GPU TPC: In deterministic mode, do not run looper interpolation on GPU in extra kernel

davidrohr · davidrohr · commit d048aa064c11 · 2024-02-28T19:21:57.000+01:00
diff --git a/GPU/GPUTracking/Base/GPUReconstruction.cxx b/GPU/GPUTracking/Base/GPUReconstruction.cxx
@@ -260,6 +260,9 @@ int GPUReconstruction::InitPhaseBeforeDevice()
 #ifdef GPUCA_HAVE_O2HEADERS
     mProcessingSettings.overrideClusterizerFragmentLen = TPC_MAX_FRAGMENT_LEN_GPU;
     param().rec.tpc.nWaysOuter = true;
+    if (param().rec.tpc.looperInterpolationInExtraPass == -1) {
+      param().rec.tpc.looperInterpolationInExtraPass = 0;
+    }
     if (mProcessingSettings.createO2Output > 1) {
       mProcessingSettings.createO2Output = 1;
     }
@@ -300,7 +303,7 @@ int GPUReconstruction::InitPhaseBeforeDevice()
     mProcessingSettings.delayedOutput = false;
   }
   if (!mProcessingSettings.fullMergerOnGPU && GetRecoStepsGPU() & RecoStep::TPCMerging) {
-    param().rec.tpc.loopInterpolationInExtraPass = 0;
+    param().rec.tpc.looperInterpolationInExtraPass = 0;
     if (param().rec.tpc.retryRefit == 1) {
       param().rec.tpc.retryRefit = 2;
     }
diff --git a/GPU/GPUTracking/Base/GPUReconstructionIncludes.h b/GPU/GPUTracking/Base/GPUReconstructionIncludes.h
@@ -37,30 +37,30 @@
 #define RANDOM_ERROR
 //#define RANDOM_ERROR || rand() % 500 == 1
 
-#define GPUCA_GPUReconstructionUpdateDefaults()                                           \
-  if (mProcessingSettings.trackletConstructorInPipeline < 0) {                            \
-    mProcessingSettings.trackletConstructorInPipeline = GPUCA_CONSTRUCTOR_IN_PIPELINE;    \
-  }                                                                                       \
-  if (mProcessingSettings.trackletSelectorInPipeline < 0) {                               \
-    mProcessingSettings.trackletSelectorInPipeline = GPUCA_SELECTOR_IN_PIPELINE;          \
-  }                                                                                       \
-  if (mProcessingSettings.trackletSelectorSlices < 0) {                                   \
-    mProcessingSettings.trackletSelectorSlices = GPUCA_TRACKLET_SELECTOR_SLICE_COUNT;     \
-  }                                                                                       \
-  if (mProcessingSettings.alternateBorderSort < 0) {                                      \
-    mProcessingSettings.alternateBorderSort = GPUCA_ALTERNATE_BORDER_SORT;                \
-  }                                                                                       \
-  if (mProcessingSettings.mergerSortTracks < 0) {                                         \
-    mProcessingSettings.mergerSortTracks = GPUCA_SORT_BEFORE_FIT;                         \
-  }                                                                                       \
-  if (param().rec.tpc.loopInterpolationInExtraPass < 0) {                                 \
-    param().rec.tpc.loopInterpolationInExtraPass = GPUCA_MERGER_SPLIT_LOOP_INTERPOLATION; \
-  }                                                                                       \
-  if (mProcessingSettings.tpcCompressionGatherModeKernel < 0) {                           \
-    mProcessingSettings.tpcCompressionGatherModeKernel = GPUCA_COMP_GATHER_KERNEL;        \
-  }                                                                                       \
-  if (mProcessingSettings.tpcCompressionGatherMode < 0) {                                 \
-    mProcessingSettings.tpcCompressionGatherMode = GPUCA_COMP_GATHER_MODE;                \
+#define GPUCA_GPUReconstructionUpdateDefaults()                                             \
+  if (mProcessingSettings.trackletConstructorInPipeline < 0) {                              \
+    mProcessingSettings.trackletConstructorInPipeline = GPUCA_CONSTRUCTOR_IN_PIPELINE;      \
+  }                                                                                         \
+  if (mProcessingSettings.trackletSelectorInPipeline < 0) {                                 \
+    mProcessingSettings.trackletSelectorInPipeline = GPUCA_SELECTOR_IN_PIPELINE;            \
