refactor: change direct submission interfaces to accept monitor fence

- unify Linux and Windows default settings
- unify override default code
- correct size estimation when fence is required
- call virtual function once for both estimation and dispatch

Related-To: NEO-8395

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

shared/source/direct_submission/direct_submission_hw.h

@@ -48,7 +48,7 @@ enum class DirectSubmissionSfenceMode : int32_t {
 
 namespace UllsDefaults {
 inline constexpr bool defaultDisableCacheFlush = true;
-inline constexpr bool defaultDisableMonitorFence = false;
+inline constexpr bool defaultDisableMonitorFence = true;
 } // namespace UllsDefaults
 
 struct BatchBuffer;
@@ -145,8 +145,8 @@ class DirectSubmissionHw {
 
     void setReturnAddress(void *returnCmd, uint64_t returnAddress);
 
-    void *dispatchWorkloadSection(BatchBuffer &batchBuffer);
-    size_t getSizeDispatch(bool relaxedOrderingSchedulerRequired, bool returnPtrsRequired);
+    void *dispatchWorkloadSection(BatchBuffer &batchBuffer, bool dispatchMonitorFence);
+    size_t getSizeDispatch(bool relaxedOrderingSchedulerRequired, bool returnPtrsRequired, bool dispatchMonitorFence);
 
     void dispatchPrefetchMitigation();
     size_t getSizePrefetchMitigation();

shared/source/direct_submission/direct_submission_hw.inl

@@ -51,6 +51,9 @@ DirectSubmissionHw<GfxFamily, Dispatcher>::DirectSubmissionHw(const DirectSubmis
 
     disableCacheFlush = UllsDefaults::defaultDisableCacheFlush;
     disableMonitorFence = UllsDefaults::defaultDisableMonitorFence;
+    if (DebugManager.flags.DirectSubmissionDisableMonitorFence.get() != -1) {
+        this->disableMonitorFence = DebugManager.flags.DirectSubmissionDisableMonitorFence.get();
+    }
 
     if (DebugManager.flags.DirectSubmissionMaxRingBuffers.get() != -1) {
         this->maxRingBufferCount = DebugManager.flags.DirectSubmissionMaxRingBuffers.get();
@@ -505,7 +508,7 @@ bool DirectSubmissionHw<GfxFamily, Dispatcher>::startRingBuffer() {
         startSize += RelaxedOrderingHelper::getSizeRegistersInit<GfxFamily>();
     }
 
-    size_t requiredSize = startSize + getSizeDispatch(false, false) + getSizeEnd(false);
+    size_t requiredSize = startSize + getSizeDispatch(false, false, dispatchMonitorFenceRequired(true)) + getSizeEnd(false);
     if (ringCommandStream.getAvailableSpace() < requiredSize) {
         switchRingBuffers();
     }
@@ -662,7 +665,7 @@ inline size_t DirectSubmissionHw<GfxFamily, Dispatcher>::getSizeEnd(bool relaxed
 }
 
 template <typename GfxFamily, typename Dispatcher>
-inline size_t DirectSubmissionHw<GfxFamily, Dispatcher>::getSizeDispatch(bool relaxedOrderingSchedulerRequired, bool returnPtrsRequired) {
+inline size_t DirectSubmissionHw<GfxFamily, Dispatcher>::getSizeDispatch(bool relaxedOrderingSchedulerRequired, bool returnPtrsRequired, bool dispatchMonitorFence) {
     size_t size = getSizeSemaphoreSection(relaxedOrderingSchedulerRequired);
     if (workloadMode == 0) {
         size += getSizeStartSection();
@@ -677,7 +680,7 @@ inline size_t DirectSubmissionHw<GfxFamily, Dispatcher>::getSizeDispatch(bool re
     if (!disableCacheFlush) {
         size += Dispatcher::getSizeCacheFlush(rootDeviceEnvironment);
     }
-    if (!disableMonitorFence) {
+    if (dispatchMonitorFence) {
         size += Dispatcher::getSizeMonitorFence(rootDeviceEnvironment);
     }
 
@@ -695,7 +698,7 @@ void DirectSubmissionHw<GfxFamily, Dispatcher>::updateRelaxedOrderingQueueSize(u
 }
 
 template <typename GfxFamily, typename Dispatcher>
-void *DirectSubmissionHw<GfxFamily, Dispatcher>::dispatchWorkloadSection(BatchBuffer &batchBuffer) {
+void *DirectSubmissionHw<GfxFamily, Dispatcher>::dispatchWorkloadSection(BatchBuffer &batchBuffer, bool dispatchMonitorFence) {
     void *currentPosition = ringCommandStream.getSpace(0);
 
     if (DebugManager.flags.DirectSubmissionPrintBuffers.get()) {
@@ -765,7 +768,7 @@ void *DirectSubmissionHw<GfxFamily, Dispatcher>::dispatchWorkloadSection(BatchBu
         Dispatcher::dispatchCacheFlush(ringCommandStream, this->rootDeviceEnvironment, gpuVaForMiFlush);
     }
 
