GPU/TPC: Simplify pad indexing in noisy-pad filter

Author: fweig

GPU/GPUTracking/DataTypes/TPCPadBitMap.h

@@ -68,7 +68,7 @@ struct TPCPadBitMap {
   {
    public:
     using T = uint32_t;
-    static constexpr int32_t NWORDS = (TPC_PADS_IN_SECTOR + sizeof(T) * 8 - 1) / sizeof(T);
+    static constexpr int32_t NWORDS = (TPC_REAL_PADS_IN_SECTOR + sizeof(T) * 8 - 1) / sizeof(T);
     GPUdi() SectorBitMap()
     {
       reset();

GPU/GPUTracking/DataTypes/TPCPadGainCalib.h

@@ -120,12 +120,14 @@ struct TPCPadGainCalib {
 
     GPUd() void reset()
     {
-      for (uint16_t p = 0; p < TPC_PADS_IN_SECTOR; p++) {
+      for (uint16_t p = 0; p < TPC_REAL_PADS_IN_SECTOR; p++) {
         set(p, 1.0f);
       }
     }
 
    private:
+    T mGainCorrection[TPC_REAL_PADS_IN_SECTOR];
+
     GPUd() T pack(float f) const
     {
       f = CAMath::Clamp(f, mMinCorrectionFactor, mMaxCorrectionFactor);
@@ -140,8 +142,6 @@ struct TPCPadGainCalib {
       return mMinCorrectionFactor + (mMaxCorrectionFactor - mMinCorrectionFactor) * float(c) / float(NumOfSteps);
     }
 
-    T mGainCorrection[TPC_PADS_IN_SECTOR];
-
     GPUdi() T& at(uint16_t globalPad)
     {
       return mGainCorrection[globalPad];

GPU/GPUTracking/Definitions/clusterFinderDefs.h

@@ -32,15 +32,19 @@
 #endif
 
 // Padding of 2 and 3 respectively would be enough. But this ensures that
-// rows are always aligned along cache lines. Likewise for TPC_PADS_PER_ROW.
+// rows are always aligned along cache lines. Likewise for TPC_CLUSTERER_ROW_PAD_CAPACITY.
 #define GPUCF_PADDING_PAD 8
 #define GPUCF_PADDING_TIME 4
-#define TPC_PADS_PER_ROW 144
+// Largest possible number of pads in a TPC row
+#define TPC_CLUSTERER_ROW_PAD_CAPACITY 144
 
-#define TPC_ROWS_PER_CRU 18
-#define TPC_PADS_PER_ROW_PADDED (TPC_PADS_PER_ROW + GPUCF_PADDING_PAD)
-#define TPC_NUM_OF_PADS (GPUCA_ROW_COUNT * TPC_PADS_PER_ROW_PADDED + GPUCF_PADDING_PAD)
-#define TPC_PADS_IN_SECTOR 14560
+// Stride between rows as stored internally by the clusterizer
+#define TPC_CLUSTERER_ROW_STRIDE (TPC_CLUSTERER_ROW_PAD_CAPACITY + GPUCF_PADDING_PAD)
+// Number of pads in a sector as stored internally by the clusterizer.
+// This includes fake pads for constant strides between rows
+#define TPC_CLUSTERER_STRIDED_PAD_COUNT (GPUCA_ROW_COUNT * TPC_CLUSTERER_ROW_STRIDE + GPUCF_PADDING_PAD)
+// Real of number of pads in a sector
+#define TPC_REAL_PADS_IN_SECTOR 14560
 #define TPC_FEC_IDS_IN_SECTOR 23296
 #define TPC_MAX_FRAGMENT_LEN_GPU 4000
 #define TPC_MAX_FRAGMENT_LEN_HOST 1000