+  }                                                                                         \
+  if (mProcessingSettings.trackletSelectorSlices < 0) {                                     \
+    mProcessingSettings.trackletSelectorSlices = GPUCA_TRACKLET_SELECTOR_SLICE_COUNT;       \
+  }                                                                                         \
+  if (mProcessingSettings.alternateBorderSort < 0) {                                        \
+    mProcessingSettings.alternateBorderSort = GPUCA_ALTERNATE_BORDER_SORT;                  \
+  }                                                                                         \
+  if (mProcessingSettings.mergerSortTracks < 0) {                                           \
+    mProcessingSettings.mergerSortTracks = GPUCA_SORT_BEFORE_FIT;                           \
+  }                                                                                         \
+  if (param().rec.tpc.looperInterpolationInExtraPass < 0) {                                 \
+    param().rec.tpc.looperInterpolationInExtraPass = GPUCA_MERGER_SPLIT_LOOP_INTERPOLATION; \
+  }                                                                                         \
+  if (mProcessingSettings.tpcCompressionGatherModeKernel < 0) {                             \
+    mProcessingSettings.tpcCompressionGatherModeKernel = GPUCA_COMP_GATHER_KERNEL;          \
+  }                                                                                         \
+  if (mProcessingSettings.tpcCompressionGatherMode < 0) {                                   \
+    mProcessingSettings.tpcCompressionGatherMode = GPUCA_COMP_GATHER_MODE;                  \
   }
 
 #endif
diff --git a/GPU/GPUTracking/Definitions/GPUSettingsList.h b/GPU/GPUTracking/Definitions/GPUSettingsList.h
@@ -135,7 +135,7 @@ AddOptionRTC(mergerCovSource, unsigned char, 2, "", 0, "Method to obtain covaria
 AddOptionRTC(mergerInterpolateErrors, unsigned char, 1, "", 0, "Use interpolation instead of extrapolation for chi2 based cluster rejection")
 AddOptionRTC(mergeCE, unsigned char, 1, "", 0, "Merge tracks accross the central electrode")
 AddOptionRTC(retryRefit, char, 1, "", 0, "Retry refit with seeding errors and without cluster rejection when fit fails")
-AddOptionRTC(loopInterpolationInExtraPass, char, -1, "", 0, "Perform loop interpolation in an extra pass")
+AddOptionRTC(looperInterpolationInExtraPass, char, -1, "", 0, "Perform looper interpolation in an extra pass")
 AddOptionRTC(mergerReadFromTrackerDirectly, char, 1, "", 0, "Forward data directly from tracker to merger on GPU")
 AddOptionRTC(dropSecondaryLegsInOutput, char, 1, "", 0, "Do not store secondary legs of looping track in TrackTPC")
 AddOptionRTC(enablePID, char, 1, "", 0, "Enable PID response")
diff --git a/GPU/GPUTracking/Global/GPUChainTrackingMerger.cxx b/GPU/GPUTracking/Global/GPUChainTrackingMerger.cxx
@@ -225,7 +225,7 @@ int GPUChainTracking::RunTPCTrackingMerger(bool synchronizeOutput)
   if (param().rec.tpc.retryRefit == 1) {
     runKernel<GPUTPCGMMergerTrackFit>(GetGridAuto(0), krnlRunRangeNone, krnlEventNone, -1);
   }
-  if (param().rec.tpc.loopInterpolationInExtraPass) {
+  if (param().rec.tpc.looperInterpolationInExtraPass) {
     runKernel<GPUTPCGMMergerFollowLoopers>(GetGridAuto(0), krnlRunRangeNone, krnlEventNone);
   }
   if (doGPU && !doGPUall) {
diff --git a/GPU/GPUTracking/Merger/GPUTPCGMTrackParam.cxx b/GPU/GPUTracking/Merger/GPUTPCGMTrackParam.cxx
@@ -708,7 +708,7 @@ GPUdic(0, 1) int GPUTPCGMTrackParam::FollowCircle(const GPUTPCGMMerger* GPUrestr
   if (Merger->Param().rec.tpc.disableRefitAttachment & 4) {
     return 1;
   }
-  if (Merger->Param().rec.tpc.loopInterpolationInExtraPass && phase2 == false) {
+  if (Merger->Param().rec.tpc.looperInterpolationInExtraPass && phase2 == false) {
     StoreAttachMirror(Merger, slice, iRow, iTrack, toAlpha, toY, toX, toSlice, toRow, inFlyDirection, prop.GetAlpha());
     return 1;
   }
@@ -811,7 +811,7 @@ GPUdni() void GPUTPCGMTrackParam::AttachClustersMirror(const GPUTPCGMMerger* GPU
   if (Merger->Param().rec.tpc.disableRefitAttachment & 8) {
     return;
   }
-  if (Merger->Param().rec.tpc.loopInterpolationInExtraPass && phase2 == false) {
