[SYCL] Fix case when host-task is used for queue with reused context (#2829)

Author: s-kanaev

sycl/source/detail/scheduler/commands.cpp

@@ -485,7 +485,7 @@ void Command::processDepEvent(EventImplPtr DepEvent, const DepDesc &Dep) {
 
   // 1. Async work is not supported for host device.
   // 2. The event handle can be null in case of, for example, alloca command,
-  //    which is currently synchrounious, so don't generate OpenCL event.
+  //    which is currently synchronous, so don't generate OpenCL event.
   //    Though, this event isn't host one as it's context isn't host one.
   if (DepEvent->is_host() || DepEvent->getHandleRef() == nullptr) {
     // call to waitInternal() is in waitForPreparedHostEvents() as it's called
@@ -2039,7 +2039,8 @@ cl_int ExecCGCommand::enqueueImp() {
             Req->MSYCLMemObj->MRecord->MAllocaCommands;
 
         for (AllocaCommandBase *AllocaCmd : AllocaCmds)
-          if (HostTask->MQueue == AllocaCmd->getQueue()) {
+          if (HostTask->MQueue->getContextImplPtr() ==
+              AllocaCmd->getQueue()->getContextImplPtr()) {
             auto MemArg =
                 reinterpret_cast<pi_mem>(AllocaCmd->getMemAllocation());
             ReqToMem.emplace_back(std::make_pair(Req, MemArg));

sycl/test/on-device/host-interop-task/interop-task.cpp

@@ -150,6 +150,46 @@ void test2() {
   }
 }
 
+// Same as above but with queue constructed out of context
+void test2_1() {
+  static constexpr int COUNT = 4;
+  buffer<int, 1> Buffer1{BUFFER_SIZE};
+  buffer<int, 1> Buffer2{BUFFER_SIZE};
+
+  auto Device = default_selector().select_device();
+  auto Context = context(Device);
+  // init the buffer with a'priori invalid data
+  {
+    queue Q(Context, Device);
+    init<int, -1, -2>(Buffer1, Buffer2, Q);
+  }
+
+  // Repeat a couple of times
+  for (size_t Idx = 0; Idx < COUNT; ++Idx) {
+    queue Q(Context, Device);
+    copy(Buffer1, Buffer2, Q);
+    modify(Buffer2, Q);
+    copy(Buffer2, Buffer1, Q);
+  }
+
+  {
+    auto Acc = Buffer1.get_access<mode::read>();
+
+    for (size_t Idx = 0; Idx < Acc.get_count(); ++Idx) {
+      std::cout << "First buffer [" << Idx << "] = " << Acc[Idx] << std::endl;
+      assert((Acc[Idx] == COUNT - 1) && "Invalid data in the first buffer");
+    }
+  }
+  {
+    auto Acc = Buffer2.get_access<mode::read>();
+
+    for (size_t Idx = 0; Idx < Acc.get_count(); ++Idx) {
+      std::cout << "Second buffer [" << Idx << "] = " << Acc[Idx] << std::endl;
+      assert((Acc[Idx] == COUNT - 1) && "Invalid data in the second buffer");
+    }
+  }
+}
+
 // A test that does a clEnqueueWait inside the interop scope, for an event
 // captured outside the command group. The OpenCL event can be set after the
 // command group finishes. Must not deadlock according to implementation and
@@ -251,6 +291,7 @@ void test6() {
 int main() {
   test1();
   test2();
+  test2_1();
   test3();
   test4();
   test5();