Move flush buffer initialization on device into __init

Signed-off-by: Mikhail Lychkov <[email protected]>

Author: mlychkov

sycl/include/CL/sycl/stream.hpp

@@ -841,6 +841,7 @@ class __SYCL_EXPORT stream {
     // overhead on FPGA target. That is why use global atomic variable to
     // calculate offsets.
     WIOffset = GlobalOffset[1].fetch_add(FlushBufferSize);
+    SetFlushBufOffset(GlobalFlushBuf, WIOffset, 0);
   }
 
   void __finalize() {

sycl/source/detail/scheduler/scheduler.cpp

@@ -75,9 +75,11 @@ EventImplPtr Scheduler::addCG(std::unique_ptr<detail::CG> CommandGroup,
 
   if (IsKernel) {
     Streams = ((CGExecKernel *)CommandGroup.get())->getStreams();
-    // Initializing stream's flush buffer on the host side.
-    for (auto StreamImplPtr : Streams) {
-      StreamImplPtr->fill(Queue);
+    if (Queue->is_host()) {
+      // Initializing stream's flush buffer on the host side.
+      for (auto StreamImplPtr : Streams) {
+        StreamImplPtr->fill(Queue);
+      }
     }
   }