GPU/GPUTracking/Global/GPUChainTrackingClusterizer.cxx

@@ -861,7 +861,7 @@ int32_t GPUChainTracking::RunTPCClusterizer(bool synchronizeOutput)
         runKernel<GPUMemClean16>({GetGridAutoStep(lane, RecoStep::TPCClusterFinding)}, clustererShadow.mPchargeMap, TPCMapMemoryLayout<ChargeMapType>::items(GetProcessingSettings().overrideClusterizerFragmentLen) * sizeof(ChargeMapType));
         runKernel<GPUMemClean16>({GetGridAutoStep(lane, RecoStep::TPCClusterFinding)}, clustererShadow.mPpeakMap, TPCMapMemoryLayout<PeakMapType>::items(GetProcessingSettings().overrideClusterizerFragmentLen) * sizeof(PeakMapType));
         if (fragment.index == 0) {
-          runKernel<GPUMemClean16>({GetGridAutoStep(lane, RecoStep::TPCClusterFinding)}, clustererShadow.mPpadIsNoisy, TPC_PADS_IN_SECTOR * sizeof(*clustererShadow.mPpadIsNoisy));
+          runKernel<GPUMemClean16>({GetGridAutoStep(lane, RecoStep::TPCClusterFinding)}, clustererShadow.mPpadIsNoisy, TPC_CLUSTERER_STRIDED_PAD_COUNT * sizeof(*clustererShadow.mPpadIsNoisy));
         }
         DoDebugAndDump(RecoStep::TPCClusterFinding, GPUChainTrackingDebugFlags::TPCClustererZeroedCharges, clusterer, &GPUTPCClusterFinder::DumpChargeMap, *mDebugFile, "Zeroed Charges");
 
@@ -965,7 +965,7 @@ int32_t GPUChainTracking::RunTPCClusterizer(bool synchronizeOutput)
           const int32_t nBlocks = GPUTPCCFCheckPadBaseline::GetNBlocks(doGPU);
 
           runKernel<GPUTPCCFCheckPadBaseline>({GetGridBlk(nBlocks, lane), {iSector}});
-          getKernelTimer<GPUTPCCFCheckPadBaseline>(RecoStep::TPCClusterFinding, iSector, TPC_PADS_IN_SECTOR * fragment.lengthWithoutOverlap() * sizeof(PackedCharge), false);
+          getKernelTimer<GPUTPCCFCheckPadBaseline>(RecoStep::TPCClusterFinding, iSector, TPC_REAL_PADS_IN_SECTOR * fragment.lengthWithoutOverlap() * sizeof(PackedCharge), false);
         }
 
         runKernel<GPUTPCCFPeakFinder>({GetGrid(clusterer.mPmemory->counters.nPositions, lane), {iSector}});

GPU/GPUTracking/TPCClusterFinder/CfArray2D.h

@@ -49,7 +49,7 @@ class TilingLayout
   enum {
     Height = Grid::Height,
     Width = Grid::Width,
-    WidthInTiles = (TPC_NUM_OF_PADS + Width - 1) / Width,
+    WidthInTiles = (TPC_CLUSTERER_STRIDED_PAD_COUNT + Width - 1) / Width,
   };
 
   GPUdi() static tpccf::SizeT idx(const CfChargePos& p)
@@ -65,7 +65,7 @@ class TilingLayout
 
   GPUd() static size_t items(size_t fragmentLen)
   {
-    return (TPC_NUM_OF_PADS + Width - 1) / Width * Width * (TPC_MAX_FRAGMENT_LEN_PADDED(fragmentLen) + Height - 1) / Height * Height;
+    return (TPC_CLUSTERER_STRIDED_PAD_COUNT + Width - 1) / Width * Width * (TPC_MAX_FRAGMENT_LEN_PADDED(fragmentLen) + Height - 1) / Height * Height;
   }
 };
 
