performance: Optimize ULLS start on submit path

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

shared/source/direct_submission/direct_submission_hw.h

@@ -82,8 +82,6 @@ class DirectSubmissionHw {
 
     MOCKABLE_VIRTUAL bool stopRingBuffer(bool blocking);
 
-    bool startRingBuffer();
-
     MOCKABLE_VIRTUAL bool dispatchCommandBuffer(BatchBuffer &batchBuffer, FlushStampTracker &flushStamp);
     uint32_t getDispatchErrorCode();
 
@@ -121,6 +119,7 @@ class DirectSubmissionHw {
     virtual bool dispatchMonitorFenceRequired(bool requireMonitorFence);
     virtual void getTagAddressValue(TagData &tagData) = 0;
     void unblockGpu();
+    bool submitCommandBufferToGpu(bool needStart, uint64_t gpuAddress, size_t size);
     bool copyCommandBufferIntoRing(BatchBuffer &batchBuffer);
 
     void cpuCachelineFlush(void *ptr, size_t size);
@@ -135,6 +134,9 @@ class DirectSubmissionHw {
     void dispatchStartSection(uint64_t gpuStartAddress);
     size_t getSizeStartSection();
 
+    size_t getUllsStateSize();
+    void dispatchUllsState();
+
     void dispatchSwitchRingBufferSection(uint64_t nextBufferGpuAddress);
     size_t getSizeSwitchRingBufferSection();
 

shared/source/direct_submission/direct_submission_hw.inl

@@ -499,55 +499,6 @@ bool DirectSubmissionHw<GfxFamily, Dispatcher>::initialize(bool submitOnInit, bo
     return ret;
 }
 
-template <typename GfxFamily, typename Dispatcher>
-bool DirectSubmissionHw<GfxFamily, Dispatcher>::startRingBuffer() {
-    if (ringStart) {
-        return true;
-    }
-
-    size_t startSize = getSizeSemaphoreSection(false);
-    if (!this->partitionConfigSet) {
-        startSize += getSizePartitionRegisterConfigurationSection();
-    }
-    if (this->miMemFenceRequired && !this->systemMemoryFenceAddressSet) {
-        startSize += getSizeSystemMemoryFenceAddress();
-    }
-    if (this->relaxedOrderingEnabled && !this->relaxedOrderingInitialized) {
-        startSize += RelaxedOrderingHelper::getSizeRegistersInit<GfxFamily>();
-    }
-
-    size_t requiredSize = startSize + getSizeDispatch(false, false, dispatchMonitorFenceRequired(true)) + getSizeEnd(false);
-    if (ringCommandStream.getAvailableSpace() < requiredSize) {
-        switchRingBuffers(nullptr);
-    }
-    uint64_t gpuStartVa = ringCommandStream.getCurrentGpuAddressPosition();
-
-    if (!this->partitionConfigSet) {
-        dispatchPartitionRegisterConfiguration();
-        this->partitionConfigSet = true;
-    }
-
-    if (this->miMemFenceRequired && !this->systemMemoryFenceAddressSet) {
-        dispatchSystemMemoryFenceAddress();
-        this->systemMemoryFenceAddressSet = true;
-    }
-
-    if (this->relaxedOrderingEnabled && !this->relaxedOrderingInitialized) {
-        preinitializeRelaxedOrderingSections();
-        dispatchStaticRelaxedOrderingScheduler();
-        initRelaxedOrderingRegisters();
-
-        this->relaxedOrderingInitialized = true;
-    }
-
-    currentQueueWorkCount++;
-    dispatchSemaphoreSection(currentQueueWorkCount);
-
-    ringStart = submit(gpuStartVa, startSize);
-
-    return ringStart;
-}
-
 template <typename GfxFamily, typename Dispatcher>
 bool DirectSubmissionHw<GfxFamily, Dispatcher>::stopRingBuffer(bool blocking) {
     if (!ringStart) {
@@ -940,15 +891,46 @@ bool DirectSubmissionHw<GfxFamily, Dispatcher>::copyCommandBufferIntoRing(BatchB
     return ret;
 }
 
+template <typename GfxFamily, typename Dispatcher>
+size_t DirectSubmissionHw<GfxFamily, Dispatcher>::getUllsStateSize() {
+    size_t startSize = 0u;
+    if (!this->partitionConfigSet) {
+        startSize += getSizePartitionRegisterConfigurationSection();
+    }
+    if (this->miMemFenceRequired && !this->systemMemoryFenceAddressSet) {
+        startSize += getSizeSystemMemoryFenceAddress();
+    }
+    if (this->relaxedOrderingEnabled && !this->relaxedOrderingInitialized) {
+        startSize += RelaxedOrderingHelper::getSizeRegistersInit<GfxFamily>();
+    }
+    return startSize;
+}
+
+template <typename GfxFamily, typename Dispatcher>
+void DirectSubmissionHw<GfxFamily, Dispatcher>::dispatchUllsState() {
+    if (!this->partitionConfigSet) {
+        dispatchPartitionRegisterConfiguration();
+        this->partitionConfigSet = true;
+    }
+    if (this->miMemFenceRequired && !this->systemMemoryFenceAddressSet) {
+        dispatchSystemMemoryFenceAddress();
+        this->systemMemoryFenceAddressSet = true;
+    }
+    if (this->relaxedOrderingEnabled && !this->relaxedOrderingInitialized) {
+        preinitializeRelaxedOrderingSections();
+        dispatchStaticRelaxedOrderingScheduler();
+        initRelaxedOrderingRegisters();
+
+        this->relaxedOrderingInitialized = true;
+    }
+}
+
 template <typename GfxFamily, typename Dispatcher>
 bool DirectSubmissionHw<GfxFamily, Dispatcher>::dispatchCommandBuffer(BatchBuffer &batchBuffer, FlushStampTracker &flushStamp) {
     if (batchBuffer.ringBufferRestartRequest) {
         this->stopRingBuffer(false);
     }
 
