOR_NAMESPACE {
                           tracks_view,
                           this->device_hitToTuple_.data());  //CHECK
 
-      HitToTuple::template launchFinalize<Acc1D>(this->device_hitToTuple_.data(), queue);
+      HitToTuple::template launchFinalize<Acc1D>(this->device_hitToTupleView_, queue);
       alpaka::exec<Acc1D>(
           queue, workDiv1D, Kernel_fillHitInTracks<TrackerTraits>{}, tracks_view, this->device_hitToTuple_.data());
 #ifdef GPU_DEBUG

diff --git a/RecoTracker/PixelSeeding/plugins/alpaka/CAHitNtupletGeneratorKernels.h b/RecoTracker/PixelSeeding/plugins/alpaka/CAHitNtupletGeneratorKernels.h
@@ -253,6 +253,8 @@ namespace ALPAKA_ACCELERATOR_NAMESPACE {
 
     // workspace
     cms::alpakatools::device_buffer<Device, HitToTuple> device_hitToTuple_;
+    cms::alpakatools::device_buffer<Device, uint32_t[]> device_hitToTupleStorage_;
+    typename HitToTuple::View device_hitToTupleView_;
     cms::alpakatools::device_buffer<Device, TupleMultiplicity> device_tupleMultiplicity_;
     cms::alpakatools::device_buffer<Device, CACell[]> device_theCells_;
     cms::alpakatools::device_buffer<Device, OuterHitOfCellContainer[]> device_isOuterHitOfCell_;

diff --git a/RecoTracker/PixelSeeding/plugins/alpaka/CAStructures.h b/RecoTracker/PixelSeeding/plugins/alpaka/CAStructures.h
@@ -31,7 +31,7 @@ namespace caStructures {
   template <typename TrackerTraits>
   using HitToTupleT =
       cms::alpakatools::OneToManyAssocRandomAccess<typename TrackerTraits::tindex_type,
-                                                   TrackerTraits::maxNumberOfHits,
+                                                   -1,
                                                    TrackerTraits::maxHitsForContainers>;  // 3.5 should be enough
 
   template <typename TrackerTraits>