NN clusterizer: Improving speed of NN input-filling kernel (#14749)

* Improve GPU filling kernel speed

* Adjusting parameter bounds and additional GPU kernel optimizations

* Adding back if statement for early exit

* const'ing + fixing CPU kernel

* Remiving print statements

* Fixing CI build issue

Author: ChSonnabend

GPU/GPUTracking/Definitions/GPUDefParametersDefaults.h

@@ -491,7 +491,7 @@
 
   #define GPUCA_LB_GPUTPCNNClusterizerKernels_runCfClusterizer GPUCA_LB_GPUTPCNNClusterizerKernels
   #define GPUCA_LB_GPUTPCNNClusterizerKernels_fillInputNNCPU GPUCA_LB_GPUTPCNNClusterizerKernels
-  #define GPUCA_LB_GPUTPCNNClusterizerKernels_fillInputNNGPU GPUCA_LB_GPUTPCNNClusterizerKernels
+  #define GPUCA_LB_GPUTPCNNClusterizerKernels_fillInputNNGPU 1024
   #define GPUCA_LB_GPUTPCNNClusterizerKernels_determineClass1Labels GPUCA_LB_GPUTPCNNClusterizerKernels
   #define GPUCA_LB_GPUTPCNNClusterizerKernels_determineClass2Labels GPUCA_LB_GPUTPCNNClusterizerKernels
   #define GPUCA_LB_GPUTPCNNClusterizerKernels_publishClass1Regression GPUCA_LB_GPUTPCNNClusterizerKernels

GPU/GPUTracking/Global/GPUChainTrackingClusterizer.cxx

@@ -1031,7 +1031,7 @@ int32_t GPUChainTracking::RunTPCClusterizer(bool synchronizeOutput)
             // Filling the data
             if (mRec->IsGPU() || GetProcessingSettings().nn.nnClusterizerForceGpuInputFill) {
               // Fills element by element of each input matrix -> better parallelizability, but worse on CPU due to unnecessary computations
-              runKernel<GPUTPCNNClusterizerKernels, GPUTPCNNClusterizerKernels::fillInputNNGPU>({GetGrid(iSize * clustererNNShadow.mNnClusterizerRowTimeSizeFull, lane), krnlRunRangeNone}, iSector, clustererNNShadow.mNnInferenceInputDType, propagateMCLabels, batchStart);
+              runKernel<GPUTPCNNClusterizerKernels, GPUTPCNNClusterizerKernels::fillInputNNGPU>({GetGrid(iSize * clustererNNShadow.mNnClusterizerRowTimeSizeThreads , lane), krnlRunRangeNone}, iSector, clustererNNShadow.mNnInferenceInputDType, propagateMCLabels, batchStart);
             } else {
               // Fills the whole input matrix at once -> better performance on CPU, but worse parallelizability
               runKernel<GPUTPCNNClusterizerKernels, GPUTPCNNClusterizerKernels::fillInputNNCPU>({GetGrid(iSize, lane), krnlRunRangeNone}, iSector, clustererNNShadow.mNnInferenceInputDType, propagateMCLabels, batchStart);

GPU/GPUTracking/TPCClusterFinder/GPUTPCNNClusterizer.h

@@ -71,6 +71,7 @@ class GPUTPCNNClusterizer : public GPUProcessor
   uint32_t mNnClusterizerPadTimeSize = 0;
   uint32_t mNnClusterizerRowTimeSize = 0;
   uint32_t mNnClusterizerRowTimeSizeFull = 0;
+  uint32_t mNnClusterizerRowTimeSizeThreads = 0;
 
   // Boundary lookup table
   // int32_t mBoundaryMapSizeRow = 0;

GPU/GPUTracking/TPCClusterFinder/GPUTPCNNClusterizerHost.cxx

@@ -98,6 +98,7 @@ void GPUTPCNNClusterizerHost::initClusterizer(const GPUSettingsProcessingNNclust
   clustererNN.mNnClusterizerPadTimeSize = clustererNN.mNnClusterizerFullPadSize * clustererNN.mNnClusterizerFullTimeSize;
   clustererNN.mNnClusterizerRowTimeSize = clustererNN.mNnClusterizerFullRowSize * clustererNN.mNnClusterizerFullTimeSize;
   clustererNN.mNnClusterizerRowTimeSizeFull = clustererNN.mNnClusterizerRowTimeSize + (settings.nnClusterizerAddIndexData ? 3 : 0);
+  clustererNN.mNnClusterizerRowTimeSizeThreads = clustererNN.mNnClusterizerRowTimeSize + (settings.nnClusterizerAddIndexData ? 1 : 0);
   clustererNN.mNnClusterizerElementSize = clustererNN.mNnClusterizerChargeArraySize + (settings.nnClusterizerAddIndexData ? 3 : 0);
   // clustererNN.mBoundaryMapSizeRow = 3 * clustererNN.mNnClusterizerSizeInputRow + o2::tpc::constants::MAXGLOBALPADROW;
   // clustererNN.mBoundaryPadding = 11; // padding on each side to account for pad_offset. N=11 since then mIsBoundary = 24320 ~< (1.5 x 2^14 = 24576) && N must be bigger than (NPads[row(end_iroc + 1)] - NPads[row(end_iroc)])/2 (=6) for pad_offset to work