Enable aggregating command buffers with multiple osContexts

- Store inspectionId for each osContext in GraphicsAllocation
- Pass osContextId to aggregateCommandBuffer and use it to select inspectionId

Change-Id: I2c377ad7577a8c882cc89c1205430cb581c2c0d5
Signed-off-by: Maciej Dziuban <maciej.dziuban@intel.com>

runtime/command_stream/command_stream_receiver_hw.inl

@@ -505,7 +505,7 @@ inline void CommandStreamReceiverHw<GfxFamily>::flushBatchedSubmissions() {
 
     auto &commandBufferList = this->submissionAggregator->peekCmdBufferList();
     if (!commandBufferList.peekIsEmpty()) {
-        auto &device = commandBufferList.peekHead()->device;
+        const auto totalMemoryBudget = static_cast<size_t>(commandBufferList.peekHead()->device.getDeviceInfo().globalMemSize / 2);
 
         ResidencyContainer surfacesForSubmit;
         ResourcePackage resourcePackage;
@@ -515,7 +515,7 @@ inline void CommandStreamReceiverHw<GfxFamily>::flushBatchedSubmissions() {
 
         while (!commandBufferList.peekIsEmpty()) {
             size_t totalUsedSize = 0u;
-            this->submissionAggregator->aggregateCommandBuffers(resourcePackage, totalUsedSize, (size_t)device.getDeviceInfo().globalMemSize * 5 / 10);
+            this->submissionAggregator->aggregateCommandBuffers(resourcePackage, totalUsedSize, totalMemoryBudget, osContext->getContextId());
             auto primaryCmdBuffer = commandBufferList.removeFrontOne();
             auto nextCommandBuffer = commandBufferList.peekHead();
             auto currentBBendLocation = primaryCmdBuffer->batchBufferEndLocation;

runtime/command_stream/submissions_aggregator.cpp

@@ -13,7 +13,7 @@ void OCLRT::SubmissionAggregator::recordCommandBuffer(CommandBuffer *commandBuff
     this->cmdBuffers.pushTailOne(*commandBuffer);
 }
 
-void OCLRT::SubmissionAggregator::aggregateCommandBuffers(ResourcePackage &resourcePackage, size_t &totalUsedSize, size_t totalMemoryBudget) {
+void OCLRT::SubmissionAggregator::aggregateCommandBuffers(ResourcePackage &resourcePackage, size_t &totalUsedSize, size_t totalMemoryBudget, uint32_t osContextId) {
     auto primaryCommandBuffer = this->cmdBuffers.peekHead();
     auto currentInspection = this->inspectionId;
 
@@ -27,8 +27,8 @@ void OCLRT::SubmissionAggregator::aggregateCommandBuffers(ResourcePackage &resou
 
     //primary command buffers must fix to budget
     for (auto &graphicsAllocation : primaryCommandBuffer->surfaces) {
-        if (graphicsAllocation->inspectionId < currentInspection) {
+        if (graphicsAllocation->getInspectionId(osContextId) < currentInspection) {
-            graphicsAllocation->inspectionId = currentInspection;
+            graphicsAllocation->setInspectionId(currentInspection, osContextId);
             resourcePackage.push_back(graphicsAllocation);
             totalUsedSize += graphicsAllocation->getUnderlyingBufferSize();
         }
@@ -62,16 +62,16 @@ void OCLRT::SubmissionAggregator::aggregateCommandBuffers(ResourcePackage &resou
             if (graphicsAllocation == primaryBatchGraphicsAllocation) {
                 continue;
             }
-            if (graphicsAllocation->inspectionId < currentInspection) {
+            if (graphicsAllocation->getInspectionId(osContextId) < currentInspection) {
-                graphicsAllocation->inspectionId = currentInspection;
+                graphicsAllocation->setInspectionId(currentInspection, osContextId);
                 newResources.push_back(graphicsAllocation);
                 nextCommandBufferNewResourcesSize += graphicsAllocation->getUnderlyingBufferSize();
             }
         }
 
