Synchronize event in the CUDAProductBase destructor (#391)

Otherwise there are possibilities for weird races, e.g. combination of
non-ExternalWork producers, consumed-but-not-read CUDAProducts, CUDA
streams executing work later than expected (= on the next event).

CUDADataFormats/Common/interface/CUDAProductBase.h

@@ -17,7 +17,10 @@ namespace impl {
 class CUDAProductBase {
 public:
   CUDAProductBase() = default;  // Needed only for ROOT dictionary generation
+  ~CUDAProductBase();
 
+  CUDAProductBase(const CUDAProductBase&) = delete;
+  CUDAProductBase& operator=(const CUDAProductBase&) = delete;
   CUDAProductBase(CUDAProductBase&& other)
       : stream_{std::move(other.stream_)},
         event_{std::move(other.event_)},

CUDADataFormats/Common/src/CUDAProductBase.cc

@@ -9,3 +9,15 @@ bool CUDAProductBase::isAvailable() const {
   }
   return cudautils::eventIsOccurred(event_->id());
 }
+
+CUDAProductBase::~CUDAProductBase() {
+  // Make sure that the production of the product in the GPU is
+  // complete before destructing the product. This is to make sure
+  // that the EDM stream does not move to the next event before all
+  // asynchronous processing of the current is complete.
+  if (event_) {
+    // TODO: a callback notifying a WaitingTaskHolder (or similar)
+    // would avoid blocking the CPU, but would also require more work.
+    event_->synchronize();
+  }
+}

CUDADataFormats/Track/BuildFile.xml

@@ -1,5 +1,6 @@
 <use name="cuda"/>
 <use name="rootcore"/>
+<use name="CUDADataFormats/Common"/>
 <use name="DataFormats/Common"/>
 <use name="HeterogeneousCore/CUDAUtilities"/>
 <use name="eigen"/>

CUDADataFormats/Vertex/BuildFile.xml

@@ -1,5 +1,6 @@
 <use name="cuda"/>
 <use name="rootcore"/>
+<use name="CUDADataFormats/Common"/>
 <use name="DataFormats/Common"/>
 <use name="HeterogeneousCore/CUDAUtilities"/>
 <use name="eigen"/>