@@ -74,12 +74,12 @@ class LinearLayout
  public:
   GPUdi() static tpccf::SizeT idx(const CfChargePos& p)
   {
-    return TPC_NUM_OF_PADS * p.timePadded + p.gpad;
+    return TPC_CLUSTERER_STRIDED_PAD_COUNT * p.timePadded + p.gpad;
   }
 
   GPUd() static size_t items(size_t fragmentLen)
   {
-    return TPC_NUM_OF_PADS * TPC_MAX_FRAGMENT_LEN_PADDED(fragmentLen);
+    return TPC_CLUSTERER_STRIDED_PAD_COUNT * TPC_MAX_FRAGMENT_LEN_PADDED(fragmentLen);
   }
 };
 

GPU/GPUTracking/TPCClusterFinder/CfChargePos.h

@@ -42,8 +42,8 @@ struct CfChargePos {
 
   GPUdi() bool valid() const { return timePadded >= 0; }
 
-  GPUdi() tpccf::Row row() const { return gpad / TPC_PADS_PER_ROW_PADDED; }
-  GPUdi() tpccf::Pad pad() const { return gpad % TPC_PADS_PER_ROW_PADDED - GPUCF_PADDING_PAD; }
+  GPUdi() tpccf::Row row() const { return gpad / TPC_CLUSTERER_ROW_STRIDE; }
+  GPUdi() tpccf::Pad pad() const { return gpad % TPC_CLUSTERER_ROW_STRIDE - GPUCF_PADDING_PAD; }
   GPUdi() tpccf::TPCFragmentTime time() const { return timePadded - GPUCF_PADDING_TIME; }
   GPUdi() tpccf::TPCFragmentTime globalTime() const { return timePadded; }
 
@@ -52,7 +52,7 @@ struct CfChargePos {
   // index between 0 and TPC_NUM_OF_PADS.
   static constexpr GPUdi() tpccf::GlobalPad tpcGlobalPadIdx(tpccf::Row row, tpccf::Pad pad)
   {
-    return TPC_PADS_PER_ROW_PADDED * row + pad + GPUCF_PADDING_PAD;
+    return TPC_CLUSTERER_ROW_STRIDE * row + pad + GPUCF_PADDING_PAD;
   }
 };
 

GPU/GPUTracking/TPCClusterFinder/GPUTPCCFCheckPadBaseline.cxx

@@ -50,8 +50,10 @@ GPUd() void GPUTPCCFCheckPadBaseline::CheckBaselineGPU(int32_t nBlocks, int32_t
   const CfFragment& fragment = clusterer.mPmemory->fragment;
   CfArray2D<PackedCharge> chargeMap(reinterpret_cast<PackedCharge*>(clusterer.mPchargeMap));
 
+  constexpr GPUTPCGeometry geo;
+
   const auto iRow = iBlock;
-  const auto rowinfo = GetRowInfo(iRow);
+  const auto nPads = geo.NPads(iRow);
   const CfChargePos basePos{(Row)iRow, 0, 0};
 
   int32_t totalCharges = 0;
@@ -62,7 +64,7 @@ GPUd() void GPUTPCCFCheckPadBaseline::CheckBaselineGPU(int32_t nBlocks, int32_t
   const int16_t iPadOffset = iThread % MaxNPadsPerRow;
   const int16_t iTimeOffset = iThread / MaxNPadsPerRow;
   const int16_t iPadHandle = iThread;
-  const bool handlePad = iPadHandle < rowinfo.nPads;
+  const bool handlePad = iPadHandle < nPads;
 
   const auto firstTB = fragment.firstNonOverlapTimeBin();
   const auto lastTB = fragment.lastNonOverlapTimeBin();
@@ -73,7 +75,7 @@ GPUd() void GPUTPCCFCheckPadBaseline::CheckBaselineGPU(int32_t nBlocks, int32_t
 
     const CfChargePos pos = basePos.delta({iPadOffset, iTime});
 
