Include product instance name for the data product copied implicitly …

…to host
cms-sw · Feb 17, 2023 · 36d6142 · 36d6142
1 parent 150d6d1
commit 36d6142
Show file tree

Hide file tree

Showing 3 changed files with 27 additions and 8 deletions.
diff --git a/HeterogeneousCore/AlpakaCore/interface/alpaka/ProducerBase.h b/HeterogeneousCore/AlpakaCore/interface/alpaka/ProducerBase.h
@@ -78,9 +78,10 @@ namespace ALPAKA_ACCELERATOR_NAMESPACE {
     // can think of it later if really needed
     template <typename TProduct, typename TToken, edm::Transition Tr>
     edm::EDPutTokenT<TToken> deviceProduces(std::string instanceName) {
-      edm::EDPutTokenT<TToken> token = Base::template produces<TToken, Tr>(std::move(instanceName));
-
-      if constexpr (not detail::useProductDirectly<TProduct>) {
+      if constexpr (detail::useProductDirectly<TProduct>) {
+        return Base::template produces<TToken, Tr>(std::move(instanceName));
+      } else {
+        edm::EDPutTokenT<TToken> token = Base::template produces<TToken, Tr>(instanceName);
         this->registerTransformAsync(
             token,
             [](TToken const& deviceProduct, edm::WaitingTaskWithArenaHolder holder) {
@@ -100,9 +101,10 @@ namespace ALPAKA_ACCELERATOR_NAMESPACE {
               // Wrap possibly move-only type into a copyable type
               return std::make_shared<TplType>(std::move(productOnHost), sentry.finish());
             },
-            [](auto tplPtr) { return std::move(std::get<0>(*tplPtr)); });
+            [](auto tplPtr) { return std::move(std::get<0>(*tplPtr)); },
+            std::move(instanceName));
+        return token;
       }
-      return token;
     }
   };
 

diff --git a/HeterogeneousCore/AlpakaTest/plugins/alpaka/TestAlpakaStreamProducer.cc b/HeterogeneousCore/AlpakaTest/plugins/alpaka/TestAlpakaStreamProducer.cc
@@ -18,7 +18,8 @@ namespace ALPAKA_ACCELERATOR_NAMESPACE {
    * This class demonstrates a stream EDProducer that
    * - consumes a host EDProduct
    * - consumes a device ESProduct
-   * - produces a device EDProduct (that can get transferred to host automatically)
+   * - produces a device EDProduct (that gets transferred to host automatically if needed)
+   * - optionally uses a product instance label
    */
   class TestAlpakaStreamProducer : public stream::EDProducer<> {
   public:
@@ -27,7 +28,7 @@ namespace ALPAKA_ACCELERATOR_NAMESPACE {
               EDM_STRINGIZE(ALPAKA_ACCELERATOR_NAMESPACE))} {
       getToken_ = consumes(config.getParameter<edm::InputTag>("source"));
       esToken_ = esConsumes();
-      devicePutToken_ = produces();
+      devicePutToken_ = produces(config.getParameter<std::string>("productInstanceName"));
     }
 
     void produce(device::Event& iEvent, device::EventSetup const& iSetup) override {
@@ -45,6 +46,7 @@ namespace ALPAKA_ACCELERATOR_NAMESPACE {
     static void fillDescriptions(edm::ConfigurationDescriptions& descriptions) {
       edm::ParameterSetDescription desc;
       desc.add<edm::InputTag>("source");
+      desc.add<std::string>("productInstanceName", "");
 
       edm::ParameterSetDescription psetSize;
       psetSize.add<int32_t>("alpaka_serial_sync");

diff --git a/HeterogeneousCore/AlpakaTest/test/testAlpakaModules_cfg.py b/HeterogeneousCore/AlpakaTest/test/testAlpakaModules_cfg.py
@@ -71,6 +71,14 @@
         alpaka_cuda_async = cms.int32(25)
     )
 )
+process.alpakaStreamInstanceProducer = cms.EDProducer("TestAlpakaStreamProducer@alpaka",
+    source = cms.InputTag("intProduct"),
+    productInstanceName = cms.string("testInstance"),
+    size = cms.PSet(
+        alpaka_serial_sync = cms.int32(6),
+        alpaka_cuda_async = cms.int32(36)
+    )
+)
 process.alpakaStreamSynchronizingProducer = cms.EDProducer("TestAlpakaStreamSynchronizingProducer@alpaka",
     source = cms.InputTag("alpakaGlobalProducer")
 )
@@ -83,19 +91,24 @@
     source = cms.InputTag("alpakaStreamProducer"),
     expectSize = cms.int32(5)
 )
+process.alpakaStreamInstanceConsumer = cms.EDAnalyzer("TestAlpakaAnalyzer",
+    source = cms.InputTag("alpakaStreamInstanceProducer", "testInstance"),
+    expectSize = cms.int32(6)
+)
 process.alpakaStreamSynchronizingConsumer = cms.EDAnalyzer("TestAlpakaAnalyzer",
     source = cms.InputTag("alpakaStreamSynchronizingProducer"),
     expectSize = cms.int32(10)
 )
 
 if args.moduleBackend != "":
     for name in ["ESProducerA", "ESProducerB", "ESProducerC", "ESProducerD",
-                 "GlobalProducer", "StreamProducer", "StreamSynchronizingProducer"]:
+                 "GlobalProducer", "StreamProducer", "StreamInstanceProducer", "StreamSynchronizingProducer"]:
         mod = getattr(process, "alpaka"+name)
         mod.alpaka = cms.untracked.PSet(backend = cms.untracked.string(args.moduleBackend))
 if args.expectBackend == "cuda_async":
     process.alpakaGlobalConsumer.expectSize = 20
     process.alpakaStreamConsumer.expectSize = 25
+    process.alpakaStreamInstanceConsumer.expectSize = 36
     process.alpakaStreamSynchronizingConsumer.expectSize = 20
 
 process.output = cms.OutputModule('PoolOutputModule',
@@ -112,11 +125,13 @@
     process.intProduct,
     process.alpakaGlobalProducer,
     process.alpakaStreamProducer,
+    process.alpakaStreamInstanceProducer,
     process.alpakaStreamSynchronizingProducer
 )
 process.p = cms.Path(
     process.alpakaGlobalConsumer+
     process.alpakaStreamConsumer+
+    process.alpakaStreamInstanceConsumer+
     process.alpakaStreamSynchronizingConsumer,
     process.t
 )