refactor: add batch buffer attribute for explicit monitor fence dispatch

Related-To: NEO-8395

Signed-off-by: Zbigniew Zdanowicz <zbigniew.zdanowicz@intel.com>

level_zero/core/source/cmdqueue/cmdqueue.cpp

@@ -123,7 +123,7 @@ NEO::SubmissionStatus CommandQueueImp::submitBatchBuffer(size_t offset, NEO::Res
 
     NEO::BatchBuffer batchBuffer(this->startingCmdBuffer->getGraphicsAllocation(), offset, 0, 0, nullptr, false,
                                  NEO::QueueThrottle::HIGH, NEO::QueueSliceCount::defaultSliceCount,
-                                 this->startingCmdBuffer->getUsed(), this->startingCmdBuffer, endingCmdPtr, csr->getNumClients(), true, false);
+                                 this->startingCmdBuffer->getUsed(), this->startingCmdBuffer, endingCmdPtr, csr->getNumClients(), true, false, true);
     batchBuffer.disableFlatRingBuffer = true;
 
     if (this->startingCmdBuffer != &this->commandStream) {

shared/source/command_stream/command_stream_receiver_hw_base.inl

@@ -237,7 +237,7 @@ CompletionStamp CommandStreamReceiverHw<GfxFamily>::flushBcsTask(LinearStream &c
     BatchBuffer batchBuffer{streamToSubmit.getGraphicsAllocation(), startOffset, 0, taskStartAddress, nullptr,
                             false, QueueThrottle::MEDIUM, NEO::QueueSliceCount::defaultSliceCount,
                             streamToSubmit.getUsed(), &streamToSubmit, bbEndLocation, this->getNumClients(), (submitCSR || dispatchBcsFlags.hasStallingCmds),
-                            dispatchBcsFlags.hasRelaxedOrderingDependencies};
+                            dispatchBcsFlags.hasRelaxedOrderingDependencies, false};
 
     updateStreamTaskCount(streamToSubmit, taskCount + 1);
 
@@ -629,7 +629,7 @@ CompletionStamp CommandStreamReceiverHw<GfxFamily>::flushTask(
     BatchBuffer batchBuffer{streamToSubmit.getGraphicsAllocation(), startOffset, chainedBatchBufferStartOffset, taskStartAddress, chainedBatchBuffer,
                             dispatchFlags.lowPriority, dispatchFlags.throttle, dispatchFlags.sliceCount,
                             streamToSubmit.getUsed(), &streamToSubmit, bbEndLocation, this->getNumClients(), (submitCSR || dispatchFlags.hasStallingCmds || hasStallingCmdsOnTaskStream),
-                            dispatchFlags.hasRelaxedOrderingDependencies};
+                            dispatchFlags.hasRelaxedOrderingDependencies, dispatchFlags.blocking};
 
     updateStreamTaskCount(streamToSubmit, taskCount + 1);
 
@@ -1225,7 +1225,7 @@ TaskCountType CommandStreamReceiverHw<GfxFamily>::flushBcsTask(const BlitPropert
     uint64_t taskStartAddress = commandStream.getGpuBase() + commandStreamStart;
 
     BatchBuffer batchBuffer{commandStream.getGraphicsAllocation(), commandStreamStart, 0, taskStartAddress, nullptr, false, QueueThrottle::MEDIUM, QueueSliceCount::defaultSliceCount,
-                            commandStream.getUsed(), &commandStream, endingCmdPtr, this->getNumClients(), hasStallingCmds, isRelaxedOrderingDispatch};
+                            commandStream.getUsed(), &commandStream, endingCmdPtr, this->getNumClients(), hasStallingCmds, isRelaxedOrderingDispatch, blocking};
 
     updateStreamTaskCount(commandStream, newTaskCount);
 
@@ -1347,7 +1347,7 @@ SubmissionStatus CommandStreamReceiverHw<GfxFamily>::flushSmallTask(LinearStream
 
