feature: Dispatch monitor fence from wait

Resolves: NEO-8240
Related-To: NEO-8067

Signed-off-by: Lukasz Jobczyk <lukasz.jobczyk@intel.com>

shared/source/direct_submission/direct_submission_hw.h

@@ -96,6 +96,8 @@ class DirectSubmissionHw {
         return relaxedOrderingEnabled;
     }
 
+    virtual void flushMonitorFence(){};
+
   protected:
     static constexpr size_t prefetchSize = 8 * MemoryConstants::cacheLineSize;
     static constexpr size_t prefetchNoops = prefetchSize / sizeof(uint32_t);
@@ -110,6 +112,7 @@ class DirectSubmissionHw {
     size_t getSizeNewResourceHandler();
     virtual void handleStopRingBuffer(){};
     virtual void ensureRingCompletion(){};
+    void switchRingBuffersNeeded(size_t size);
     virtual uint64_t switchRingBuffers();
     virtual void handleSwitchRingBuffers() = 0;
     GraphicsAllocation *switchRingBuffersAllocations();

shared/source/direct_submission/direct_submission_hw.inl

@@ -960,9 +960,7 @@ bool DirectSubmissionHw<GfxFamily, Dispatcher>::dispatchCommandBuffer(BatchBuffe
         }
     }
 
-    if (ringCommandStream.getAvailableSpace() < requiredMinimalSize) {
-        switchRingBuffers();
-    }
+    this->switchRingBuffersNeeded(requiredMinimalSize);
 
     if (this->relaxedOrderingEnabled && batchBuffer.hasStallingCmds && this->relaxedOrderingSchedulerRequired) {
         dispatchRelaxedOrderingQueueStall();
@@ -1027,6 +1025,13 @@ bool DirectSubmissionHw<GfxFamily, Dispatcher>::isNewResourceHandleNeeded() {
     return newResourcesBound;
 }
 
+template <typename GfxFamily, typename Dispatcher>
+void DirectSubmissionHw<GfxFamily, Dispatcher>::switchRingBuffersNeeded(size_t size) {
+    if (this->ringCommandStream.getAvailableSpace() < size) {
+        this->switchRingBuffers();
+    }
+}
+
 template <typename GfxFamily, typename Dispatcher>
 inline uint64_t DirectSubmissionHw<GfxFamily, Dispatcher>::switchRingBuffers() {
     GraphicsAllocation *nextRingBuffer = switchRingBuffersAllocations();

shared/source/direct_submission/windows/wddm_direct_submission.h

@@ -23,6 +23,8 @@ class WddmDirectSubmission : public DirectSubmissionHw<GfxFamily, Dispatcher> {
 
     ~WddmDirectSubmission() override;
 
+    void flushMonitorFence() override;
+
   protected:
     bool allocateOsResources() override;
     bool submit(uint64_t gpuAddress, size_t size) override;
@@ -31,6 +33,7 @@ class WddmDirectSubmission : public DirectSubmissionHw<GfxFamily, Dispatcher> {
     void handleCompletionFence(uint64_t completionValue, MonitoredFence &fence);
     void ensureRingCompletion() override;
     void handleSwitchRingBuffers() override;
+    void handleStopRingBuffer() override;
     uint64_t updateTagValue() override;
     void getTagAddressValue(TagData &tagData) override;
     bool isCompleted(uint32_t ringBufferIndex) override;

shared/source/direct_submission/windows/wddm_direct_submission.inl

@@ -41,6 +41,10 @@ WddmDirectSubmission<GfxFamily, Dispatcher>::WddmDirectSubmission(const DirectSu
     if (this->miMemFenceRequired) {
         this->gpuVaForAdditionalSynchronizationWA = this->completionFenceAllocation->getGpuAddress() + 8u;
     }
+
+    if (DebugManager.flags.DirectSubmissionDisableMonitorFence.get() != -1) {
+        this->disableMonitorFence = DebugManager.flags.DirectSubmissionDisableMonitorFence.get();
+    }
 }
 
 template <typename GfxFamily, typename Dispatcher>
@@ -53,6 +57,31 @@ WddmDirectSubmission<GfxFamily, Dispatcher>::~WddmDirectSubmission() {
     wddm->getWddmInterface()->destroyMonitorFence(ringFence);
 }
 
