Allow Device creating multiple CSRs [2/n]

Create new OsContext per CSR

Change-Id: I8dad7fc1ab450e560f78eba3152b5913791e59a3
Signed-off-by: Dunajski, Bartosz <bartosz.dunajski@intel.com>

runtime/device/device.cpp

@@ -86,8 +86,8 @@ Device::~Device() {
         performanceCounters->shutdown();
     }
 
-    for (auto &csr : commandStreamReceiver) {
-        csr->flushBatchedSubmissions();
+    for (auto &engine : engines) {
+        engine.commandStreamReceiver->flushBatchedSubmissions();
     }
 
     if (deviceInfo.sourceLevelDebuggerActive && executionEnvironment->sourceLevelDebugger) {
@@ -116,21 +116,20 @@ bool Device::createDeviceImpl(const HardwareInfo *pHwInfo, Device &outDevice) {
     executionEnvironment->initializeMemoryManager(outDevice.getEnabled64kbPages(), outDevice.getEnableLocalMemory(),
                                                   outDevice.getDeviceIndex(), deviceCsrIndex);
 
-    outDevice.osContext = new OsContext(executionEnvironment->osInterface.get(), outDevice.getDeviceIndex());
-    executionEnvironment->memoryManager->registerOsContext(outDevice.osContext);
-
-    outDevice.commandStreamReceiver.resize(1);
-    outDevice.commandStreamReceiver[deviceCsrIndex] = executionEnvironment->commandStreamReceivers[outDevice.getDeviceIndex()][deviceCsrIndex].get();
-    if (!outDevice.commandStreamReceiver[deviceCsrIndex]->initializeTagAllocation()) {
+    auto osContext = executionEnvironment->memoryManager->createAndRegisterOsContext();
+    auto commandStreamReceiver = executionEnvironment->commandStreamReceivers[outDevice.getDeviceIndex()][deviceCsrIndex].get();
+    if (!commandStreamReceiver->initializeTagAllocation()) {
         return false;
     }
 
+    outDevice.engines.emplace_back(commandStreamReceiver, osContext);
+
     auto pDevice = &outDevice;
     if (!pDevice->osTime) {
-        pDevice->osTime = OSTime::create(outDevice.commandStreamReceiver[deviceCsrIndex]->getOSInterface());
+        pDevice->osTime = OSTime::create(commandStreamReceiver->getOSInterface());
     }
-    pDevice->driverInfo.reset(DriverInfo::create(outDevice.commandStreamReceiver[deviceCsrIndex]->getOSInterface()));
-    pDevice->tagAddress = reinterpret_cast<uint32_t *>(outDevice.commandStreamReceiver[deviceCsrIndex]->getTagAllocation()->getUnderlyingBuffer());
+    pDevice->driverInfo.reset(DriverInfo::create(commandStreamReceiver->getOSInterface()));
+    pDevice->tagAddress = reinterpret_cast<uint32_t *>(commandStreamReceiver->getTagAllocation()->getUnderlyingBuffer());
 
     pDevice->initializeCaps();
 
@@ -142,8 +141,8 @@ bool Device::createDeviceImpl(const HardwareInfo *pHwInfo, Device &outDevice) {
     }
 
     uint32_t deviceHandle = 0;
-    if (outDevice.commandStreamReceiver[deviceCsrIndex]->getOSInterface()) {
-        deviceHandle = outDevice.commandStreamReceiver[deviceCsrIndex]->getOSInterface()->getDeviceHandle();
+    if (commandStreamReceiver->getOSInterface()) {
+        deviceHandle = commandStreamReceiver->getOSInterface()->getDeviceHandle();
     }
 
     if (pDevice->deviceInfo.sourceLevelDebuggerActive) {
@@ -160,14 +159,14 @@ bool Device::createDeviceImpl(const HardwareInfo *pHwInfo, Device &outDevice) {
         if (!pDevice->preemptionAllocation) {
             return false;
         }
-        outDevice.commandStreamReceiver[deviceCsrIndex]->setPreemptionCsrAllocation(pDevice->preemptionAllocation);
+        commandStreamReceiver->setPreemptionCsrAllocation(pDevice->preemptionAllocation);
         auto sipType = SipKernel::getSipKernelType(pHwInfo->pPlatform->eRenderCoreFamily, pDevice->isSourceLevelDebuggerActive());
         initSipKernel(sipType, *pDevice);
     }
 