         if (nextCommandBuffer->batchBuffer.commandBufferAllocation && (nextCommandBuffer->batchBuffer.commandBufferAllocation != primaryBatchGraphicsAllocation)) {
-            if (nextCommandBuffer->batchBuffer.commandBufferAllocation->inspectionId < currentInspection) {
+            if (nextCommandBuffer->batchBuffer.commandBufferAllocation->getInspectionId(osContextId) < currentInspection) {
-                nextCommandBuffer->batchBuffer.commandBufferAllocation->inspectionId = currentInspection;
+                nextCommandBuffer->batchBuffer.commandBufferAllocation->setInspectionId(currentInspection, osContextId);
                 newResources.push_back(nextCommandBuffer->batchBuffer.commandBufferAllocation);
                 nextCommandBufferNewResourcesSize += nextCommandBuffer->batchBuffer.commandBufferAllocation->getUnderlyingBufferSize();
             }

runtime/command_stream/submissions_aggregator.h

@@ -43,7 +43,7 @@ struct BatchBuffer {
 };
 
 struct CommandBuffer : public IDNode<CommandBuffer> {
-    CommandBuffer(Device &);
+    CommandBuffer(Device &device);
     ResidencyContainer surfaces;
     BatchBuffer batchBuffer;
     void *batchBufferEndLocation = nullptr;
@@ -62,7 +62,7 @@ using ResourcePackage = StackVec<GraphicsAllocation *, 128>;
 class SubmissionAggregator {
   public:
     void recordCommandBuffer(CommandBuffer *commandBuffer);
-    void aggregateCommandBuffers(ResourcePackage &resourcePackage, size_t &totalUsedSize, size_t totalMemoryBudget);
+    void aggregateCommandBuffers(ResourcePackage &resourcePackage, size_t &totalUsedSize, size_t totalMemoryBudget, uint32_t osContextId);
     CommandBufferList &peekCmdBufferList() { return cmdBuffers; }
 
   protected:

runtime/memory_manager/graphics_allocation.h

@@ -123,6 +123,8 @@ class GraphicsAllocation : public IDNode<GraphicsAllocation> {
     void updateTaskCount(uint32_t newTaskCount, uint32_t contextId);
     uint32_t getTaskCount(uint32_t contextId) const { return usageInfos[contextId].taskCount; }
     void resetTaskCount(uint32_t contextId) { updateTaskCount(objectNotUsed, contextId); }
+    uint32_t getInspectionId(uint32_t contextId) { return usageInfos[contextId].inspectionId; }
+    void setInspectionId(uint32_t newInspectionId, uint32_t contextId) { usageInfos[contextId].inspectionId = newInspectionId; }
 
     void updateResidencyTaskCount(uint32_t newTaskCount, uint32_t contextId) { usageInfos[contextId].residencyTaskCount = newTaskCount; }
     uint32_t getResidencyTaskCount(uint32_t contextId) const { return usageInfos[contextId].residencyTaskCount; }
@@ -135,6 +137,7 @@ class GraphicsAllocation : public IDNode<GraphicsAllocation> {
     struct UsageInfo {
         uint32_t taskCount = objectNotUsed;
         uint32_t residencyTaskCount = objectNotResident;
+        uint32_t inspectionId = 0u;
     };
 
     //this variable can only be modified from SubmissionAggregator
@@ -148,7 +151,6 @@ class GraphicsAllocation : public IDNode<GraphicsAllocation> {
     uint32_t reuseCount = 0; // GraphicsAllocation can be reused by shared resources
     bool evictable = true;
     MemoryPool::Type memoryPool = MemoryPool::MemoryNull;
-    uint32_t inspectionId = 0;
     AllocationType allocationType = AllocationType::UNKNOWN;
     bool aubWritable = true;
     bool allocDumpable = false;

unit_tests/command_stream/submissions_aggregator_tests.cpp

@@ -72,13 +72,13 @@ TEST(SubmissionsAggregator, givenTwoCommandBuffersWhenMergeResourcesIsCalledThen
     size_t totalMemoryBudget = -1;
     ResourcePackage resourcePackage;
 
