[SYCL] Join thread pool threads on Scheduler destruction (intel#3090)

Threads from default queue thread pool can access scheduler after its destruction has begun.
To avoid such illegal access, wait for all threads to join in scheduler destructor.

Author: Alexander Batashev

sycl/source/detail/queue_impl.hpp

@@ -366,6 +366,12 @@ class queue_impl {
     return *MHostTaskThreadPool;
   }
 
+  void stopThreadPool() {
+    if (MHostTaskThreadPool) {
+      MHostTaskThreadPool->finishAndWait();
+    }
+  }
+
   /// Gets the native handle of the SYCL queue.
   ///
   /// \return a native handle.

sycl/source/detail/scheduler/scheduler.cpp

@@ -306,6 +306,7 @@ Scheduler::Scheduler() {
 }
 
 Scheduler::~Scheduler() {
+  DefaultHostQueue->stopThreadPool();
   // By specification there are several possible sync points: buffer
   // destruction, wait() method of a queue or event. Stream doesn't introduce
   // any synchronization point. It is guaranteed that stream is flushed and

sycl/unittests/scheduler/StreamInitDependencyOnHost.cpp

@@ -22,6 +22,18 @@ class MockHandler : public sycl::handler {
     static_cast<sycl::handler *>(this)->MCGType = Type;
   }
 
+  template <typename KernelType, typename ArgType, int Dims,
+            typename KernelName>
+  void setHostKernel(KernelType Kernel) {
+    static_cast<sycl::handler *>(this)->MHostKernel.reset(
+        new sycl::detail::HostKernel<KernelType, ArgType, Dims, KernelName>(
+            Kernel));
+  }
+
+  template <int Dims> void setNDRangeDesc(sycl::nd_range<Dims> Range) {
+    static_cast<sycl::handler *>(this)->MNDRDesc.set(std::move(Range));
+  }
+
   void addStream(const detail::StreamImplPtr &Stream) {
     sycl::handler::addStream(Stream);
   }
@@ -87,6 +99,13 @@ TEST_F(SchedulerTest, StreamInitDependencyOnHost) {
   MockHandler MockCGH(HQueueImpl, true);
   MockCGH.setType(detail::CG::KERNEL);
 
+  auto EmptyKernel = [](sycl::nd_item<1>) {};
+  MockCGH
+      .setHostKernel<decltype(EmptyKernel), sycl::nd_item<1>, 1, class Empty>(
+          EmptyKernel);
+  MockCGH.setNDRangeDesc(
+      sycl::nd_range<1>{sycl::range<1>{1}, sycl::range<1>{1}});
+
   // Emulating construction of stream object inside command group
   detail::StreamImplPtr StreamImpl =
       std::make_shared<detail::stream_impl>(1024, 200, MockCGH);