+template <typename GfxFamily, typename Dispatcher>
+inline void WddmDirectSubmission<GfxFamily, Dispatcher>::flushMonitorFence() {
+    this->startRingBuffer();
+
+    size_t requiredMinimalSize = this->getSizeSemaphoreSection(false) +
+                                 Dispatcher::getSizeMonitorFence(this->rootDeviceEnvironment) +
+                                 this->getSizeNewResourceHandler() +
+                                 this->getSizeSwitchRingBufferSection() +
+                                 this->getSizeEnd(false);
+    this->switchRingBuffersNeeded(requiredMinimalSize);
+
+    this->handleNewResourcesSubmission();
+
+    TagData currentTagData = {};
+    this->getTagAddressValue(currentTagData);
+    Dispatcher::dispatchMonitorFence(this->ringCommandStream, currentTagData.tagAddress, currentTagData.tagValue, this->rootDeviceEnvironment, this->useNotifyForPostSync, this->partitionedMode, this->dcFlushRequired);
+
+    this->dispatchSemaphoreSection(this->currentQueueWorkCount + 1);
+    this->handleResidency();
+    this->unblockGpu();
+    this->currentQueueWorkCount++;
+
+    this->updateTagValue();
+}
+
 template <typename GfxFamily, typename Dispatcher>
 void WddmDirectSubmission<GfxFamily, Dispatcher>::ensureRingCompletion() {
     WddmDirectSubmission<GfxFamily, Dispatcher>::handleCompletionFence(ringFence.lastSubmittedFence, ringFence);
@@ -96,19 +125,38 @@ bool WddmDirectSubmission<GfxFamily, Dispatcher>::handleResidency() {
     return true;
 }
 
+template <typename GfxFamily, typename Dispatcher>
+void WddmDirectSubmission<GfxFamily, Dispatcher>::handleStopRingBuffer() {
+    if (this->disableMonitorFence) {
+        MonitoredFence &currentFence = osContextWin->getResidencyController().getMonitoredFence();
+        currentFence.lastSubmittedFence = currentFence.currentFenceValue;
+        currentFence.currentFenceValue++;
+    }
+}
+
 template <typename GfxFamily, typename Dispatcher>
 void WddmDirectSubmission<GfxFamily, Dispatcher>::handleSwitchRingBuffers() {
+    if (this->disableMonitorFence) {
+        MonitoredFence &currentFence = osContextWin->getResidencyController().getMonitoredFence();
+        currentFence.lastSubmittedFence = currentFence.currentFenceValue;
+        currentFence.currentFenceValue++;
+
+        this->ringBuffers[this->currentRingBuffer].completionFence = currentFence.lastSubmittedFence;
+    }
 }
 
 template <typename GfxFamily, typename Dispatcher>
 uint64_t WddmDirectSubmission<GfxFamily, Dispatcher>::updateTagValue() {
-    MonitoredFence &currentFence = osContextWin->getResidencyController().getMonitoredFence();
+    if (!this->disableMonitorFence) {
+        MonitoredFence &currentFence = osContextWin->getResidencyController().getMonitoredFence();
 
-    currentFence.lastSubmittedFence = currentFence.currentFenceValue;
-    currentFence.currentFenceValue++;
-    this->ringBuffers[this->currentRingBuffer].completionFence = currentFence.lastSubmittedFence;
+        currentFence.lastSubmittedFence = currentFence.currentFenceValue;
+        currentFence.currentFenceValue++;
+        this->ringBuffers[this->currentRingBuffer].completionFence = currentFence.lastSubmittedFence;
 
-    return currentFence.lastSubmittedFence;
+        return currentFence.lastSubmittedFence;
+    }
+    return 0ull;
 }
 
 template <typename GfxFamily, typename Dispatcher>