     if (DebugManager.flags.EnableExperimentalCommandBuffer.get() > 0) {
-        outDevice.commandStreamReceiver[deviceCsrIndex]->setExperimentalCmdBuffer(
-            std::unique_ptr<ExperimentalCommandBuffer>(new ExperimentalCommandBuffer(outDevice.commandStreamReceiver[deviceCsrIndex], pDevice->getDeviceInfo().profilingTimerResolution)));
+        commandStreamReceiver->setExperimentalCmdBuffer(std::unique_ptr<ExperimentalCommandBuffer>(
+            new ExperimentalCommandBuffer(commandStreamReceiver, pDevice->getDeviceInfo().profilingTimerResolution)));
     }
 
     return true;
@@ -230,7 +229,7 @@ unique_ptr_if_unused<Device> Device::release() {
 
 bool Device::isSimulation() const {
     bool simulation = hwInfo.capabilityTable.isSimulation(hwInfo.pPlatform->usDeviceID);
-    if (commandStreamReceiver[0]->getType() != CommandStreamReceiverType::CSR_HW) {
+    if (engines[0].commandStreamReceiver->getType() != CommandStreamReceiverType::CSR_HW) {
         simulation = true;
     }
     if (hwInfo.pSkuTable->ftrSimulationMode) {

runtime/device/device.h

@@ -11,6 +11,7 @@
 #include "runtime/device/device_info_map.h"
 #include "runtime/execution_environment/execution_environment.h"
 #include "runtime/helpers/base_object.h"
+#include "runtime/helpers/engine_control.h"
 #include "runtime/helpers/hw_info.h"
 #include "runtime/memory_manager/memory_constants.h"
 #include "runtime/os_interface/performance_counters.h"
@@ -118,7 +119,7 @@ class Device : public BaseObject<_cl_device_id> {
     SourceLevelDebugger *getSourceLevelDebugger() { return executionEnvironment->sourceLevelDebugger.get(); }
     ExecutionEnvironment *getExecutionEnvironment() const { return executionEnvironment; }
     const HardwareCapabilities &getHardwareCapabilities() const { return hardwareCapabilities; }
-    OsContext *getOsContext() const { return osContext; }
+    OsContext *getOsContext() const { return engines[0].osContext; }
     uint32_t getDeviceIndex() { return deviceIndex; }
     bool isFullRangeSvm() {
         return getHardwareInfo().capabilityTable.gpuAddressSpace == MemoryConstants::max48BitAddress;
@@ -155,7 +156,7 @@ class Device : public BaseObject<_cl_device_id> {
     std::unique_ptr<DriverInfo> driverInfo;
     std::unique_ptr<PerformanceCounters> performanceCounters;
 
-    OsContext *osContext = nullptr;
+    std::vector<EngineControl> engines;
 
     void *slmWindowStartAddress = nullptr;
 
@@ -165,7 +166,6 @@ class Device : public BaseObject<_cl_device_id> {
     EngineType engineType;
     ExecutionEnvironment *executionEnvironment = nullptr;
     uint32_t deviceIndex = 0u;
-    std::vector<CommandStreamReceiver *> commandStreamReceiver;
 };
 
 template <cl_device_info Param>
@@ -177,7 +177,7 @@ inline void Device::getCap(const void *&src,
 }
 
 inline CommandStreamReceiver &Device::getCommandStreamReceiver() {
-    return *this->commandStreamReceiver[0];
+    return *engines[0].commandStreamReceiver;
 }
 
 inline volatile uint32_t *Device::getTagAddress() const {

runtime/helpers/CMakeLists.txt

@@ -28,6 +28,7 @@ set(RUNTIME_SRCS_HELPERS_BASE
   ${CMAKE_CURRENT_SOURCE_DIR}/dispatch_info.h
   ${CMAKE_CURRENT_SOURCE_DIR}/dispatch_info_builder.h
   ${CMAKE_CURRENT_SOURCE_DIR}/enable_product.inl
+  ${CMAKE_CURRENT_SOURCE_DIR}/engine_control.h
   ${CMAKE_CURRENT_SOURCE_DIR}/error_mappers.h
   ${CMAKE_CURRENT_SOURCE_DIR}/extendable_enum.h
   ${CMAKE_CURRENT_SOURCE_DIR}/file_io.cpp

