[SYCL][Graph] Makes command graph functions thread-safe (#265)

* [SYCL][Graph] Makes command graph functions thread-safe

Addresses comments made on the first PR commit.
Mutexes are now added to Graph implementation entry points
instead of end points as was the case in the previous commit.
Adds "build_pthread_inc" lit test macro to facilitate the
compilation of the threading tests.
Removes std::barrier (std-20) dependency in threading tests.

Addresses Issue: #85

* [SYCL][Graph] Makes command graph functions thread-safe

Moves threading tests that do not require a device to run to unitests

* Update sycl/source/detail/graph_impl.cpp

Co-authored-by: Ben Tracy <[email protected]>

* [SYCL][Graph] Makes command graph functions thread-safe

Adds some comments.

* Update sycl/source/handler.cpp

Co-authored-by: Pablo Reble <[email protected]>

* Update sycl/source/detail/graph_impl.hpp

Co-authored-by: Ewan Crawford <[email protected]>

* [SYCL][Graph] Makes command graph functions thread-safe

Adds dedidacted sub-class to unitests for multi-threading unitests

* [SYCL][Graph] Makes command graph functions thread-safe

Adds comments

* [SYCL][Graph] thread-safe: bug fix after rebase

---------

Co-authored-by: Ben Tracy <[email protected]>
Co-authored-by: Pablo Reble <[email protected]>
Co-authored-by: Ewan Crawford <[email protected]>

Author: 4 people

sycl/source/detail/graph_impl.cpp

@@ -220,7 +220,6 @@ graph_impl::add(sycl::detail::CG::CGTYPE CGType,
       checkForRequirement(Req, NodePtr, UniqueDeps);
     }
   }
-
   // Add any nodes specified by event dependencies into the dependency list
   for (auto Dep : CommandGroup->getEvents()) {
     if (auto NodeImpl = MEventsMap.find(Dep); NodeImpl != MEventsMap.end()) {
@@ -474,6 +473,8 @@ void exec_graph_impl::createCommandBuffers(sycl::device Device) {
 }
 
 exec_graph_impl::~exec_graph_impl() {
+  WriteLock Lock(MMutex);
+
   // clear all recording queue if not done before (no call to end_recording)
   MGraphImpl->clearQueues();
 
@@ -499,6 +500,8 @@ exec_graph_impl::~exec_graph_impl() {
 sycl::event
 exec_graph_impl::enqueue(const std::shared_ptr<sycl::detail::queue_impl> &Queue,
                          sycl::detail::CG::StorageInitHelper CGData) {
+  WriteLock Lock(MMutex);
+
   auto CreateNewEvent([&]() {
     auto NewEvent = std::make_shared<sycl::detail::event_impl>(Queue);
     NewEvent->setContextImpl(Queue->getContextImplPtr());
@@ -612,6 +615,7 @@ node modifiable_command_graph::addImpl(const std::vector<node> &Deps) {
     DepImpls.push_back(sycl::detail::getSyclObjImpl(D));
   }
 
+  graph_impl::WriteLock Lock(impl->MMutex);
   std::shared_ptr<detail::node_impl> NodeImpl = impl->add(DepImpls);
   return sycl::detail::createSyclObjFromImpl<node>(NodeImpl);
 }
@@ -624,6 +628,7 @@ node modifiable_command_graph::addImpl(std::function<void(handler &)> CGF,
     DepImpls.push_back(sycl::detail::getSyclObjImpl(D));
   }
 
+  graph_impl::WriteLock Lock(impl->MMutex);
   std::shared_ptr<detail::node_impl> NodeImpl =
       impl->add(impl, CGF, {}, DepImpls);
   return sycl::detail::createSyclObjFromImpl<node>(NodeImpl);
@@ -635,6 +640,7 @@ void modifiable_command_graph::make_edge(node &Src, node &Dest) {
   std::shared_ptr<detail::node_impl> ReceiverImpl =
       sycl::detail::getSyclObjImpl(Dest);
 
+  graph_impl::WriteLock Lock(impl->MMutex);
   impl->makeEdge(SenderImpl, ReceiverImpl);
 }
 
@@ -666,6 +672,7 @@ bool modifiable_command_graph::begin_recording(queue &RecordingQueue) {
 
   if (QueueImpl->getCommandGraph() == nullptr) {
     QueueImpl->setCommandGraph(impl);
+    graph_impl::WriteLock Lock(impl->MMutex);
     impl->addQueue(QueueImpl);
     return true;
   }
@@ -687,12 +694,16 @@ bool modifiable_command_graph::begin_recording(
   return QueueStateChanged;
 }
 
-bool modifiable_command_graph::end_recording() { return impl->clearQueues(); }
+bool modifiable_command_graph::end_recording() {
+  graph_impl::WriteLock Lock(impl->MMutex);
+  return impl->clearQueues();
+}
 
 bool modifiable_command_graph::end_recording(queue &RecordingQueue) {
   auto QueueImpl = sycl::detail::getSyclObjImpl(RecordingQueue);
   if (QueueImpl->getCommandGraph() == impl) {
     QueueImpl->setCommandGraph(nullptr);
+    graph_impl::WriteLock Lock(impl->MMutex);
     impl->removeQueue(QueueImpl);
     return true;
   }
@@ -719,6 +730,9 @@ executable_command_graph::executable_command_graph(
     const std::shared_ptr<detail::graph_impl> &Graph, const sycl::context &Ctx)
     : MTag(rand()),
       impl(std::make_shared<detail::exec_graph_impl>(Ctx, Graph)) {
+  // Graph is read and written in this scope so we lock
+  // this graph with full priviledges.
+  graph_impl::WriteLock Lock(Graph->MMutex);
   finalizeImpl(); // Create backend representation for executable graph
 }