Reorganize CUDAScopedContext (#355)

* Split CUDAScopedContext to *Acquire and *Produce The motivation is that acquire() and produce() need a different functionality, and are constructed differently (e.g. acquire version always needs the edm::WaitingTaskWithArenaHolder). This split should make it more difficult to make mistakes. It should also make future evolution, e.g. towards chains of TBB tasks alternating in CPU and GPU work, easier. * Rename CUDAContextToken to CUDAContextState, and change semantics Now CUDAScopedContextAcquire takes it as a parameter to constructor, and stores the state in its destructor (yielding RAII semantics). * Document the constructors.
cms-patatrack · Dec 29, 2020 · af2ffa3 · af2ffa3
1 parent acc8360
commit af2ffa3
Show file tree

Hide file tree

Showing 4 changed files with 6 additions and 8 deletions.
diff --git a/EventFilter/SiPixelRawToDigi/plugins/SiPixelDigiErrorsSoAFromCUDA.cc b/EventFilter/SiPixelRawToDigi/plugins/SiPixelDigiErrorsSoAFromCUDA.cc
@@ -43,7 +43,7 @@ void SiPixelDigiErrorsSoAFromCUDA::fillDescriptions(edm::ConfigurationDescriptio
 
 void SiPixelDigiErrorsSoAFromCUDA::acquire(const edm::Event& iEvent, const edm::EventSetup& iSetup, edm::WaitingTaskWithArenaHolder waitingTaskHolder) {
   // Do the transfer in a CUDA stream parallel to the computation CUDA stream
-  CUDAScopedContext ctx{iEvent.streamID(), std::move(waitingTaskHolder)};
+  CUDAScopedContextAcquire ctx{iEvent.streamID(), std::move(waitingTaskHolder)};
 
   const auto& gpuDigiErrors = ctx.get(iEvent, digiErrorGetToken_);
 

diff --git a/EventFilter/SiPixelRawToDigi/plugins/SiPixelDigisSoAFromCUDA.cc b/EventFilter/SiPixelRawToDigi/plugins/SiPixelDigisSoAFromCUDA.cc
@@ -47,7 +47,7 @@ void SiPixelDigisSoAFromCUDA::fillDescriptions(edm::ConfigurationDescriptions& d
 
 void SiPixelDigisSoAFromCUDA::acquire(const edm::Event& iEvent, const edm::EventSetup& iSetup, edm::WaitingTaskWithArenaHolder waitingTaskHolder) {
   // Do the transfer in a CUDA stream parallel to the computation CUDA stream
-  CUDAScopedContext ctx{iEvent.streamID(), std::move(waitingTaskHolder)};
+  CUDAScopedContextAcquire ctx{iEvent.streamID(), std::move(waitingTaskHolder)};
 
   const auto& gpuDigis = ctx.get(iEvent, digiGetToken_);
 

diff --git a/RecoLocalTracker/SiPixelClusterizer/plugins/SiPixelRawToClusterCUDA.cc b/RecoLocalTracker/SiPixelClusterizer/plugins/SiPixelRawToClusterCUDA.cc
@@ -53,7 +53,7 @@ class SiPixelRawToClusterCUDA: public edm::stream::EDProducer<edm::ExternalWork>
   edm::EDPutTokenT<CUDAProduct<SiPixelDigiErrorsCUDA>> digiErrorPutToken_;
   edm::EDPutTokenT<CUDAProduct<SiPixelClustersCUDA>> clusterPutToken_;
 
-  CUDAContextToken ctxTmp_;
+  CUDAContextState ctxState_;
 
   edm::ESWatcher<SiPixelFedCablingMapRcd> recordWatcher;
 
@@ -118,7 +118,7 @@ void SiPixelRawToClusterCUDA::fillDescriptions(edm::ConfigurationDescriptions& d
 
 
 void SiPixelRawToClusterCUDA::acquire(const edm::Event& iEvent, const edm::EventSetup& iSetup, edm::WaitingTaskWithArenaHolder waitingTaskHolder) {
-  CUDAScopedContext ctx{iEvent.streamID(), std::move(waitingTaskHolder)};
+  CUDAScopedContextAcquire ctx{iEvent.streamID(), std::move(waitingTaskHolder), ctxState_};
 
   edm::ESHandle<SiPixelFedCablingMapGPUWrapper> hgpuMap;
   iSetup.get<CkfComponentsRecord>().get(hgpuMap);
@@ -228,12 +228,10 @@ void SiPixelRawToClusterCUDA::acquire(const edm::Event& iEvent, const edm::Event
                              useQuality_, includeErrors_,
                              edm::MessageDrop::instance()->debugEnabled,
                              ctx.stream());
-
-  ctxTmp_ = ctx.toToken();
 }
 
 void SiPixelRawToClusterCUDA::produce(edm::Event& iEvent, const edm::EventSetup& iSetup) {
-  CUDAScopedContext ctx{std::move(ctxTmp_)};
+  CUDAScopedContextProduce ctx{ctxState_};
 
   auto tmp = gpuAlgo_.getResults();
   ctx.emplace(iEvent, digiPutToken_, std::move(tmp.first));

diff --git a/RecoLocalTracker/SiPixelRecHits/plugins/SiPixelRecHitCUDA.cc b/RecoLocalTracker/SiPixelRecHits/plugins/SiPixelRecHitCUDA.cc
@@ -85,7 +85,7 @@ void SiPixelRecHitCUDA::produce(edm::StreamID streamID, edm::Event& iEvent, cons
   edm::Handle<CUDAProduct<SiPixelClustersCUDA>> hclusters;
   iEvent.getByToken(token_, hclusters);
 
-  CUDAScopedContext ctx{*hclusters};
+  CUDAScopedContextProduce ctx{*hclusters};
   auto const& clusters = ctx.get(*hclusters);
 
   edm::Handle<CUDAProduct<SiPixelDigisCUDA>> hdigis;