Compute number of required blocks relative to block size.

Author: fweig

GPU/GPUTracking/Global/GPUChainTrackingClusterizer.cxx

@@ -962,8 +962,9 @@ int32_t GPUChainTracking::RunTPCClusterizer(bool synchronizeOutput)
         checkForNoisyPads &= !GetProcessingSettings().disableTPCNoisyPadFilter;
 
         if (checkForNoisyPads) {
-          const int32_t padsPerBlock = doGPU ? 16 : 8; // FIXME: Don't hardcode this!!!
-          const int32_t nBlocks = TPC_PADS_IN_SECTOR / padsPerBlock;
+          const auto props = getKernelProperties<GPUTPCCFCheckPadBaseline>();
+
+          const int32_t nBlocks = GPUTPCCFCheckPadBaseline::GetNBlocks(doGPU, props.nThreads);
 
           runKernel<GPUTPCCFCheckPadBaseline>({GetGridBlk(nBlocks, lane), {iSector}});
           getKernelTimer<GPUTPCCFCheckPadBaseline>(RecoStep::TPCClusterFinding, iSector, TPC_PADS_IN_SECTOR * fragment.lengthWithoutOverlap() * sizeof(PackedCharge), false);

GPU/GPUTracking/TPCClusterFinder/GPUTPCCFCheckPadBaseline.h

@@ -49,6 +49,13 @@ class GPUTPCCFCheckPadBaseline : public GPUKernelTemplate
     return gpudatatypes::RecoStep::TPCClusterFinding;
   }
 
+  static int32_t GetNBlocks(bool isGPU, int32_t nThreads)
+  {
+    const int32_t padsPerBlock = isGPU ? nThreads / TimebinsPerCacheline : PadsPerCacheline;
+    const int32_t nBlocks = TPC_PADS_IN_SECTOR / padsPerBlock;
+    return nBlocks;
+  }
+
   template <int32_t iKernel = defaultKernel>
   GPUd() static void Thread(int32_t nBlocks, int32_t nThreads, int32_t iBlock, int32_t iThread, GPUSharedMemory& smem, processorType& clusterer);