Refactor hipStreamWaitEvent

- Null streams use same flow as non-null.
- Add HIP_SYNC_STREAM_WAIT
- Resolve null stream.

src/hip_hcc.cpp

@@ -88,6 +88,8 @@ int HIP_HOST_COHERENT = 1;
 
 int HIP_SYNC_HOST_ALLOC = 1;
 
+int HIP_SYNC_STREAM_WAIT = 0;
+
 
 #if (__hcc_workweek__ >= 17300)
 // Make sure we have required bug fix in HCC
@@ -1249,6 +1251,7 @@ void HipReadEnv()
 
     READ_ENV_I(release, HIP_SYNC_HOST_ALLOC, 0, "Sync before and after all host memory allocations.  May help stability");
     READ_ENV_I(release, HIP_SYNC_NULL_STREAM, 0, "Synchronize on host for null stream submissions");
+    READ_ENV_I(release, HIP_SYNC_STREAM_WAIT, 0, "hipStreamWaitEvent will synchronize to host");
 
 
     READ_ENV_I(release, HIP_HOST_COHERENT, 0, "If set, all host memory will be allocated as fine-grained system memory.  This allows threadfence_system to work but prevents host memory from being cached on GPU which may have performance impact.");

src/hip_hcc_internal.h

@@ -65,6 +65,7 @@ extern int HIP_HIDDEN_FREE_MEM;
 //---
 // Chicken bits for disabling functionality to work around potential issues:
 extern int HIP_SYNC_HOST_ALLOC;
+extern int HIP_SYNC_STREAM_WAIT;
 
 extern int HIP_SYNC_NULL_STREAM;
 

src/hip_stream.cpp

@@ -93,18 +93,15 @@ hipError_t hipStreamWaitEvent(hipStream_t stream, hipEvent_t event, unsigned int
 
     } else if (event->_state != hipEventStatusUnitialized) {
 
-        if (stream != hipStreamNull) {
-
+        if (HIP_SYNC_STREAM_WAIT || (HIP_SYNC_NULL_STREAM && (stream == 0))) {
+            // conservative wait on host for the specified event to complete:
+            event->locked_waitComplete((event->_flags & hipEventBlockingSync) ? hc::hcWaitModeBlocked : hc::hcWaitModeActive);
+        } else {
+            stream = ihipSyncAndResolveStream(stream);
             // This will user create_blocking_marker to wait on the specified queue.
             stream->locked_streamWaitEvent(event);
-
-        } else {
-            // TODO-hcc Convert to use create_blocking_marker(...) functionality.
-            // Currently we have a super-conservative version of this - block on host, and drain the queue.
-            // This should create a barrier packet in the target queue.
-            // TODO-HIP_SYNC_NULL_STREAM
-            stream->locked_wait();
         }
+
     } // else event not recorded, return immediately and don't create marker.
 
     return ihipLogStatus(e);