-    if (!disableMonitorFence || this->dispatchMonitorFenceRequired(batchBuffer.hasStallingCmds)) {
+    if (dispatchMonitorFence) {
         TagData currentTagData = {};
         getTagAddressValue(currentTagData);
         Dispatcher::dispatchMonitorFence(ringCommandStream, currentTagData.tagAddress, currentTagData.tagValue, this->rootDeviceEnvironment,
@@ -935,8 +938,9 @@ bool DirectSubmissionHw<GfxFamily, Dispatcher>::dispatchCommandBuffer(BatchBuffe
     this->startRingBuffer();
 
     bool relaxedOrderingSchedulerWillBeNeeded = (this->relaxedOrderingSchedulerRequired || batchBuffer.hasRelaxedOrderingDependencies);
+    bool dispatchMonitorFence = this->dispatchMonitorFenceRequired(batchBuffer.hasStallingCmds);
 
-    size_t dispatchSize = getSizeDispatch(relaxedOrderingSchedulerWillBeNeeded, batchBuffer.hasRelaxedOrderingDependencies);
+    size_t dispatchSize = getSizeDispatch(relaxedOrderingSchedulerWillBeNeeded, batchBuffer.hasRelaxedOrderingDependencies, dispatchMonitorFence);
 
     if (this->copyCommandBufferIntoRing(batchBuffer)) {
         dispatchSize += (batchBuffer.stream->getUsed() - batchBuffer.startOffset) - 2 * getSizeStartSection();
@@ -965,7 +969,7 @@ bool DirectSubmissionHw<GfxFamily, Dispatcher>::dispatchCommandBuffer(BatchBuffe
 
     handleNewResourcesSubmission();
 
-    void *currentPosition = dispatchWorkloadSection(batchBuffer);
+    void *currentPosition = dispatchWorkloadSection(batchBuffer, dispatchMonitorFence);
 
     cpuCachelineFlush(currentPosition, dispatchSize);
     handleResidency();
@@ -1084,7 +1088,7 @@ inline GraphicsAllocation *DirectSubmissionHw<GfxFamily, Dispatcher>::switchRing
 
 template <typename GfxFamily, typename Dispatcher>
 bool DirectSubmissionHw<GfxFamily, Dispatcher>::dispatchMonitorFenceRequired(bool hasStallingCmds) {
-    return false;
+    return !this->disableMonitorFence;
 }
 
 template <typename GfxFamily, typename Dispatcher>

shared/source/direct_submission/linux/drm_direct_submission.inl

@@ -28,12 +28,6 @@ template <typename GfxFamily, typename Dispatcher>
 DrmDirectSubmission<GfxFamily, Dispatcher>::DrmDirectSubmission(const DirectSubmissionInputParams &inputParams)
     : DirectSubmissionHw<GfxFamily, Dispatcher>(inputParams) {
 
-    this->disableMonitorFence = true;
-
-    if (DebugManager.flags.DirectSubmissionDisableMonitorFence.get() != -1) {
-        this->disableMonitorFence = DebugManager.flags.DirectSubmissionDisableMonitorFence.get();
-    }
-
     if (DebugManager.flags.OverrideUserFenceStartValue.get() != -1) {
         this->completionFenceValue = static_cast<decltype(completionFenceValue)>(DebugManager.flags.OverrideUserFenceStartValue.get());
     }

shared/source/direct_submission/windows/wddm_direct_submission.inl

@@ -41,12 +41,6 @@ WddmDirectSubmission<GfxFamily, Dispatcher>::WddmDirectSubmission(const DirectSu
     if (this->miMemFenceRequired) {
         this->gpuVaForAdditionalSynchronizationWA = this->completionFenceAllocation->getGpuAddress() + 8u;
     }
-
-    this->disableMonitorFence = true;
-
-    if (DebugManager.flags.DirectSubmissionDisableMonitorFence.get() != -1) {
-        this->disableMonitorFence = DebugManager.flags.DirectSubmissionDisableMonitorFence.get();
-    }
 }
 
 template <typename GfxFamily, typename Dispatcher>
@@ -157,7 +151,7 @@ uint64_t WddmDirectSubmission<GfxFamily, Dispatcher>::updateTagValue(bool hasSta
 
 template <typename GfxFamily, typename Dispatcher>
 bool WddmDirectSubmission<GfxFamily, Dispatcher>::dispatchMonitorFenceRequired(bool hasStallingCmds) {
-    return hasStallingCmds;
+    return !this->disableMonitorFence || hasStallingCmds;
 }
 
 template <typename GfxFamily, typename Dispatcher>