     BatchBuffer batchBuffer{commandStreamTask.getGraphicsAllocation(), commandStreamStartTask, 0, taskStartAddress,
                             nullptr, false, QueueThrottle::MEDIUM, QueueSliceCount::defaultSliceCount,
-                            commandStreamTask.getUsed(), &commandStreamTask, endingCmdPtr, this->getNumClients(), true, false};
+                            commandStreamTask.getUsed(), &commandStreamTask, endingCmdPtr, this->getNumClients(), true, false, true};
 
     this->latestSentTaskCount = taskCount + 1;
     auto submissionStatus = flushHandler(batchBuffer, getResidencyAllocations());
@@ -2159,7 +2159,7 @@ CompletionStamp CommandStreamReceiverHw<GfxFamily>::handleImmediateFlushSendBatc
     BatchBuffer batchBuffer{streamToSubmit.getGraphicsAllocation(), startOffset, chainedBatchBufferStartOffset, taskStartAddress, chainedBatchBuffer,
                             immediateLowPriority, immediateThrottle, immediateSliceCount,
                             streamToSubmit.getUsed(), &streamToSubmit, flushData.endPtr, this->getNumClients(), hasStallingCmds,
-                            dispatchFlags.hasRelaxedOrderingDependencies};
+                            dispatchFlags.hasRelaxedOrderingDependencies, dispatchFlags.blockingAppend};
     updateStreamTaskCount(streamToSubmit, taskCount + 1);
 
     auto submissionStatus = flushHandler(batchBuffer, this->getResidencyAllocations());

shared/source/command_stream/submissions_aggregator.cpp

@@ -97,13 +97,13 @@ NEO::BatchBuffer::BatchBuffer(GraphicsAllocation *commandBufferAllocation, size_
                               size_t chainedBatchBufferStartOffset, uint64_t taskStartAddress, GraphicsAllocation *chainedBatchBuffer,
                               bool lowPriority, QueueThrottle throttle, uint64_t sliceCount,
                               size_t usedSize, LinearStream *stream, void *endCmdPtr, uint32_t numCsrClients, bool hasStallingCmds,
-                              bool hasRelaxedOrderingDependencies)
+                              bool hasRelaxedOrderingDependencies, bool dispatchMonitorFence)
     : commandBufferAllocation(commandBufferAllocation), startOffset(startOffset),
       chainedBatchBufferStartOffset(chainedBatchBufferStartOffset), taskStartAddress(taskStartAddress), chainedBatchBuffer(chainedBatchBuffer),
       lowPriority(lowPriority),
       throttle(throttle), sliceCount(sliceCount),
       usedSize(usedSize), stream(stream), endCmdPtr(endCmdPtr), numCsrClients(numCsrClients), hasStallingCmds(hasStallingCmds),
-      hasRelaxedOrderingDependencies(hasRelaxedOrderingDependencies) {}
+      hasRelaxedOrderingDependencies(hasRelaxedOrderingDependencies), dispatchMonitorFence(dispatchMonitorFence) {}
 
 NEO::CommandBuffer::CommandBuffer(Device &device) : device(device) {
     flushStamp.reset(new FlushStampTracker(false));

shared/source/command_stream/submissions_aggregator.h

@@ -33,7 +33,8 @@ struct BatchBuffer {
                 void *endCmdPtr,
                 uint32_t numCsrClients,
                 bool hasStallingCmds,
-                bool hasRelaxedOrderingDependencies);
+                bool hasRelaxedOrderingDependencies,
+                bool dispatchMonitorFence);
     BatchBuffer() {}
     GraphicsAllocation *commandBufferAllocation = nullptr;
     size_t startOffset = 0u;
@@ -55,6 +56,7 @@ struct BatchBuffer {
     bool hasRelaxedOrderingDependencies = false;
     bool ringBufferRestartRequest = false;
     bool disableFlatRingBuffer = false;
+    bool dispatchMonitorFence = false;
 };
 
 struct CommandBuffer : public IDNode<CommandBuffer> {

shared/test/common/helpers/batch_buffer_helper.h

@@ -26,7 +26,8 @@ struct BatchBufferHelper {
             nullptr,                            // endCmdPtr
             0,                                  // numCsrClients
             false,                              // hasStallingCmds
-            false                               // hasRelaxedOrderingDependencies
+            false,                              // hasRelaxedOrderingDependencies
+            false                               // dispatchMonitorFence
         );
     }