runtime/helpers/engine_control.h

@@ -0,0 +1,22 @@
+/*
+ * Copyright (C) 2018 Intel Corporation
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ */
+
+#pragma once
+
+namespace OCLRT {
+class CommandStreamReceiver;
+class OsContext;
+
+struct EngineControl {
+    EngineControl() = default;
+    EngineControl(CommandStreamReceiver *commandStreamReceiver, OsContext *osContext)
+        : commandStreamReceiver(commandStreamReceiver), osContext(osContext){};
+
+    CommandStreamReceiver *commandStreamReceiver = nullptr;
+    OsContext *osContext = nullptr;
+};
+} // namespace OCLRT

runtime/memory_manager/graphics_allocation.cpp

@@ -20,16 +20,16 @@ GraphicsAllocation::GraphicsAllocation(void *cpuPtrIn, uint64_t gpuAddress, uint
                                                                                                                                                               size(sizeIn),
                                                                                                                                                               cpuPtr(cpuPtrIn),
                                                                                                                                                               gpuAddress(gpuAddress),
-                                                                                                                                                              usageInfos(osContextCount),
                                                                                                                                                               isShareable(isShareable) {
+    usageInfos.resize(maxOsContextCount);
 }
 
 GraphicsAllocation::GraphicsAllocation(void *cpuPtrIn, size_t sizeIn, osHandle sharedHandleIn, uint32_t osContextCount, bool isShareable) : size(sizeIn),
                                                                                                                                             cpuPtr(cpuPtrIn),
                                                                                                                                             gpuAddress(castToUint64(cpuPtrIn)),
                                                                                                                                             sharedHandle(sharedHandleIn),
-                                                                                                                                            usageInfos(osContextCount),
                                                                                                                                             isShareable(isShareable) {
+    usageInfos.resize(maxOsContextCount);
 }
 GraphicsAllocation::~GraphicsAllocation() = default;
 

runtime/memory_manager/graphics_allocation.h

@@ -152,7 +152,7 @@ class GraphicsAllocation : public IDNode<GraphicsAllocation> {
     bool aubWritable = true;
     bool allocDumpable = false;
     bool memObjectsAllocationWithWritableFlags = false;
-    StackVec<UsageInfo, maxOsContextCount> usageInfos;
+    std::vector<UsageInfo> usageInfos;
     std::atomic<uint32_t> registeredContextsNum{0};
     bool isShareable = false;
 };

runtime/memory_manager/memory_manager.cpp

@@ -19,6 +19,7 @@
 #include "runtime/memory_manager/host_ptr_manager.h"
 #include "runtime/memory_manager/internal_allocation_storage.h"
 #include "runtime/os_interface/os_context.h"
+#include "runtime/os_interface/os_interface.h"
 #include "runtime/utilities/stackvec.h"
 #include "runtime/utilities/tag_allocator.h"
 
@@ -200,13 +201,16 @@ bool MemoryManager::isMemoryBudgetExhausted() const {
     return false;
 }
 
-void MemoryManager::registerOsContext(OsContext *contextToRegister) {
-    auto contextId = contextToRegister->getContextId();
+OsContext *MemoryManager::createAndRegisterOsContext() {
+    auto contextId = ++latestContextId;
     if (contextId + 1 > registeredOsContexts.size()) {
         registeredOsContexts.resize(contextId + 1);
     }
-    contextToRegister->incRefInternal();
-    registeredOsContexts[contextToRegister->getContextId()] = contextToRegister;
+    auto osContext = new OsContext(executionEnvironment.osInterface.get(), contextId);
+    osContext->incRefInternal();
+    registeredOsContexts[contextId] = osContext;
+
+    return osContext;
 }
 
 bool MemoryManager::getAllocationData(AllocationData &allocationData, const AllocationFlags &flags, const DevicesBitfield devicesBitfield,

runtime/memory_manager/memory_manager.h

@@ -24,6 +24,7 @@ class GraphicsAllocation;
 class HostPtrManager;
 class CommandStreamReceiver;
 class OsContext;
+class OSInterface;
 class TimestampPacket;
 
 struct HwPerfCounter;
@@ -246,7 +247,7 @@ class MemoryManager {
         ::alignedFree(ptr);
     }
 
-    void registerOsContext(OsContext *contextToRegister);
+    OsContext *createAndRegisterOsContext();
     uint32_t getOsContextCount() { return static_cast<uint32_t>(registeredOsContexts.size()); }
     CommandStreamReceiver *getCommandStreamReceiver(uint32_t contextId);
     HostPtrManager *getHostPtrManager() const { return hostPtrManager.get(); }
@@ -270,6 +271,7 @@ class MemoryManager {
     ExecutionEnvironment &executionEnvironment;
     std::vector<OsContext *> registeredOsContexts;
     std::unique_ptr<HostPtrManager> hostPtrManager;
+    uint32_t latestContextId = std::numeric_limits<uint32_t>::max();
 };
 
 std::unique_ptr<DeferredDeleter> createDeferredDeleter();

unit_tests/memory_manager/memory_manager_tests.cpp

@@ -197,7 +197,9 @@ TEST_F(MemoryAllocatorTest, allocateSystemAligned) {
 TEST_F(MemoryAllocatorTest, allocateGraphics) {
     unsigned int alignment = 4096;
 