-    submissionsAggregator.aggregateCommandBuffers(resourcePackage, totalUsedSize, totalMemoryBudget);
+    submissionsAggregator.aggregateCommandBuffers(resourcePackage, totalUsedSize, totalMemoryBudget, 0u);
 
     EXPECT_EQ(0u, totalUsedSize);
 
     submissionsAggregator.recordCommandBuffer(cmdBuffer);
 
-    submissionsAggregator.aggregateCommandBuffers(resourcePackage, totalUsedSize, totalMemoryBudget);
+    submissionsAggregator.aggregateCommandBuffers(resourcePackage, totalUsedSize, totalMemoryBudget, 0u);
 
     EXPECT_EQ(15u, totalUsedSize);
     totalUsedSize = 0;
@@ -92,7 +92,7 @@ TEST(SubmissionsAggregator, givenTwoCommandBuffersWhenMergeResourcesIsCalledThen
     EXPECT_EQ(5u, cmdBuffer->surfaces.size());
     EXPECT_EQ(4u, cmdBuffer2->surfaces.size());
 
-    submissionsAggregator.aggregateCommandBuffers(resourcePackage, totalUsedSize, totalMemoryBudget);
+    submissionsAggregator.aggregateCommandBuffers(resourcePackage, totalUsedSize, totalMemoryBudget, 0u);
 
     //command buffer 2 is aggregated to command buffer 1
     auto primaryBatchInstepctionId = submissionsAggregator.peekCommandBuffersList().peekHead()->inspectionId;
@@ -153,7 +153,7 @@ TEST(SubmissionsAggregator, givenSubmissionAggregatorWhenThreeCommandBuffersAreS
     EXPECT_EQ(4u, cmdBuffer2->surfaces.size());
     EXPECT_EQ(2u, cmdBuffer3->surfaces.size());
 
-    submissionsAggregator.aggregateCommandBuffers(resourcePackage, totalUsedSize, totalMemoryBudget);
+    submissionsAggregator.aggregateCommandBuffers(resourcePackage, totalUsedSize, totalMemoryBudget, 0u);
 
     //command buffer 3 and 2 is aggregated to command buffer 1
     auto primaryBatchInstepctionId = submissionsAggregator.peekCommandBuffersList().peekHead()->inspectionId;
@@ -207,7 +207,7 @@ TEST(SubmissionsAggregator, givenMultipleCommandBuffersWhenTheyAreAggreagateWith
     submissionsAggregator.recordCommandBuffer(cmdBuffer2);
     submissionsAggregator.recordCommandBuffer(cmdBuffer3);
 
-    submissionsAggregator.aggregateCommandBuffers(resourcePackage, totalUsedSize, totalMemoryBudget);
+    submissionsAggregator.aggregateCommandBuffers(resourcePackage, totalUsedSize, totalMemoryBudget, 0u);
 
     //command buffer 2 is aggregated to command buffer 1, comand buffer 3 becomes command buffer 2
     EXPECT_EQ(submissionsAggregator.peekCommandBuffersList().peekHead(), cmdBuffer);
@@ -263,7 +263,7 @@ TEST(SubmissionsAggregator, givenMultipleCommandBuffersWhenAggregateIsCalledMult
     submissionsAggregator.recordCommandBuffer(cmdBuffer2);
     submissionsAggregator.recordCommandBuffer(cmdBuffer3);
 
-    submissionsAggregator.aggregateCommandBuffers(resourcePackage, totalUsedSize, totalMemoryBudget);
+    submissionsAggregator.aggregateCommandBuffers(resourcePackage, totalUsedSize, totalMemoryBudget, 0u);
 
     //command buffers not aggregated due to too low limit
     EXPECT_EQ(submissionsAggregator.peekCommandBuffersList().peekHead(), cmdBuffer);
@@ -275,7 +275,7 @@ TEST(SubmissionsAggregator, givenMultipleCommandBuffersWhenAggregateIsCalledMult
     resourcePackage.clear();
     totalUsedSize = 0;
 