-    smem.charges[iTimeOffset][iPadOffset] = iTime < lastTB && iPadOffset < rowinfo.nPads ? chargeMap[pos].unpack() : 0;
+    smem.charges[iTimeOffset][iPadOffset] = iTime < lastTB && iPadOffset < nPads ? chargeMap[pos].unpack() : 0;
 
     GPUbarrier();
 
@@ -91,7 +93,7 @@ GPUd() void GPUTPCCFCheckPadBaseline::CheckBaselineGPU(int32_t nBlocks, int32_t
   }
 
   if (handlePad) {
-    updatePadBaseline(rowinfo.globalPadOffset + iPadOffset, clusterer, totalCharges, maxConsecCharges, maxCharge);
+    updatePadBaseline(basePos.gpad + iPadHandle, clusterer, totalCharges, maxConsecCharges, maxCharge);
   }
 #endif
 }
@@ -102,11 +104,10 @@ GPUd() void GPUTPCCFCheckPadBaseline::CheckBaselineCPU(int32_t nBlocks, int32_t
   const CfFragment& fragment = clusterer.mPmemory->fragment;
   CfArray2D<PackedCharge> chargeMap(reinterpret_cast<PackedCharge*>(clusterer.mPchargeMap));
 
-  int32_t basePad = iBlock * PadsPerCacheline;
-  int32_t padsPerRow;
-  CfChargePos basePos = padToCfChargePos<PadsPerCacheline>(basePad, clusterer, padsPerRow);
+  CfChargePos basePos(iBlock * PadsPerCacheline, 0);
 
-  if (not basePos.valid()) {
+  constexpr GPUTPCGeometry geo;
+  if (basePos.pad() >= geo.NPads(basePos.row())) {
     return;
   }
 
@@ -153,45 +154,11 @@ GPUd() void GPUTPCCFCheckPadBaseline::CheckBaselineCPU(int32_t nBlocks, int32_t
   }
 
   for (tpccf::Pad localpad = 0; localpad < PadsPerCacheline; localpad++) {
-    updatePadBaseline(basePad + localpad, clusterer, totalCharges[localpad], maxConsecCharges[localpad], maxCharge[localpad]);
+    updatePadBaseline(basePos.gpad + localpad, clusterer, totalCharges[localpad], maxConsecCharges[localpad], maxCharge[localpad]);
   }
 #endif
 }
 
-template <int32_t PadsPerBlock>
-GPUd() CfChargePos GPUTPCCFCheckPadBaseline::padToCfChargePos(int32_t& pad, const GPUTPCClusterFinder& clusterer, int32_t& padsPerRow)
-{
-  constexpr GPUTPCGeometry geo;
-
-  int32_t padOffset = 0;
-  for (Row r = 0; r < GPUCA_ROW_COUNT; r++) {
-    int32_t npads = geo.NPads(r);
-    int32_t padInRow = pad - padOffset;
-    if (0 <= padInRow && padInRow < npads) {
-      int32_t cachelineOffset = padInRow % PadsPerBlock;
-      pad -= cachelineOffset;
-      padsPerRow = npads;
-      return CfChargePos{r, Pad(padInRow - cachelineOffset), 0};
-    }
-    padOffset += npads;
-  }
-
-  padsPerRow = 0;
-  return CfChargePos{0, 0, INVALID_TIME_BIN};
-}
-
-GPUd() GPUTPCCFCheckPadBaseline::RowInfo GPUTPCCFCheckPadBaseline::GetRowInfo(int16_t row)
-{
-  constexpr GPUTPCGeometry geo;
-
-  int16_t padOffset = 0;
-  for (int16_t r = 0; r < row; r++) {
-    padOffset += geo.NPads(r);
-  }
-
-  return RowInfo{padOffset, geo.NPads(row)};
-}
-
 GPUd() void GPUTPCCFCheckPadBaseline::updatePadBaseline(int32_t pad, const GPUTPCClusterFinder& clusterer, int32_t totalCharges, int32_t consecCharges, Charge maxCharge)
 {
   const CfFragment& fragment = clusterer.mPmemory->fragment;