+    memoryManager->createAndRegisterOsContext();
     auto allocation = memoryManager->allocateGraphicsMemory(sizeof(char));
+
     ASSERT_NE(nullptr, allocation);
     // initial taskCount must be -1. if not, we may kill allocation before it will be used
     EXPECT_EQ((uint32_t)-1, allocation->getTaskCount(0));
@@ -1224,6 +1226,7 @@ TEST_F(MemoryManagerWithCsrTest, givenAllocationThatWasUsedAndIsCompletedWhenche
 }
 
 TEST_F(MemoryManagerWithCsrTest, givenAllocationThatWasUsedAndIsNotCompletedWhencheckGpuUsageAndDestroyGraphicsAllocationsIsCalledThenItIsAddedToTemporaryAllocationList) {
+    memoryManager->createAndRegisterOsContext();
     auto usedAllocationAndNotGpuCompleted = memoryManager->allocateGraphicsMemory(4096);
 
     auto tagAddress = csr->getTagAddress();
@@ -1393,24 +1396,21 @@ TEST(GraphicsAllocation, givenSharedHandleBasedConstructorWhenGraphicsAllocation
 }
 
 TEST(ResidencyDataTest, givenOsContextWhenItIsRegisteredToMemoryManagerThenRefCountIncreases) {
-    auto osContext = new OsContext(nullptr, 0u);
     ExecutionEnvironment executionEnvironment;
-    OsAgnosticMemoryManager memoryManager(false, false, executionEnvironment);
-    memoryManager.registerOsContext(osContext);
+    MockMemoryManager memoryManager(false, false, executionEnvironment);
+    memoryManager.createAndRegisterOsContext();
     EXPECT_EQ(1u, memoryManager.getOsContextCount());
-    EXPECT_EQ(1, osContext->getRefInternalCount());
+    EXPECT_EQ(1, memoryManager.registeredOsContexts[0]->getRefInternalCount());
 }
 
 TEST(ResidencyDataTest, givenTwoOsContextsWhenTheyAreRegistredFromHigherToLowerThenProperSizeIsReturned) {
-    auto osContext2 = new OsContext(nullptr, 1u);
-    auto osContext = new OsContext(nullptr, 0u);
     ExecutionEnvironment executionEnvironment;
-    OsAgnosticMemoryManager memoryManager(false, false, executionEnvironment);
-    memoryManager.registerOsContext(osContext2);
-    memoryManager.registerOsContext(osContext);
+    MockMemoryManager memoryManager(false, false, executionEnvironment);
+    memoryManager.createAndRegisterOsContext();
+    memoryManager.createAndRegisterOsContext();
     EXPECT_EQ(2u, memoryManager.getOsContextCount());
-    EXPECT_EQ(1, osContext->getRefInternalCount());
-    EXPECT_EQ(1, osContext2->getRefInternalCount());
+    EXPECT_EQ(1, memoryManager.registeredOsContexts[0]->getRefInternalCount());
+    EXPECT_EQ(1, memoryManager.registeredOsContexts[1]->getRefInternalCount());
 }
 
 TEST(ResidencyDataTest, givenResidencyDataWhenUpdateCompletionDataIsCalledThenItIsProperlyUpdated) {