-    submissionsAggregator.aggregateCommandBuffers(resourcePackage, totalUsedSize, totalMemoryBudget);
+    submissionsAggregator.aggregateCommandBuffers(resourcePackage, totalUsedSize, totalMemoryBudget, 0u);
     //all cmd buffers are merged to 1
     EXPECT_EQ(cmdBuffer3->inspectionId, cmdBuffer2->inspectionId);
     EXPECT_EQ(cmdBuffer->inspectionId, cmdBuffer2->inspectionId);
@@ -313,7 +313,7 @@ TEST(SubmissionsAggregator, givenMultipleCommandBuffersWithDifferentGraphicsAllo
     submissionsAggregator.recordCommandBuffer(cmdBuffer);
     submissionsAggregator.recordCommandBuffer(cmdBuffer2);
 
-    submissionsAggregator.aggregateCommandBuffers(resourcePackage, totalUsedSize, totalMemoryBudget);
+    submissionsAggregator.aggregateCommandBuffers(resourcePackage, totalUsedSize, totalMemoryBudget, 0u);
 
     EXPECT_EQ(4u, resourcePackage.size());
     EXPECT_EQ(15u, totalUsedSize);
@@ -347,7 +347,7 @@ TEST(SubmissionsAggregator, givenTwoCommandBufferWhereSecondContainsFirstOnResou
     submissionsAggregator.recordCommandBuffer(cmdBuffer);
     submissionsAggregator.recordCommandBuffer(cmdBuffer2);
 
-    submissionsAggregator.aggregateCommandBuffers(resourcePackage, totalUsedSize, totalMemoryBudget);
+    submissionsAggregator.aggregateCommandBuffers(resourcePackage, totalUsedSize, totalMemoryBudget, 0u);
 
     //resource pack shuold have 3 surfaces
     EXPECT_EQ(3u, resourcePackage.size());
@@ -379,7 +379,7 @@ TEST(SubmissionsAggregator, givenTwoCommandBufferWhereSecondContainsTheFirstComm
     submissionsAggregator.recordCommandBuffer(cmdBuffer);
     submissionsAggregator.recordCommandBuffer(cmdBuffer2);
 
-    submissionsAggregator.aggregateCommandBuffers(resourcePackage, totalUsedSize, totalMemoryBudget);
+    submissionsAggregator.aggregateCommandBuffers(resourcePackage, totalUsedSize, totalMemoryBudget, 0u);
 
     //resource pack shuold have 3 surfaces
     EXPECT_EQ(2u, resourcePackage.size());
@@ -408,7 +408,7 @@ TEST(SubmissionsAggregator, givenCommandBuffersRequiringDifferentCoherencySettin
     ResourcePackage resourcePackage;
     size_t totalUsedSize = 0;
     size_t totalMemoryBudget = 200;
-    submissionsAggregator.aggregateCommandBuffers(resourcePackage, totalUsedSize, totalMemoryBudget);
+    submissionsAggregator.aggregateCommandBuffers(resourcePackage, totalUsedSize, totalMemoryBudget, 0u);
     EXPECT_EQ(1u, totalUsedSize);
     EXPECT_EQ(1u, resourcePackage.size());
     EXPECT_NE(cmdBuffer->inspectionId, cmdBuffer2->inspectionId);
@@ -437,7 +437,7 @@ TEST(SubmissionsAggregator, givenCommandBuffersRequiringDifferentThrottleSetting
     ResourcePackage resourcePackage;
     size_t totalUsedSize = 0;
     size_t totalMemoryBudget = 200;
-    submissionsAggregator.aggregateCommandBuffers(resourcePackage, totalUsedSize, totalMemoryBudget);
+    submissionsAggregator.aggregateCommandBuffers(resourcePackage, totalUsedSize, totalMemoryBudget, 0u);
     EXPECT_EQ(1u, totalUsedSize);
     EXPECT_EQ(1u, resourcePackage.size());
     EXPECT_NE(cmdBuffer->inspectionId, cmdBuffer2->inspectionId);
@@ -466,7 +466,7 @@ TEST(SubmissionsAggregator, givenCommandBuffersRequiringDifferentPrioritySetting
     ResourcePackage resourcePackage;
     size_t totalUsedSize = 0;
     size_t totalMemoryBudget = 200;
-    submissionsAggregator.aggregateCommandBuffers(resourcePackage, totalUsedSize, totalMemoryBudget);
+    submissionsAggregator.aggregateCommandBuffers(resourcePackage, totalUsedSize, totalMemoryBudget, 0u);
     EXPECT_EQ(1u, totalUsedSize);
     EXPECT_EQ(1u, resourcePackage.size());
     EXPECT_NE(cmdBuffer->inspectionId, cmdBuffer2->inspectionId);
@@ -479,14 +479,72 @@ TEST(SubmissionsAggregator, dontAllocateFlushStamp) {
     EXPECT_EQ(nullptr, cmdBuffer.flushStamp->getStampReference());
 }
 
