GPU: Several fixes for sanitizers and build with clang

Common/MathUtils/include/MathUtils/detail/basicMath.h

@@ -113,7 +113,11 @@ GPUdi() int nint(double x)
 template <>
 GPUdi() bool finite(double x)
 {
+#ifdef __FAST_MATH__
+  return false;
+#else
   return std::isfinite(x);
+#endif
 }
 template <>
 GPUdi() double log(double x)

GPU/GPUTracking/Base/GPUReconstruction.cxx

@@ -147,8 +147,12 @@ int32_t GPUReconstruction::Init()
   if (InitDevice()) {
     return 1;
   }
-  mHostMemoryPoolEnd = (char*)mHostMemoryBase + mHostMemorySize;
-  mDeviceMemoryPoolEnd = (char*)mDeviceMemoryBase + mDeviceMemorySize;
+  if (mProcessingSettings.memoryAllocationStrategy == GPUMemoryResource::ALLOCATION_GLOBAL) {
+    mHostMemoryPoolEnd = (char*)mHostMemoryBase + mHostMemorySize;
+    mDeviceMemoryPoolEnd = (char*)mDeviceMemoryBase + mDeviceMemorySize;
+  } else {
+    mHostMemoryPoolEnd = mDeviceMemoryPoolEnd = nullptr;
+  }
   if (InitPhasePermanentMemory()) {
     return 1;
   }
@@ -860,14 +864,18 @@ void GPUReconstruction::ClearAllocatedMemory(bool clearOutputs)
       FreeRegisteredMemory(i);
     }
   }
-  mHostMemoryPool = GPUProcessor::alignPointer<GPUCA_MEMALIGN>(mHostMemoryPermanent);
-  mDeviceMemoryPool = GPUProcessor::alignPointer<GPUCA_MEMALIGN>(mDeviceMemoryPermanent);
   mUnmanagedChunks.clear();
-  mVolatileMemoryStart = nullptr;
   mNonPersistentMemoryStack.clear();
   mNonPersistentIndividualAllocations.clear();
-  mHostMemoryPoolEnd = mHostMemoryPoolBlocked ? mHostMemoryPoolBlocked : ((char*)mHostMemoryBase + mHostMemorySize);
-  mDeviceMemoryPoolEnd = mDeviceMemoryPoolBlocked ? mDeviceMemoryPoolBlocked : ((char*)mDeviceMemoryBase + mDeviceMemorySize);
+  mVolatileMemoryStart = nullptr;
+  if (mProcessingSettings.memoryAllocationStrategy == GPUMemoryResource::ALLOCATION_GLOBAL) {
+    mHostMemoryPool = GPUProcessor::alignPointer<GPUCA_MEMALIGN>(mHostMemoryPermanent);
+    mDeviceMemoryPool = GPUProcessor::alignPointer<GPUCA_MEMALIGN>(mDeviceMemoryPermanent);
+    mHostMemoryPoolEnd = mHostMemoryPoolBlocked ? mHostMemoryPoolBlocked : ((char*)mHostMemoryBase + mHostMemorySize);
+    mDeviceMemoryPoolEnd = mDeviceMemoryPoolBlocked ? mDeviceMemoryPoolBlocked : ((char*)mDeviceMemoryBase + mDeviceMemorySize);
+  } else {
+    mHostMemoryPool = mDeviceMemoryPool = mHostMemoryPoolEnd = mDeviceMemoryPoolEnd = nullptr;
+  }
 }
 
 void GPUReconstruction::UpdateMaxMemoryUsed()

GPU/GPUTracking/Base/GPUReconstructionCPU.cxx

@@ -112,7 +112,12 @@ inline void GPUReconstructionCPUBackend::runKernelBackendInternal<GPUMemClean16,
 template <class T, int32_t I, typename... Args>
 void GPUReconstructionCPUBackend::runKernelBackend(const krnlSetupArgs<T, I, Args...>& args)
 {
+#pragma GCC diagnostic push
+#if defined(__clang__)
+#pragma GCC diagnostic ignored "-Wunused-lambda-capture" // this is not alway captured below
+#endif
   std::apply([this, &args](auto&... vals) { runKernelBackendInternal<T, I, Args...>(args.s, vals...); }, args.v);
+#pragma GCC diagnostic push
 }
 