-    if (!this->startRingBuffer()) {
-        return false;
-    }
     lastSubmittedThrottle = batchBuffer.throttle;
     bool relaxedOrderingSchedulerWillBeNeeded = (this->relaxedOrderingSchedulerRequired || batchBuffer.hasRelaxedOrderingDependencies);
     bool inputRequiredMonitorFence = false;
@@ -959,7 +941,7 @@ bool DirectSubmissionHw<GfxFamily, Dispatcher>::dispatchCommandBuffer(BatchBuffe
     }
     bool dispatchMonitorFence = this->dispatchMonitorFenceRequired(inputRequiredMonitorFence);
 
-    size_t dispatchSize = getSizeDispatch(relaxedOrderingSchedulerWillBeNeeded, batchBuffer.hasRelaxedOrderingDependencies, dispatchMonitorFence);
+    size_t dispatchSize = this->getUllsStateSize() + getSizeDispatch(relaxedOrderingSchedulerWillBeNeeded, batchBuffer.hasRelaxedOrderingDependencies, dispatchMonitorFence);
 
     if (this->copyCommandBufferIntoRing(batchBuffer)) {
         dispatchSize += (batchBuffer.stream->getUsed() - batchBuffer.startOffset) - 2 * getSizeStartSection();
@@ -978,8 +960,14 @@ bool DirectSubmissionHw<GfxFamily, Dispatcher>::dispatchCommandBuffer(BatchBuffe
         }
     }
 
+    auto needStart = !this->ringStart;
+    this->ringStart = true;
+    auto startVA = ringCommandStream.getCurrentGpuAddressPosition();
+
     this->switchRingBuffersNeeded(requiredMinimalSize, batchBuffer.allocationsForResidency);
 
+    this->dispatchUllsState();
+
     if (this->relaxedOrderingEnabled && batchBuffer.hasStallingCmds && this->relaxedOrderingSchedulerRequired) {
         dispatchRelaxedOrderingQueueStall();
     }
@@ -991,9 +979,10 @@ bool DirectSubmissionHw<GfxFamily, Dispatcher>::dispatchCommandBuffer(BatchBuffe
     void *currentPosition = dispatchWorkloadSection(batchBuffer, dispatchMonitorFence);
 
     cpuCachelineFlush(currentPosition, dispatchSize);
-    handleResidency();
 
-    this->unblockGpu();
+    if (!this->submitCommandBufferToGpu(needStart, startVA, requiredMinimalSize)) {
+        return false;
+    }
 
     cpuCachelineFlush(semaphorePtr, MemoryConstants::cacheLineSize);
     currentQueueWorkCount++;
@@ -1008,6 +997,17 @@ bool DirectSubmissionHw<GfxFamily, Dispatcher>::dispatchCommandBuffer(BatchBuffe
     return ringStart;
 }
 
+template <typename GfxFamily, typename Dispatcher>
+bool DirectSubmissionHw<GfxFamily, Dispatcher>::submitCommandBufferToGpu(bool needStart, uint64_t gpuAddress, size_t size) {
+    if (needStart) {
+        return this->submit(gpuAddress, size);
+    } else {
+        handleResidency();
+        this->unblockGpu();
+        return true;
+    }
+}
+
 template <typename GfxFamily, typename Dispatcher>
 inline void DirectSubmissionHw<GfxFamily, Dispatcher>::setReturnAddress(void *returnCmd, uint64_t returnAddress) {
     using MI_BATCH_BUFFER_START = typename GfxFamily::MI_BATCH_BUFFER_START;

shared/source/direct_submission/windows/wddm_direct_submission.inl

@@ -55,7 +55,9 @@ WddmDirectSubmission<GfxFamily, Dispatcher>::~WddmDirectSubmission() {
 
 template <typename GfxFamily, typename Dispatcher>
 inline void WddmDirectSubmission<GfxFamily, Dispatcher>::flushMonitorFence() {
-    this->startRingBuffer();
+    auto needStart = !this->ringStart;
+    this->ringStart = true;
+    auto startVA = this->ringCommandStream.getCurrentGpuAddressPosition();
 
     size_t requiredMinimalSize = this->getSizeSemaphoreSection(false) +
                                  Dispatcher::getSizeMonitorFence(this->rootDeviceEnvironment) +
@@ -71,8 +73,7 @@ inline void WddmDirectSubmission<GfxFamily, Dispatcher>::flushMonitorFence() {
     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->submitCommandBufferToGpu(needStart, startVA, requiredMinimalSize);
     this->currentQueueWorkCount++;
 
     this->updateTagValueImpl(this->currentRingBuffer);