RelaxedOrdering: Queue size limit

Related-To: NEO-7458

Signed-off-by: Dunajski, Bartosz <bartosz.dunajski@intel.com>

shared/source/debug_settings/debug_variables_base.inl

@@ -334,6 +334,7 @@ DECLARE_DEBUG_VARIABLE(int32_t, DirectSubmissionMaxRingBuffers, -1, "-1: default
 DECLARE_DEBUG_VARIABLE(int32_t, DirectSubmissionDisablePrefetcher, -1, "-1: default, 0 - disable, 1 - enable. If enabled, disable prefetcher is being dispatched")
 DECLARE_DEBUG_VARIABLE(int32_t, DirectSubmissionRelaxedOrdering, -1, "-1: default, 0 - disable, 1 - enable. If enabled, tasks sent to direct submission ring may be dispatched out of order")
 DECLARE_DEBUG_VARIABLE(int32_t, DirectSubmissionRelaxedOrderingForBcs, -1, "-1: default, 0 - disable, 1 - enable. If set, enable RelaxedOrdering feature for BCS engine")
+DECLARE_DEBUG_VARIABLE(int32_t, DirectSubmissionRelaxedOrderingQueueSizeLimit, -1, "-1: default, >0: Max gpu queue size. If limit is reached, scheduler wont consume new work")
 DECLARE_DEBUG_VARIABLE(bool, DirectSubmissionPrintBuffers, false, "Print address of submitted command buffers")
 
 /*FEATURE FLAGS*/

shared/source/direct_submission/direct_submission_hw.inl

@@ -90,6 +90,8 @@ void DirectSubmissionHw<GfxFamily, Dispatcher>::dispatchStaticRelaxedOrderingSch
     uint64_t schedulerStartAddress = schedulerCmdStream.getGpuBase();
     uint64_t deferredTasksListGpuVa = deferredTasksListAllocation->getGpuAddress();
 
+    uint64_t loopSectionStartAddress = schedulerStartAddress + RelaxedOrderingHelper::StaticSchedulerSizeAndOffsetSection<GfxFamily>::loopStartSectionStart;
+
     // 1. Init section
     {
         EncodeMiPredicate<GfxFamily>::encode(schedulerCmdStream, MiPredicateType::Disable);
@@ -182,7 +184,7 @@ void DirectSubmissionHw<GfxFamily, Dispatcher>::dispatchStaticRelaxedOrderingSch
 
         EncodeBatchBufferStartOrEnd<GfxFamily>::programConditionalRegRegBatchBufferStart(
             schedulerCmdStream,
-            schedulerStartAddress + RelaxedOrderingHelper::StaticSchedulerSizeAndOffsetSection<GfxFamily>::loopStartSectionStart,
+            loopSectionStartAddress,
             AluRegisters::R_1, AluRegisters::R_2, CompareOperation::NotEqual, false);
 
         LriHelper<GfxFamily>::program(&schedulerCmdStream, CS_GPR_R2, 0, true);
@@ -193,9 +195,19 @@ void DirectSubmissionHw<GfxFamily, Dispatcher>::dispatchStaticRelaxedOrderingSch
     {
         *schedulerCmdStream.getSpaceForCmd<typename GfxFamily::MI_ARB_CHECK>() = GfxFamily::cmdInitArbCheck;
 
+        uint32_t queueSizeLimit = 2;
+        if (DebugManager.flags.DirectSubmissionRelaxedOrderingQueueSizeLimit.get() != -1) {
+            queueSizeLimit = static_cast<uint32_t>(DebugManager.flags.DirectSubmissionRelaxedOrderingQueueSizeLimit.get());
+        }
+
         EncodeBatchBufferStartOrEnd<GfxFamily>::programConditionalDataRegBatchBufferStart(
             schedulerCmdStream,
-            schedulerStartAddress + RelaxedOrderingHelper::StaticSchedulerSizeAndOffsetSection<GfxFamily>::loopStartSectionStart,
+            loopSectionStartAddress,
+            CS_GPR_R1, queueSizeLimit, CompareOperation::GreaterOrEqual, false);
+
+        EncodeBatchBufferStartOrEnd<GfxFamily>::programConditionalDataRegBatchBufferStart(
+            schedulerCmdStream,
+            loopSectionStartAddress,
             CS_GPR_R5, 1, CompareOperation::Equal, false);
     }
 
@@ -246,7 +258,7 @@ void DirectSubmissionHw<GfxFamily, Dispatcher>::dispatchRelaxedOrderingScheduler
         schedulerCmdStream.getSpace(sizeof(typename GfxFamily::MI_BATCH_BUFFER_START)); // skip patching
     }
 