unit_tests/mocks/mock_device.cpp

@@ -19,7 +19,7 @@ MockDevice::MockDevice(const HardwareInfo &hwInfo)
     executionEnvironment->commandStreamReceivers.resize(getDeviceIndex() + 1);
     executionEnvironment->commandStreamReceivers[getDeviceIndex()].push_back(std::unique_ptr<CommandStreamReceiver>(commandStreamReceiver));
     this->executionEnvironment->memoryManager = std::move(this->mockMemoryManager);
-    this->commandStreamReceiver.push_back(commandStreamReceiver);
+    this->engines.emplace_back(commandStreamReceiver, nullptr);
 }
 MockDevice::MockDevice(const HardwareInfo &hwInfo, ExecutionEnvironment *executionEnvironment, uint32_t deviceIndex)
     : Device(hwInfo, executionEnvironment, deviceIndex) {
@@ -49,7 +49,7 @@ void MockDevice::resetCommandStreamReceiver(CommandStreamReceiver *newCsr) {
     executionEnvironment->commandStreamReceivers[getDeviceIndex()][0].reset(newCsr);
     executionEnvironment->commandStreamReceivers[getDeviceIndex()][0]->initializeTagAllocation();
     executionEnvironment->commandStreamReceivers[getDeviceIndex()][0]->setPreemptionCsrAllocation(preemptionAllocation);
-    this->commandStreamReceiver[0] = newCsr;
+    this->engines[0].commandStreamReceiver = newCsr;
     UNRECOVERABLE_IF(getDeviceIndex() != 0u);
     this->tagAddress = executionEnvironment->commandStreamReceivers[getDeviceIndex()][0]->getTagAddress();
 }

unit_tests/mocks/mock_device.h

@@ -99,7 +99,7 @@ class MockDevice : public Device {
                 size_t alignment = 256 * MemoryConstants::kiloByte;
                 bool uncacheable = getWaTable()->waCSRUncachable;
                 this->preemptionAllocation = executionEnvironment->memoryManager->allocateGraphicsMemory(requiredSize, alignment, false, uncacheable);
-                this->commandStreamReceiver[0]->setPreemptionCsrAllocation(preemptionAllocation);
+                this->engines[0].commandStreamReceiver->setPreemptionCsrAllocation(preemptionAllocation);
             }
         }
     }

unit_tests/mocks/mock_memory_manager.h

@@ -20,6 +20,7 @@ class MockMemoryManager : public OsAgnosticMemoryManager {
     using MemoryManager::allocateGraphicsMemory;
     using MemoryManager::allocateGraphicsMemoryInPreferredPool;
     using MemoryManager::getAllocationData;
+    using MemoryManager::registeredOsContexts;
     using MemoryManager::timestampPacketAllocator;
     using OsAgnosticMemoryManager::OsAgnosticMemoryManager;
     MockMemoryManager(ExecutionEnvironment &executionEnvironment) : OsAgnosticMemoryManager(false, false, executionEnvironment) {

unit_tests/os_interface/windows/device_command_stream_tests.cpp

@@ -256,11 +256,9 @@ TEST(WddmPreemptionHeaderTests, givenWddmCommandStreamReceiverWhenPreemptionIsOf
 
     auto commandBuffer = executionEnvironment->memoryManager->allocateGraphicsMemory(4096);
     LinearStream cs(commandBuffer);
-
-    BatchBuffer batchBuffer{cs.getGraphicsAllocation(), 0, 0, nullptr, false, false, QueueThrottle::MEDIUM, cs.getUsed(), &cs};
     OsContext *osContext = new OsContext(executionEnvironment->osInterface.get(), 0u);
     osContext->incRefInternal();
-    executionEnvironment->memoryManager->registerOsContext(osContext);
+    BatchBuffer batchBuffer{cs.getGraphicsAllocation(), 0, 0, nullptr, false, false, QueueThrottle::MEDIUM, cs.getUsed(), &cs};
     executionEnvironment->commandStreamReceivers[0][0]->flush(batchBuffer, EngineType::ENGINE_RCS,
                                                               executionEnvironment->commandStreamReceivers[0][0]->getResidencyAllocations(), *osContext);
     auto commandHeader = wddm->submitResult.commandHeaderSubmitted;
@@ -288,7 +286,6 @@ TEST(WddmPreemptionHeaderTests, givenWddmCommandStreamReceiverWhenPreemptionIsOn
     BatchBuffer batchBuffer{cs.getGraphicsAllocation(), 0, 0, nullptr, false, false, QueueThrottle::MEDIUM, cs.getUsed(), &cs};
     OsContext *osContext = new OsContext(executionEnvironment->osInterface.get(), 0u);
     osContext->incRefInternal();
-    executionEnvironment->memoryManager->registerOsContext(osContext);
     executionEnvironment->commandStreamReceivers[0][0]->flush(batchBuffer, EngineType::ENGINE_RCS, executionEnvironment->commandStreamReceivers[0][0]->getResidencyAllocations(), *osContext);
     auto commandHeader = wddm->submitResult.commandHeaderSubmitted;
     COMMAND_BUFFER_HEADER *pHeader = reinterpret_cast<COMMAND_BUFFER_HEADER *>(commandHeader);