 template <class T, int32_t I>

GPU/GPUTracking/Global/GPUChainTrackingSectorTracker.cxx

@@ -150,7 +150,7 @@ int32_t GPUChainTracking::RunTPCTrackingSectors_internal()
   if (param().rec.tpc.occupancyMapTimeBins || param().rec.tpc.sysClusErrorC12Norm) {
     uint32_t& occupancyTotal = *mInputsHost->mTPCClusterOccupancyMap;
     occupancyTotal = CAMath::Float2UIntRn(mRec->MemoryScalers()->nTPCHits / (mIOPtrs.settingsTF && mIOPtrs.settingsTF->hasNHBFPerTF ? mIOPtrs.settingsTF->nHBFPerTF : 128));
-    mRec->UpdateParamOccupancyMap(param().rec.tpc.occupancyMapTimeBins ? mInputsHost->mTPCClusterOccupancyMap + 2 : nullptr, param().rec.tpc.occupancyMapTimeBins ? mInputsShadow->mTPCClusterOccupancyMap + 2 : nullptr, occupancyTotal, streamInitAndOccMap);
+    mRec->UpdateParamOccupancyMap(param().rec.tpc.occupancyMapTimeBins ? mInputsHost->mTPCClusterOccupancyMap + 2 : nullptr, doGPU && param().rec.tpc.occupancyMapTimeBins ? mInputsShadow->mTPCClusterOccupancyMap + 2 : nullptr, occupancyTotal, streamInitAndOccMap);
   }
 
   int32_t streamMap[NSECTORS];

GPU/GPUTracking/Merger/GPUTPCGMMerger.cxx

@@ -1706,20 +1706,20 @@ GPUd() void GPUTPCGMMerger::CollectMergedTracks(int32_t nBlocks, int32_t nThread
       nHits = nFilteredHits;
     }
 
-    uint32_t iOutTrackFirstCluster = CAMath::AtomicAdd(&mMemory->nOutputTrackClusters, (uint32_t)nHits);
+    const uint32_t iOutTrackFirstCluster = CAMath::AtomicAdd(&mMemory->nOutputTrackClusters, (uint32_t)nHits);
     if (iOutTrackFirstCluster >= mNMaxOutputTrackClusters) {
       raiseError(GPUErrors::ERROR_MERGER_HIT_OVERFLOW, iOutTrackFirstCluster, mNMaxOutputTrackClusters);
       CAMath::AtomicExch(&mMemory->nOutputTrackClusters, mNMaxOutputTrackClusters);
       continue;
     }
 
-    GPUTPCGMMergedTrackHit* cl = mClusters + iOutTrackFirstCluster;
-    GPUTPCGMMergedTrackHitXYZ* clXYZ = mClustersXYZ + iOutTrackFirstCluster;
+    GPUTPCGMMergedTrackHit* const cl = mClusters + iOutTrackFirstCluster;
 
     for (int32_t i = 0; i < nHits; i++) {
       uint8_t state;
       if (Param().par.earlyTpcTransform) {
         const GPUTPCClusterData& c = GetConstantMem()->tpcTrackers[trackClusters[i].sector].ClusterData()[trackClusters[i].id - GetConstantMem()->tpcTrackers[trackClusters[i].sector].Data().ClusterIdOffset()];
+        GPUTPCGMMergedTrackHitXYZ* const clXYZ = mClustersXYZ + iOutTrackFirstCluster;
         clXYZ[i].x = c.x;
         clXYZ[i].y = c.y;
         clXYZ[i].z = c.z;
@@ -1760,7 +1760,7 @@ GPUd() void GPUTPCGMMerger::CollectMergedTracks(int32_t nBlocks, int32_t nThread
     mergedTrack.SetCSide(p2.CSide());
 
     GPUTPCGMBorderTrack b;
-    const float toX = Param().par.earlyTpcTransform ? clXYZ[0].x : GPUTPCGeometry::Row2X(cl[0].row);
+    const float toX = Param().par.earlyTpcTransform ? mClustersXYZ[iOutTrackFirstCluster].x : GPUTPCGeometry::Row2X(cl[0].row);
     if (p2.TransportToX(this, toX, Param().bzCLight, b, GPUCA_MAX_SIN_PHI, false)) {
       p1.X() = toX;
       p1.Y() = b.Par()[0];
@@ -1791,12 +1791,13 @@ GPUd() void GPUTPCGMMerger::CollectMergedTracks(int32_t nBlocks, int32_t nThread
     if (Param().rec.tpc.mergeCE) {
       bool CEside;
       if (Param().par.earlyTpcTransform) {
+        const GPUTPCGMMergedTrackHitXYZ* const clXYZ = mClustersXYZ + iOutTrackFirstCluster;
         CEside = (mergedTrack.CSide() != 0) ^ (clXYZ[0].z > clXYZ[nHits - 1].z);
       } else {
         auto& cls = mConstantMem->ioPtrs.clustersNative->clustersLinear;
         CEside = cls[cl[0].num].getTime() < cls[cl[nHits - 1].num].getTime();
       }
-      MergeCEFill(trackParts[CEside ? lastTrackIndex : firstTrackIndex], cl[CEside ? (nHits - 1) : 0], &clXYZ[CEside ? (nHits - 1) : 0], iOutputTrack);
+      MergeCEFill(trackParts[CEside ? lastTrackIndex : firstTrackIndex], cl[CEside ? (nHits - 1) : 0], Param().par.earlyTpcTransform ? &(mClustersXYZ + iOutTrackFirstCluster)[CEside ? (nHits - 1) : 0] : nullptr, iOutputTrack);
     }
   } // itr
 }