+TEST(SubmissionsAggregator, givenMultipleOsContextsWhenAggregatingGraphicsAllocationsThenUseInspectionIdCorrespondingWithOsContextId) {
+    SubmissionAggregator submissionsAggregator;
+    ResourcePackage resourcePackage;
+    const auto totalMemoryBudget = 3u;
+    size_t totalUsedSize = 0;
+    std::unique_ptr<Device> device(MockDevice::createWithNewExecutionEnvironment<MockDevice>(nullptr));
+    CommandBuffer *cmdBuffer0 = new CommandBuffer(*device);
+    CommandBuffer *cmdBuffer1 = new CommandBuffer(*device);
+
+    MockGraphicsAllocation alloc0(nullptr, 1);
+    MockGraphicsAllocation alloc1(nullptr, 1);
+    MockGraphicsAllocation alloc2(nullptr, 1);
+    MockGraphicsAllocation alloc3(nullptr, 1);
+
+    cmdBuffer0->surfaces.push_back(&alloc0);
+    cmdBuffer0->surfaces.push_back(&alloc1);
+    cmdBuffer1->surfaces.push_back(&alloc2);
+    cmdBuffer1->surfaces.push_back(&alloc3);
+
+    submissionsAggregator.recordCommandBuffer(cmdBuffer0);
+    submissionsAggregator.recordCommandBuffer(cmdBuffer1);
+    submissionsAggregator.aggregateCommandBuffers(resourcePackage, totalUsedSize, totalMemoryBudget, 3u);
+    EXPECT_EQ(1u, alloc0.getInspectionId(3u));
+    EXPECT_EQ(1u, alloc1.getInspectionId(3u));
+    EXPECT_EQ(1u, alloc2.getInspectionId(3u));
+    EXPECT_EQ(1u, alloc3.getInspectionId(3u));
+}
+
+TEST(SubmissionsAggregator, givenMultipleOsContextsWhenAggregatingGraphicsAllocationsThenDoNotUpdateInspectionIdsOfOtherContexts) {
+    SubmissionAggregator submissionsAggregator;
+    ResourcePackage resourcePackage;
+    const auto totalMemoryBudget = 2u;
+    size_t totalUsedSize = 0;
+    std::unique_ptr<Device> device(MockDevice::createWithNewExecutionEnvironment<MockDevice>(nullptr));
+    CommandBuffer *cmdBuffer0 = new CommandBuffer(*device);
+    CommandBuffer *cmdBuffer1 = new CommandBuffer(*device);
+
+    MockGraphicsAllocation alloc0(nullptr, 1);
+    MockGraphicsAllocation alloc1(nullptr, 1);
+
+    cmdBuffer0->surfaces.push_back(&alloc0);
+    cmdBuffer0->surfaces.push_back(&alloc1);
+
+    submissionsAggregator.recordCommandBuffer(cmdBuffer0);
+    submissionsAggregator.recordCommandBuffer(cmdBuffer1);
+    submissionsAggregator.aggregateCommandBuffers(resourcePackage, totalUsedSize, totalMemoryBudget, 3u);
+
+    for (auto osContextId = 0u; osContextId < alloc1.usageInfos.size(); osContextId++) {
+        if (osContextId != 3u) {
+            EXPECT_EQ(0u, alloc0.getInspectionId(osContextId));
+        }
+    }
+    for (auto osContextId = 0u; osContextId < alloc0.usageInfos.size(); osContextId++) {
+        if (osContextId != 3u) {
+            EXPECT_EQ(0u, alloc0.getInspectionId(osContextId));
+        }
+    }
+}
+
 struct SubmissionsAggregatorTests : public ::testing::Test {
     void SetUp() override {
         device.reset(MockDevice::createWithNewExecutionEnvironment<MockDevice>(platformDevices[0]));
         context.reset(new MockContext(device.get()));
     }
 
