Fix for kernel synchronization

The completion future of a particular kernel is lost if there are
multiple kernels in the stream. This can cause a racing condition where
the signal associated with the unreferenced completion_future might get
released by hcc runtime.


[ROCm/clr commit: 6773a64b22]

projects/clr/hipamd/include/hcc_detail/hip_hcc.h

@@ -440,9 +440,13 @@ public:
     hc::accelerator_view        _av;
     unsigned                    _flags;
 
-private: // Critical Data.  THis MUST be accessed through LockedAccessor_StreamCrit_t
+private:
+    // Critical Data.  THis MUST be accessed through LockedAccessor_StreamCrit_t
     ihipStreamCritical_t        _criticalData;
 
+    // Array of dependency completion_future.
+    std::vector<hc::completion_future> _depFutures;
+
 private:
     void                        enqueueBarrier(hsa_queue_t* queue, ihipSignal_t *depSignal);
     void                        waitCopy(LockedAccessor_StreamCrit_t &crit, ihipSignal_t *signal);

projects/clr/hipamd/src/hip_hcc.cpp

@@ -181,6 +181,8 @@ void ihipStream_t::wait(LockedAccessor_StreamCrit_t &crit, bool assertQueueEmpty
     // Reset the stream to "empty" - next command will not set up an inpute dependency on any older signal.
     crit->_last_command_type = ihipCommandCopyH2D;
     crit->_last_copy_signal = NULL;
+
+    _depFutures.clear();
 }
 
 
@@ -428,6 +430,9 @@ int ihipStream_t::preCopyCommand(LockedAccessor_StreamCrit_t &crit, ihipSignal_t
             needSync = 1;
             hsa_signal_t *hsaSignal = (static_cast<hsa_signal_t*> (crit->_last_kernel_future.get_native_handle()));
             if (hsaSignal) {
+                // Keep reference to the kernel future in order to keep the
+                // dependent signal alive.
+                _depFutures.push_back(crit->_last_kernel_future);
                 *waitSignal = * hsaSignal;
             } else {
                 assert(0); // if NULL signal, and we return 1, hsa_amd_memory_copy_async will fail.  Confirm this never happens.