+  if (Merger->Param().rec.tpc.looperInterpolationInExtraPass && phase2 == false) {
     StoreAttachMirror(Merger, slice, iRow, iTrack, 0, toY, 0, -1, 0, 0, prop.GetAlpha());
     return;
   }
diff --git a/GPU/GPUTracking/Standalone/Benchmark/standalone.cxx b/GPU/GPUTracking/Standalone/Benchmark/standalone.cxx
@@ -451,7 +451,7 @@ int SetupReconstruction()
       procSet.tpcInputWithClusterRejection = 1;
     }
     recSet.tpc.disableRefitAttachment = 0xFF;
-    recSet.tpc.loopInterpolationInExtraPass = 0;
+    recSet.tpc.looperInterpolationInExtraPass = 0;
     recSet.maxTrackQPtB5 = CAMath::Min(recSet.maxTrackQPtB5, recSet.tpc.rejectQPtB5);
     recSet.useMatLUT = true;
     recAsync->SetSettings(&grp, &recSet, &procSet, &steps);

Original file line number	Diff line number	Diff line change
`@@ -260,6 +260,9 @@ int GPUReconstruction::InitPhaseBeforeDevice()`
`260`	`260`	`#ifdef GPUCA_HAVE_O2HEADERS`
`261`	`261`	`mProcessingSettings.overrideClusterizerFragmentLen = TPC_MAX_FRAGMENT_LEN_GPU;`
`262`	`262`	`param().rec.tpc.nWaysOuter = true;`
	`263`	`+ if (param().rec.tpc.looperInterpolationInExtraPass == -1) {`
	`264`	`+ param().rec.tpc.looperInterpolationInExtraPass = 0;`
	`265`	`+ }`
`263`	`266`	`if (mProcessingSettings.createO2Output > 1) {`
`264`	`267`	`mProcessingSettings.createO2Output = 1;`
`265`	`268`	`}`
`@@ -300,7 +303,7 @@ int GPUReconstruction::InitPhaseBeforeDevice()`
`300`	`303`	`mProcessingSettings.delayedOutput = false;`
`301`	`304`	`}`
`302`	`305`	`if (!mProcessingSettings.fullMergerOnGPU && GetRecoStepsGPU() & RecoStep::TPCMerging) {`
`303`		`- param().rec.tpc.loopInterpolationInExtraPass = 0;`
	`306`	`+ param().rec.tpc.looperInterpolationInExtraPass = 0;`
`304`	`307`	`if (param().rec.tpc.retryRefit == 1) {`
`305`	`308`	`param().rec.tpc.retryRefit = 2;`
`306`	`309`	`}`
Original file line number	Diff line number	Diff line change
`@@ -225,7 +225,7 @@ int GPUChainTracking::RunTPCTrackingMerger(bool synchronizeOutput)`
`225`	`225`	`if (param().rec.tpc.retryRefit == 1) {`
`226`	`226`	`runKernel<GPUTPCGMMergerTrackFit>(GetGridAuto(0), krnlRunRangeNone, krnlEventNone, -1);`
`227`	`227`	`}`
`228`		`- if (param().rec.tpc.loopInterpolationInExtraPass) {`
	`228`	`+ if (param().rec.tpc.looperInterpolationInExtraPass) {`
`229`	`229`	`runKernel<GPUTPCGMMergerFollowLoopers>(GetGridAuto(0), krnlRunRangeNone, krnlEventNone);`
`230`	`230`	`}`
`231`	`231`	`if (doGPU && !doGPUall) {`
Original file line number	Diff line number	Diff line change
`@@ -708,7 +708,7 @@ GPUdic(0, 1) int GPUTPCGMTrackParam::FollowCircle(const GPUTPCGMMerger* GPUrestr`
`708`	`708`	`if (Merger->Param().rec.tpc.disableRefitAttachment & 4) {`
`709`	`709`	`return 1;`
`710`	`710`	`}`
`711`		`- if (Merger->Param().rec.tpc.loopInterpolationInExtraPass && phase2 == false) {`
	`711`	`+ if (Merger->Param().rec.tpc.looperInterpolationInExtraPass && phase2 == false) {`
`712`	`712`	`StoreAttachMirror(Merger, slice, iRow, iTrack, toAlpha, toY, toX, toSlice, toRow, inFlyDirection, prop.GetAlpha());`
`713`	`713`	`return 1;`
`714`	`714`	`}`
`@@ -811,7 +811,7 @@ GPUdni() void GPUTPCGMTrackParam::AttachClustersMirror(const GPUTPCGMMerger* GPU`
`811`	`811`	`if (Merger->Param().rec.tpc.disableRefitAttachment & 8) {`
`812`	`812`	`return;`
`813`	`813`	`}`
`814`		`- if (Merger->Param().rec.tpc.loopInterpolationInExtraPass && phase2 == false) {`
	`814`	`+ if (Merger->Param().rec.tpc.looperInterpolationInExtraPass && phase2 == false) {`
`815`	`815`	`StoreAttachMirror(Merger, slice, iRow, iTrack, 0, toY, 0, -1, 0, 0, prop.GetAlpha());`
`816`	`816`	`return;`
`817`	`817`	`}`
Original file line number	Diff line number	Diff line change
`@@ -451,7 +451,7 @@ int SetupReconstruction()`
`451`	`451`	`procSet.tpcInputWithClusterRejection = 1;`
`452`	`452`	`}`
`453`	`453`	`recSet.tpc.disableRefitAttachment = 0xFF;`
`454`		`- recSet.tpc.loopInterpolationInExtraPass = 0;`
	`454`	`+ recSet.tpc.looperInterpolationInExtraPass = 0;`
`455`	`455`	`recSet.maxTrackQPtB5 = CAMath::Min(recSet.maxTrackQPtB5, recSet.tpc.rejectQPtB5);`
`456`	`456`	`recSet.useMatLUT = true;`
`457`	`457`	`recAsync->SetSettings(&grp, &recSet, &procSet, &steps);`