-    template <typename T>
+    void overrideCsr(CommandStreamReceiver *newCsr) {
-    void overrideCsr(T *newCsr) {
         device->resetCommandStreamReceiver(newCsr);
         newCsr->overrideDispatchPolicy(DispatchMode::BatchedDispatch);
     }

unit_tests/memory_manager/graphics_allocation_tests.cpp

@@ -10,6 +10,13 @@
 
 using namespace OCLRT;
 
+TEST(GraphicsAllocationTest, givenGraphicsAllocationWhenIsCreatedThenAllInspectionIdsAreSetToZero) {
+    MockGraphicsAllocation graphicsAllocation(nullptr, 0u, 0u, maxOsContextCount, true);
+    for (auto i = 0u; i < maxOsContextCount; i++) {
+        EXPECT_EQ(0u, graphicsAllocation.getInspectionId(i));
+    }
+}
+
 TEST(GraphicsAllocationTest, givenGraphicsAllocationWhenIsCreatedThenTaskCountsAreInitializedProperly) {
     GraphicsAllocation graphicsAllocation1(nullptr, 0u, 0u, 0u, maxOsContextCount, true);
     GraphicsAllocation graphicsAllocation2(nullptr, 0u, 0u, maxOsContextCount, true);

unit_tests/mocks/mock_graphics_allocation.h

@@ -14,14 +14,19 @@ class MockGraphicsAllocation : public GraphicsAllocation {
     using GraphicsAllocation::GraphicsAllocation;
     using GraphicsAllocation::objectNotResident;
     using GraphicsAllocation::objectNotUsed;
+    using GraphicsAllocation::usageInfos;
 
     MockGraphicsAllocation() : MockGraphicsAllocation(true) {}
     MockGraphicsAllocation(bool shareable) : GraphicsAllocation(nullptr, 0u, 0, maxOsContextCount, shareable) {}
     MockGraphicsAllocation(void *buffer, size_t sizeIn) : GraphicsAllocation(buffer, castToUint64(buffer), 0llu, sizeIn, maxOsContextCount, false) {
     }
-    void resetInspectionId() {
+
-        this->inspectionId = 0;
+    void resetInspectionIds() {
+        for (auto &usageInfo : usageInfos) {
+            usageInfo.inspectionId = 0u;
+        }
     }
+
     void overrideMemoryPool(MemoryPool::Type pool) {
         this->memoryPool = pool;
     }

unit_tests/mocks/mock_program.cpp

@@ -36,7 +36,7 @@ void GlobalMockSipProgram::resetAllocationState() {
     for (uint32_t index = 0u; index < maxOsContextCount; index++) {
         this->kernelInfoArray[0]->kernelAllocation->resetResidencyTaskCount(index);
     }
-    static_cast<MockGraphicsAllocation *>(this->kernelInfoArray[0]->kernelAllocation)->resetInspectionId();
+    static_cast<MockGraphicsAllocation *>(this->kernelInfoArray[0]->kernelAllocation)->resetInspectionIds();
 }
 void GlobalMockSipProgram::initSipProgram() {
     cl_int retVal = 0;