SWDEV-405983 - adding in HIP_LAUNCH_BLOCKING

Change-Id: I3f9c8a745099aab05155ebe910e727693961a02f


[ROCm/clr commit: 28f0daa34f]

projects/clr/rocclr/platform/command.cpp

@@ -412,7 +412,8 @@ NDRangeKernelCommand::NDRangeKernelCommand(HostQueue& queue, const EventWaitList
                                            uint32_t gridId, uint32_t numGrids,
                                            uint64_t prevGridSum, uint64_t allGridSum,
                                            uint32_t firstDevice, bool forceProfiling) :
-    Command(queue, CL_COMMAND_NDRANGE_KERNEL, eventWaitList, AMD_SERIALIZE_KERNEL),
+    Command(queue, CL_COMMAND_NDRANGE_KERNEL, eventWaitList, AMD_SERIALIZE_KERNEL |
+                                                            (HIP_LAUNCH_BLOCKING << 1)),
     kernel_(kernel),
     sizes_(sizes),
     sharedMemBytes_(sharedMemBytes),

projects/clr/rocclr/utils/flags.hpp

@@ -166,6 +166,9 @@ release(uint, AMD_SERIALIZE_KERNEL, 0,                                        \
 release(uint, AMD_SERIALIZE_COPY, 0,                                          \
         "Serialize copies, 0x1 = Wait for completion before enqueue"          \
         "0x2 = Wait for completion after enqueue 0x3 = both")                 \
+release(uint, HIP_LAUNCH_BLOCKING, 0,                                         \
+        "Serialize kernel enqueue 0x1 = Wait for completion after enqueue,"   \
+        "same as AMD_SERIALIZE_KERNEL=2")                                     \
 release(bool, PAL_ALWAYS_RESIDENT, false,                                     \
         "Force memory resources to become resident at allocation time")       \
 release(uint, HIP_HOST_COHERENT, 0,                                           \