[llvm] [libsycl] Initial prefetch implementation for queue (PR #212104)

[Kseniya Tikhomirova via llvm-commits]

================
@@ -199,6 +199,35 @@ QueueImpl::memcpy(void *Dest, const void *Src, std::size_t NumBytes,
   return createEvent();
 }
 
+EventImplPtr QueueImpl::prefetch(void *Ptr, std::size_t NumBytes,
+                                 const std::vector<EventImplPtr> &DepEvents) {
+  checkEventsPlatformMatch(DepEvents, MDevice.getPlatformImpl());
+
+  if (NumBytes == 0) {
+    handleEventDependencies(DepEvents);
+    return createEvent();
+  }
+  if (!Ptr) {
+    throw sycl::exception(sycl::make_error_code(sycl::errc::invalid),
+                          "Nullptr argument in prefetch operation");
+  }
+
+  auto [TargetDevice, IsHost] = getAllocDevice(Ptr);
+
+  std::size_t Count = 1;
+  const void *Mems[] = {Ptr};
+  const std::size_t Sizes[] = {NumBytes};
+
+  ol_mem_migration_flags_t Flag = (MDevice.isCPU())
----------------
KseniyaTikhomirova wrote:

liboffload has no "CPU" term so far, it is using a single "host" device (backend) to represent host side. I think we should stick to this paradigm and check if MDevice != "HostDevice" then. Or we can hardcode single value here since libsycl queues are always device queues.
@sergey-semenov what do you think about this?

https://github.com/llvm/llvm-project/pull/212104