GPU/GPUTracking/Merger/GPUTPCGMTrackParam.cxx

@@ -1091,7 +1091,7 @@ GPUd() void GPUTPCGMTrackParam::RefitTrack(GPUTPCGMMergedTrack& GPUrestrict() tr
   GPUTPCGMTrackParam t = track.Param();
   float Alpha = track.Alpha();
   CADEBUG(int32_t nTrackHitsOld = nTrackHits; float ptOld = t.QPt());
-  bool ok = t.Fit(merger, iTrk, merger->Clusters() + track.FirstClusterRef(), merger->ClustersXYZ() + track.FirstClusterRef(), nTrackHits, NTolerated, Alpha, attempt, GPUCA_MAX_SIN_PHI, &track.OuterParam());
+  bool ok = t.Fit(merger, iTrk, merger->Clusters() + track.FirstClusterRef(), merger->Param().par.earlyTpcTransform ? merger->ClustersXYZ() + track.FirstClusterRef() : nullptr, nTrackHits, NTolerated, Alpha, attempt, GPUCA_MAX_SIN_PHI, &track.OuterParam());
   CADEBUG(printf("Finished Fit Track %d\n", iTrk));
   CADEBUG(printf("OUTPUT hits %d -> %d+%d = %d, QPt %f -> %f, SP %f, ok %d chi2 %f chi2ndf %f\n", nTrackHitsOld, nTrackHits, NTolerated, nTrackHits + NTolerated, ptOld, t.QPt(), t.SinPhi(), (int32_t)ok, t.Chi2(), t.Chi2() / CAMath::Max(1, nTrackHits)));
 

GPU/GPUTracking/Standalone/CMakeLists.txt

@@ -50,9 +50,6 @@ set(CMAKE_POSITION_INDEPENDENT_CODE ON)
 
 if(GPUCA_BUILD_DEBUG)
   set(CMAKE_CXX_FLAGS "-O0 -ggdb")
-  if (GPUCA_BUILD_DEBUG_SANITIZE)
-    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsanitize=address,undefined -fno-sanitize=vptr") #TODO: Check why this does not work with clang
-  endif()
   set(CMAKE_BUILD_TYPE DEBUG)
 else()
   set(CMAKE_CXX_FLAGS "-O3 -march=native -ggdb -minline-all-stringops -funroll-loops -fno-stack-protector")
@@ -67,7 +64,13 @@ else()
   set(CMAKE_BUILD_TYPE RELEASE)
   add_definitions(-DNDEBUG)
 endif()
-set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-error -Wall -Wextra -Wshadow -Wno-unused-function -Wno-unused-parameter -Wno-unused-local-typedefs -Wno-unknown-pragmas -Wno-write-strings")
+if (GPUCA_BUILD_DEBUG_SANITIZE)
+  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsanitize=address,undefined -fno-sanitize=vptr")
+  if(CMAKE_CXX_COMPILER MATCHES "clang\\+\\+")
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -shared-libasan")
+  endif()
+endif()
+set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-error -Wall -Wextra -Wshadow -Wno-unused-function -Wno-unused-parameter -Wno-unused-local-typedefs -Wno-unknown-pragmas -Wno-write-strings -Wno-vla-cxx-extension")
 set(CMAKE_SHARED_LINKER_FLAGS "${CMAKE_SHARED_LINKER_FLAGS} -rdynamic -Wl,--no-undefined")
 
 # Find mandatory packages
@@ -247,6 +250,13 @@ if(GPUCA_CONFIG_ROOT)
 endif()
 target_link_libraries(standalone_support PUBLIC Microsoft.GSL::GSL TPCFastTransformation)
 
+if (GPUCA_BUILD_DEBUG_SANITIZE AND CMAKE_CXX_COMPILER MATCHES "clang\\+\\+")
+  execute_process(COMMAND ${CMAKE_CXX_COMPILER} -print-file-name=libclang_rt.asan-x86_64.so OUTPUT_VARIABLE CLANG_ASAN_SO_PATH OUTPUT_STRIP_TRAILING_WHITESPACE)
+  get_filename_component(CLANG_ASAN_SO_PATH "${CLANG_ASAN_SO_PATH}" DIRECTORY)
+  get_filename_component(CLANG_ASAN_SO_PATH "${CLANG_ASAN_SO_PATH}" ABSOLUTE)
+  target_link_options(ca PRIVATE "-Wl,-rpath,${CLANG_ASAN_SO_PATH}")
+endif()
+
 # Installation
 install(TARGETS ca TPCFastTransformation standalone_support)
 install(FILES "cmake/makefile" DESTINATION "${CMAKE_INSTALL_PREFIX}")

GPU/GPUTracking/display/GPUDisplay.h

@@ -161,6 +161,7 @@ class GPUDisplay : public GPUDisplayInterface
   {
 #pragma GCC diagnostic push
 #pragma GCC diagnostic ignored "-Wformat-security"
+#pragma GCC diagnostic ignored "-Wformat-truncation"
     snprintf(mInfoText2, 1024, args...);
 #pragma GCC diagnostic pop
     GPUInfo("%s", mInfoText2);