-    // 7. Semaphore section
+    // 3. Semaphore section
     {
         using MI_SEMAPHORE_WAIT = typename GfxFamily::MI_SEMAPHORE_WAIT;
         using COMPARE_OPERATION = typename GfxFamily::MI_SEMAPHORE_WAIT::COMPARE_OPERATION;

shared/source/direct_submission/relaxed_ordering_helper.h

@@ -53,7 +53,7 @@ struct StaticSchedulerSizeAndOffsetSection {
                                                               (2 * sizeof(MI_LOAD_REGISTER_IMM)) + EncodeMiPredicate<GfxFamily>::getCmdSize();
 
     static constexpr uint64_t drainRequestSectionStart = tasksListLoopCheckSectionStart + tasksListLoopCheckSectionSize;
-    static constexpr uint64_t drainRequestSectionSize = sizeof(typename GfxFamily::MI_ARB_CHECK) + EncodeBatchBufferStartOrEnd<GfxFamily>::getCmdSizeConditionalDataRegBatchBufferStart();
+    static constexpr uint64_t drainRequestSectionSize = sizeof(typename GfxFamily::MI_ARB_CHECK) + (2 * EncodeBatchBufferStartOrEnd<GfxFamily>::getCmdSizeConditionalDataRegBatchBufferStart());
 
     static constexpr uint64_t schedulerLoopCheckSectionJumpStart = drainRequestSectionStart + drainRequestSectionSize;
     static constexpr uint64_t schedulerLoopCheckSectionJumpSize = 2 * sizeof(MI_LOAD_REGISTER_REG) + sizeof(MI_BATCH_BUFFER_START);

shared/test/common/test_files/igdrcl.config

@@ -491,3 +491,4 @@ ForceComputeWalkerPostSyncFlush = -1
 DirectSubmissionRelaxedOrdering = -1
 DirectSubmissionRelaxedOrderingForBcs = -1
 OverrideUserFenceStartValue = -1
+DirectSubmissionRelaxedOrderingQueueSizeLimit = -1

shared/test/unit_test/direct_submission/direct_submission_tests_2.cpp

@@ -1458,13 +1458,25 @@ bool DirectSubmissionRelaxedOrderingTests::verifyStaticSchedulerProgramming(Grap
         return false;
     }
 
+    uint32_t queueLimit = 2;
+    if (DebugManager.flags.DirectSubmissionRelaxedOrderingQueueSizeLimit.get() != -1) {
+        queueLimit = static_cast<uint32_t>(DebugManager.flags.DirectSubmissionRelaxedOrderingQueueSizeLimit.get());
+    }
+
     if (!verifyConditionalDataRegBbStart<FamilyType>(++arbCheck, schedulerStartGpuAddress + RelaxedOrderingHelper::StaticSchedulerSizeAndOffsetSection<FamilyType>::loopStartSectionStart,
+                                                     CS_GPR_R1, queueLimit, CompareOperation::GreaterOrEqual, false)) {
+        return false;
+    }
+
+    auto conditionalBbStartcmds = ptrOffset(arbCheck, EncodeBatchBufferStartOrEnd<FamilyType>::getCmdSizeConditionalDataRegBatchBufferStart());
+
+    if (!verifyConditionalDataRegBbStart<FamilyType>(conditionalBbStartcmds, schedulerStartGpuAddress + RelaxedOrderingHelper::StaticSchedulerSizeAndOffsetSection<FamilyType>::loopStartSectionStart,
                                                      CS_GPR_R5, 1, CompareOperation::Equal, false)) {
         return false;
     }
 
     // 6. Jump to scheduler loop check section (dynamic scheduler)
-    auto lrrCmd = reinterpret_cast<MI_LOAD_REGISTER_REG *>(ptrOffset(arbCheck, EncodeBatchBufferStartOrEnd<FamilyType>::getCmdSizeConditionalDataRegBatchBufferStart()));
+    auto lrrCmd = reinterpret_cast<MI_LOAD_REGISTER_REG *>(ptrOffset(conditionalBbStartcmds, EncodeBatchBufferStartOrEnd<FamilyType>::getCmdSizeConditionalDataRegBatchBufferStart()));
 
     if (!verifyLrr<FamilyType>(lrrCmd, CS_GPR_R0, CS_GPR_R9)) {
         return false;
@@ -1638,6 +1650,19 @@ HWTEST_F(DirectSubmissionRelaxedOrderingTests, whenAllocatingResourcesThenCreate
     EXPECT_EQ(directSubmission.deferredTasksListAllocation, *allocsIter);
 }
 
+HWTEST2_F(DirectSubmissionRelaxedOrderingTests, givenDebugFlagSetWhenDispatchingStaticSchedulerThenOverrideQueueSizeLimit, IsAtLeastXeHpcCore) {
+    using Dispatcher = RenderDispatcher<FamilyType>;
+
+    DebugManager.flags.DirectSubmissionRelaxedOrderingQueueSizeLimit.set(123);
+
+    MockDirectSubmissionHw<FamilyType, Dispatcher> directSubmission(*pDevice->getDefaultEngine().commandStreamReceiver);
+    directSubmission.initialize(true, false);
+
+    EXPECT_EQ(1u, directSubmission.dispatchStaticRelaxedOrderingSchedulerCalled);
+    EXPECT_TRUE(verifyStaticSchedulerProgramming<FamilyType>(*directSubmission.relaxedOrderingSchedulerAllocation,
+                                                             directSubmission.deferredTasksListAllocation->getGpuAddress()));
+}
+
 HWTEST2_F(DirectSubmissionRelaxedOrderingTests, whenInitializingThenDispatchStaticScheduler, IsAtLeastXeHpcCore) {
     using Dispatcher = RenderDispatcher<FamilyType>;