GPU/GPUTracking/TPCClusterFinder/GPUTPCCFCheckPadBaseline.h

@@ -63,7 +63,7 @@ class GPUTPCCFCheckPadBaseline : public GPUKernelTemplate
 
   static int32_t GetNBlocks(bool isGPU)
   {
-    const int32_t nBlocks = TPC_PADS_IN_SECTOR / PadsPerCacheline;
+    const int32_t nBlocks = TPC_CLUSTERER_STRIDED_PAD_COUNT / PadsPerCacheline;
     return isGPU ? GPUCA_ROW_COUNT : nBlocks;
   }
 
@@ -74,15 +74,6 @@ class GPUTPCCFCheckPadBaseline : public GPUKernelTemplate
   GPUd() static void CheckBaselineGPU(int32_t nBlocks, int32_t nThreads, int32_t iBlock, int32_t iThread, GPUSharedMemory& smem, processorType& clusterer);
   GPUd() static void CheckBaselineCPU(int32_t nBlocks, int32_t nThreads, int32_t iBlock, int32_t iThread, GPUSharedMemory& smem, processorType& clusterer);
 
-  template <int32_t PadsPerBlock>
-  GPUd() static CfChargePos padToCfChargePos(int32_t& pad, const GPUTPCClusterFinder&, int32_t& padsPerRow);
-
-  struct RowInfo {
-    int16_t globalPadOffset;
-    int16_t nPads;
-  };
-  GPUd() static RowInfo GetRowInfo(int16_t row);
-
   GPUd() static void updatePadBaseline(int32_t pad, const GPUTPCClusterFinder&, int32_t totalCharges, int32_t consecCharges, tpccf::Charge maxCharge);
 };
 

GPU/GPUTracking/TPCClusterFinder/GPUTPCCFPeakFinder.cxx

@@ -105,11 +105,11 @@ GPUd() void GPUTPCCFPeakFinder::findPeaksImpl(int32_t nBlocks, int32_t nThreads,
   // For certain configurations dummy work items are added, so the total
   // number of work items is dividable by 64.
   // These dummy items also compute the last digit but discard the result.
-  CfChargePos pos = positions[CAMath::Min(idx, (SizeT)(digitnum - 1))];
+  CfChargePos pos = positions[CAMath::Min<SizeT>(idx, digitnum - 1)];
   Charge charge = pos.valid() ? chargeMap[pos].unpack() : Charge(0);
 
-  bool hasLostBaseline = padHasLostBaseline[gainCorrection.globalPad(pos.row(), pos.pad())];
-  charge = (hasLostBaseline) ? 0.f : charge;
+  bool hasLostBaseline = padHasLostBaseline[pos.gpad];
+  charge = hasLostBaseline ? 0.f : charge;
 
   uint8_t peak = isPeak(smem, charge, pos, SCRATCH_PAD_SEARCH_N, chargeMap, calib, smem.posBcast, smem.buf);
 

GPU/GPUTracking/TPCClusterFinder/GPUTPCClusterFinder.cxx

@@ -79,7 +79,7 @@ void* GPUTPCClusterFinder::SetPointersOutput(void* mem)
 
 void* GPUTPCClusterFinder::SetPointersScratch(void* mem)
 {
-  computePointerWithAlignment(mem, mPpadIsNoisy, TPC_PADS_IN_SECTOR);
+  computePointerWithAlignment(mem, mPpadIsNoisy, TPC_CLUSTERER_STRIDED_PAD_COUNT);
   computePointerWithAlignment(mem, mPpositions, mNMaxDigitsFragment);
   computePointerWithAlignment(mem, mPpeakPositions, mNMaxPeaks);
   computePointerWithAlignment(mem, mPfilteredPeakPositions, mNMaxClusters);