Use qTot as marker for HIP clusters instead of qMax

Author: fweig

DataFormats/Detectors/TPC/include/DataFormatsTPC/ClusterNative.h

@@ -63,8 +63,9 @@ struct ClusterNative {
   static constexpr int scalePadPacked = 64;       //< ~60 is needed for 0.1mm precision, but power of two avoids rounding
   static constexpr int scaleSigmaTimePacked = 32; // 1/32nd of pad/timebin precision for cluster size
   static constexpr int scaleSigmaPadPacked = 32;
-  static constexpr int scaleSaturatedQTot = 4;
-  static constexpr int maxSaturatedQTot = USHRT_MAX * scaleSaturatedQTot;
+  static constexpr int scaleSaturatedQtot = 8;
+  static constexpr int maxRegularQtot = 25 * 1024;
+  static constexpr int maxSaturatedQtot = (USHRT_MAX - maxRegularQtot) * scaleSaturatedQtot;
 
   uint32_t timeFlagsPacked; //< Contains the time in the lower 24 bits in a packed format, contains the flags in the
                             // upper 8 bits
@@ -89,9 +90,7 @@ struct ClusterNative {
   GPUd() uint16_t getQtot() const
   {
     if (isSaturated()) [[unlikely]] {
-      // Check for overflow, so return type can stay uint16
-      auto sqtot = getSaturatedQtot();
-      return sqtot <= USHRT_MAX ? sqtot : USHRT_MAX;
+      return maxRegularQtot;
     }
     return qTot;
   }
@@ -155,19 +154,19 @@ struct ClusterNative {
     sigmaPadPacked = tmp;
   }
 
-  GPUd() bool isSaturated() const { return qMax >= 1023; }
+  GPUd() bool isSaturated() const { return qTot > maxRegularQtot; }
 
   GPUd() void setSaturatedQtot(uint32_t qtot)
   {
-    if (qtot > maxSaturatedQTot) {
-      qtot = maxSaturatedQTot;
+    if (qtot > maxSaturatedQtot) {
+      qtot = maxSaturatedQtot;
     }
-    this->qTot = (qtot + scaleSaturatedQTot / 2) / scaleSaturatedQTot;
+    this->qTot = ((qtot + scaleSaturatedQtot / 2) / scaleSaturatedQtot) + maxRegularQtot;
   }
 
   GPUd() uint32_t getSaturatedQtot() const
   {
-    return uint32_t(qTot) * scaleSaturatedQTot;
+    return uint32_t(qTot - maxRegularQtot) * scaleSaturatedQtot;
   }
 
   GPUd() void setSaturatedTailLength(uint32_t tail)

GPU/GPUTracking/DataCompression/GPUTPCClusterStatistics.cxx

@@ -129,7 +129,7 @@ void GPUTPCClusterStatistics::RunStatistics(const o2::tpc::ClusterNativeAccess*
           if (param.rec.tpc.compressionTypeMask & GPUSettings::CompressionTruncate) {
             GPUTPCCompression::truncateSignificantBitsChargeMax(tmpClusters[k].qMax, param);
             GPUTPCCompression::truncateSignificantBitsWidth(tmpClusters[k].sigmaPadPacked, param);
-            if (!tmpClusters[k].isSaturated()) {
+            if (!tmpClusters[k].isSaturated()) [[likely]] {
               GPUTPCCompression::truncateSignificantBitsCharge(tmpClusters[k].qTot, param);
               GPUTPCCompression::truncateSignificantBitsWidth(tmpClusters[k].sigmaTimePacked, param);
             }

GPU/GPUTracking/DataCompression/GPUTPCCompression.h

@@ -47,7 +47,7 @@ class GPUTPCCompression : public GPUProcessor
 #endif
 
   static constexpr uint32_t P_MAX_QMAX = 1 << 10;
-  static constexpr uint32_t P_MAX_QTOT = 1 << 16;
+  static constexpr uint32_t P_MAX_QTOT = 5 * 5 * P_MAX_QMAX;
   static constexpr uint32_t P_MAX_TIME = 1 << 24;
   static constexpr uint32_t P_MAX_PAD = 1 << 16;
   static constexpr uint32_t P_MAX_SIGMA = 1 << 8;

GPU/GPUTracking/DataCompression/GPUTPCCompressionKernels.cxx

@@ -122,7 +122,7 @@ GPUdii() void GPUTPCCompressionKernels::Thread<GPUTPCCompressionKernels::step0at
       if (param.rec.tpc.compressionTypeMask & GPUSettings::CompressionTruncate) {
         compressor.truncateSignificantBitsChargeMax(qmax, param);
         compressor.truncateSignificantBitsWidth(sigmapad, param);
-        if (!orgCl.isSaturated()) {
+        if (!orgCl.isSaturated()) [[likely]] {
           compressor.truncateSignificantBitsCharge(qtot, param);
           compressor.truncateSignificantBitsWidth(sigmatime, param);
         }
@@ -301,7 +301,7 @@ GPUdii() void GPUTPCCompressionKernels::Thread<GPUTPCCompressionKernels::step1un
         if (param.rec.tpc.compressionTypeMask & GPUSettings::CompressionTruncate) {
           compressor.truncateSignificantBitsChargeMax(qmax, param);
           compressor.truncateSignificantBitsWidth(sigmapad, param);
-          if (!orgCl.isSaturated()) {
+          if (!orgCl.isSaturated()) [[likely]] {
             compressor.truncateSignificantBitsCharge(qtot, param);
             compressor.truncateSignificantBitsWidth(sigmatime, param);
           }