Rename an overload of waitUntilComplete to waitForAllEngines

- Overloaded functions were quite different, they should have different names
- waitForLatestTaskCount function was redundant and hence its body was moved
inside waitForAllEngines

Signed-off-by: Maciej Dziuban <maciej.dziuban@intel.com>

opencl/source/command_queue/command_queue.cpp

@@ -187,15 +187,6 @@ bool CommandQueue::isCompleted(uint32_t gpgpuTaskCount, CopyEngineState bcsState
     return false;
 }
 
-void CommandQueue::waitForLatestTaskCount() {
-    TimestampPacketContainer nodesToRelease;
-    if (deferredTimestampPackets) {
-        deferredTimestampPackets->swapNodes(nodesToRelease);
-    }
-
-    waitUntilComplete(taskCount, this->bcsState.taskCount, flushStamp->peekStamp(), false);
-}
-
 void CommandQueue::waitUntilComplete(uint32_t gpgpuTaskCountToWait, uint32_t bcsTaskCountToWait, FlushStamp flushStampToWait, bool useQuickKmdSleep) {
     WAIT_ENTER()
 
@@ -918,13 +909,18 @@ void CommandQueue::aubCaptureHook(bool &blocking, bool &clearAllDependencies, co
     }
 }
 
-void CommandQueue::waitUntilComplete(bool blockedQueue, PrintfHandler *printfHandler) {
+void CommandQueue::waitForAllEngines(bool blockedQueue, PrintfHandler *printfHandler) {
     if (blockedQueue) {
         while (isQueueBlocked()) {
         }
     }
 
-    waitForLatestTaskCount();
+    TimestampPacketContainer nodesToRelease;
+    if (deferredTimestampPackets) {
+        deferredTimestampPackets->swapNodes(nodesToRelease);
+    }
+
+    waitUntilComplete(taskCount, this->bcsState.taskCount, flushStamp->peekStamp(), false);
 
     if (printfHandler) {
         printfHandler->printEnqueueOutput();

opencl/source/command_queue/command_queue.h

@@ -217,7 +217,7 @@ class CommandQueue : public BaseObject<_cl_command_queue> {
     MOCKABLE_VIRTUAL bool isQueueBlocked();
 
     MOCKABLE_VIRTUAL void waitUntilComplete(uint32_t gpgpuTaskCountToWait, uint32_t bcsTaskCountToWait, FlushStamp flushStampToWait, bool useQuickKmdSleep);
-    MOCKABLE_VIRTUAL void waitUntilComplete(bool blockedQueue, PrintfHandler *printfHandler);
+    MOCKABLE_VIRTUAL void waitForAllEngines(bool blockedQueue, PrintfHandler *printfHandler);
 
     static uint32_t getTaskLevelFromWaitList(uint32_t taskLevel,
                                              cl_uint numEventsInWaitList,
@@ -360,7 +360,6 @@ class CommandQueue : public BaseObject<_cl_command_queue> {
     MOCKABLE_VIRTUAL bool blitEnqueueImageAllowed(const size_t *origin, const size_t *region, const Image &image) const;
     void aubCaptureHook(bool &blocking, bool &clearAllDependencies, const MultiDispatchInfo &multiDispatchInfo);
     virtual bool obtainTimestampPacketForCacheFlush(bool isCacheFlushRequired) const = 0;
-    void waitForLatestTaskCount();
 
     Context *context = nullptr;
     ClDevice *device = nullptr;

opencl/source/command_queue/enqueue_common.h

@@ -409,7 +409,7 @@ void CommandQueueHw<GfxFamily>::enqueueHandler(Surface **surfacesForResidency,
     commandStreamRecieverOwnership.unlock();
 
     if (blocking) {
-        waitUntilComplete(blockQueue, (blockQueue ? nullptr : printfHandler.get()));
+        waitForAllEngines(blockQueue, (blockQueue ? nullptr : printfHandler.get()));
     }
     if (migratedMemory) {
         getGpgpuCommandStreamReceiver().flushBatchedSubmissions();
@@ -1249,7 +1249,7 @@ void CommandQueueHw<GfxFamily>::enqueueBlit(const MultiDispatchInfo &multiDispat
     commandStreamRecieverOwnership.unlock();
 
     if (blocking) {
-        waitUntilComplete(blockQueue, nullptr);
+        waitForAllEngines(blockQueue, nullptr);
     }
 }
 

opencl/source/command_queue/finish.h

@@ -19,12 +19,8 @@ cl_int CommandQueueHw<GfxFamily>::finish() {
         return CL_OUT_OF_RESOURCES;
     }
 
-    //as long as queue is blocked we need to stall.
+    // Stall until HW reaches taskCount on all its engines
-    while (isQueueBlocked())
+    waitForAllEngines(true, nullptr);
-        ;
-
-    // Stall until HW reaches CQ taskCount
-    waitForLatestTaskCount();
 
     return CL_SUCCESS;
 }

opencl/test/unit_test/command_stream/command_stream_receiver_flush_task_3_tests.cpp

@@ -1031,7 +1031,7 @@ HWTEST_F(CommandStreamReceiverFlushTaskTests, givenUpdateTaskCountFromWaitSetWhe
     mockCsr->useGpuIdleImplicitFlush = false;
     mockCsr->overrideDispatchPolicy(DispatchMode::BatchedDispatch);
 
-    commandQueue.waitUntilComplete(false, nullptr);
+    commandQueue.waitForAllEngines(false, nullptr);
 
     parseCommands<FamilyType>(mockCsr->getCS(4096u));
     auto itorPipeControl = find<typename FamilyType::PIPE_CONTROL *>(cmdList.begin(), cmdList.end());
@@ -1059,7 +1059,7 @@ HWTEST_F(CommandStreamReceiverFlushTaskTests, givenEnabledDirectSubmissionUpdate
     mockCsr->useGpuIdleImplicitFlush = false;
     mockCsr->overrideDispatchPolicy(DispatchMode::BatchedDispatch);
 
-    commandQueue.waitUntilComplete(false, nullptr);
+    commandQueue.waitForAllEngines(false, nullptr);
 
     parseCommands<FamilyType>(mockCsr->getCS(4096u));
     auto itorPipeControl = find<typename FamilyType::PIPE_CONTROL *>(cmdList.begin(), cmdList.end());

opencl/test/unit_test/program/printf_handler_tests.cpp

@@ -169,7 +169,7 @@ HWTEST_F(PrintfHandlerTests, givenPrintfHandlerWhenEnqueueIsBlockedThenDontUsePr
         using CommandQueueHw<FamilyType>::CommandQueueHw;
         using CommandQueueHw<FamilyType>::enqueueKernel;
 
-        void waitUntilComplete(bool blockedQueue, PrintfHandler *printfHandler) override {
+        void waitForAllEngines(bool blockedQueue, PrintfHandler *printfHandler) override {
             waitCalled = true;
             printfHandlerUsedForWait = printfHandler;
         }