Merge pull request #48079 from SegmentLinking/T3_profiling_study

optimize LST T3 kernel block size and rearrange T3 kernel functions

Author: cmsbuild

RecoTracker/LSTCore/src/alpaka/LSTEvent.dev.cc

@@ -402,7 +402,7 @@ void LSTEvent::createTriplets() {
   auto index_gpu_buf = cms::alpakatools::make_device_buffer<uint16_t[]>(queue_, nLowerModules_);
   alpaka::memcpy(queue_, index_gpu_buf, index_buf_h, nonZeroModules);
 
-  auto const createTriplets_workDiv = cms::alpakatools::make_workdiv<Acc3D>({max_blocks, 1, 1}, {1, 16, 16});
+  auto const createTriplets_workDiv = cms::alpakatools::make_workdiv<Acc3D>({nonZeroModules, 1, 1}, {1, 16, 16});
 
   alpaka::exec<Acc3D>(queue_,
                       createTriplets_workDiv,

RecoTracker/LSTCore/src/alpaka/Triplet.h

@@ -648,10 +648,6 @@ namespace ALPAKA_ACCELERATOR_NAMESPACE::lst {
                                                                    float& circleCenterX,
                                                                    float& circleCenterY,
                                                                    const float ptCut) {
-    //this cut reduces the number of candidates by a factor of 4, i.e., 3 out of 4 warps can end right here!
-    if (segments.mdIndices()[innerSegmentIndex][1] != segments.mdIndices()[outerSegmentIndex][0])
-      return false;
-
     unsigned int firstMDIndex = segments.mdIndices()[innerSegmentIndex][0];
     unsigned int secondMDIndex = segments.mdIndices()[outerSegmentIndex][0];
     unsigned int thirdMDIndex = segments.mdIndices()[outerSegmentIndex][1];
@@ -741,6 +737,10 @@ namespace ALPAKA_ACCELERATOR_NAMESPACE::lst {
           for (unsigned int outerSegmentArrayIndex : cms::alpakatools::uniform_elements_x(acc, nOuterSegments)) {
             unsigned int outerSegmentIndex = ranges.segmentRanges()[middleLowerModuleIndex][0] + outerSegmentArrayIndex;
 
+            //this cut reduces the number of candidates by a factor of 4, i.e., 3 out of 4 warps can end right here!
+            if (segments.mdIndices()[innerSegmentIndex][1] != segments.mdIndices()[outerSegmentIndex][0])
+              continue;
+
             uint16_t outerOuterLowerModuleIndex = segments.outerLowerModuleIndices()[outerSegmentIndex];
 
             float zOut, rtOut, betaIn, betaInCut, circleRadius, circleCenterX, circleCenterY;