Allow Device creating multiple CSRs [6/n]

- Introduce default Engine query
- Improve Deferred Deleter usage
- Remove Tag Allocation from Device

Change-Id: Iaa88d8dc0166325acf9a157dcd2217ea408ee285
Signed-off-by: Dunajski, Bartosz <[email protected]>

Author: BartoszDunajski

runtime/command_queue/command_queue.cpp

@@ -77,14 +77,14 @@ CommandQueue::CommandQueue(Context *context,
     commandQueueProperties = getCmdQueueProperties<cl_command_queue_properties>(properties);
     flushStamp.reset(new FlushStampTracker(true));
 
-    processProperties();
-
     if (device) {
-        engine = &device->getEngine(engineId);
+        engine = &device->getDefaultEngine();
         if (getCommandStreamReceiver().peekTimestampPacketWriteEnabled()) {
             timestampPacketContainer = std::make_unique<TimestampPacketContainer>();
         }
     }
+
+    processProperties();
 }
 
 CommandQueue::~CommandQueue() {

runtime/command_queue/command_queue.h

@@ -428,7 +428,7 @@ class CommandQueue : public BaseObject<_cl_command_queue> {
 
     QueuePriority priority;
     QueueThrottle throttle;
-    size_t engineId = 0;
+    uint32_t engineId = 0;
 
     bool perfCountersEnabled;
     cl_uint perfCountersConfig;

runtime/device/device.cpp

@@ -129,7 +129,6 @@ bool Device::createDeviceImpl(const HardwareInfo *pHwInfo, Device &outDevice) {
         pDevice->osTime = OSTime::create(commandStreamReceiver->getOSInterface());
     }
     pDevice->driverInfo.reset(DriverInfo::create(commandStreamReceiver->getOSInterface()));
-    pDevice->tagAddress = reinterpret_cast<uint32_t *>(commandStreamReceiver->getTagAllocation()->getUnderlyingBuffer());
 
     pDevice->initializeCaps();
 
@@ -150,6 +149,7 @@ bool Device::createDeviceImpl(const HardwareInfo *pHwInfo, Device &outDevice) {
     }
 
     outDevice.executionEnvironment->memoryManager->setForce32BitAllocations(pDevice->getDeviceInfo().force32BitAddressess);
+    outDevice.executionEnvironment->memoryManager->setDefaultEngineIndex(deviceCsrIndex);
 
     if (pDevice->preemptionMode == PreemptionMode::MidThread || pDevice->isSourceLevelDebuggerActive()) {
         size_t requiredSize = pHwInfo->capabilityTable.requiredPreemptionSurfaceSize;

runtime/device/device.h

@@ -70,7 +70,8 @@ class Device : public BaseObject<_cl_device_id> {
         deviceInfo.force32BitAddressess = value;
     }
 
-    EngineControl &getEngine(size_t engineId);
+    EngineControl &getEngine(uint32_t engineId);
+    EngineControl &getDefaultEngine();
 
     volatile uint32_t *getTagAddress() const;
 
@@ -146,7 +147,6 @@ class Device : public BaseObject<_cl_device_id> {
     HardwareCapabilities hardwareCapabilities = {};
     DeviceInfo deviceInfo;
 
-    volatile uint32_t *tagAddress = nullptr;
     GraphicsAllocation *preemptionAllocation = nullptr;
     std::unique_ptr<OSTime> osTime;
     std::unique_ptr<DriverInfo> driverInfo;
@@ -161,6 +161,7 @@ class Device : public BaseObject<_cl_device_id> {
     PreemptionMode preemptionMode;
     ExecutionEnvironment *executionEnvironment = nullptr;
     uint32_t deviceIndex = 0u;
+    uint32_t defaultEngineIndex = 0;
 };
 
 template <cl_device_info Param>
@@ -171,12 +172,12 @@ inline void Device::getCap(const void *&src,
     retSize = size = DeviceInfoTable::Map<Param>::size;
 }
 
-inline EngineControl &Device::getEngine(size_t engineId) {
+inline EngineControl &Device::getEngine(uint32_t engineId) {
     return engines[engineId];
 }
 
-inline volatile uint32_t *Device::getTagAddress() const {
-    return tagAddress;
+inline EngineControl &Device::getDefaultEngine() {
+    return engines[defaultEngineIndex];
 }
 
 inline MemoryManager *Device::getMemoryManager() const {

runtime/device_queue/device_queue_hw.inl

@@ -230,7 +230,7 @@ void DeviceQueueHw<GfxFamily>::addExecutionModelCleanUpSection(Kernel *parentKer
 
     PipeControlHelper<GfxFamily>::obtainPipeControlAndProgramPostSyncOperation(&slbCS, PIPE_CONTROL::POST_SYNC_OPERATION_WRITE_IMMEDIATE_DATA, criticalSectionAddress, ExecutionModelCriticalSection::Free);
 
-    uint64_t tagAddress = (uint64_t)device->getTagAddress();
+    uint64_t tagAddress = reinterpret_cast<uint64_t>(device->getDefaultEngine().commandStreamReceiver->getTagAddress());
 
     addPipeControlCmdWa();
 

runtime/execution_environment/execution_environment.h

@@ -24,7 +24,7 @@ class BuiltIns;
 struct HardwareInfo;
 class OSInterface;
 
-using CsrContainer = std::array<std::unique_ptr<CommandStreamReceiver>, EngineInstanceT::numGpgpuEngineInstances>;
+using CsrContainer = std::vector<std::array<std::unique_ptr<CommandStreamReceiver>, EngineInstanceT::numGpgpuEngineInstances>>;
 
 class ExecutionEnvironment : public ReferenceTrackedObject<ExecutionEnvironment> {
   private:
@@ -51,7 +51,7 @@ class ExecutionEnvironment : public ReferenceTrackedObject<ExecutionEnvironment>
     std::unique_ptr<OSInterface> osInterface;
     std::unique_ptr<MemoryManager> memoryManager;
     std::unique_ptr<AubCenter> aubCenter;
-    std::vector<CsrContainer> commandStreamReceivers;
+    CsrContainer commandStreamReceivers;
     std::unique_ptr<BuiltIns> builtins;
     std::unique_ptr<CompilerInterface> compilerInterface;
     std::unique_ptr<SourceLevelDebugger> sourceLevelDebugger;

runtime/mem_obj/mem_obj.cpp

@@ -288,7 +288,7 @@ void MemObj::releaseAllocatedMapPtr() {
 }
 
 void MemObj::waitForCsrCompletion() {
-    memoryManager->getCommandStreamReceiver(0)->waitForCompletionWithTimeout(false, TimeoutControls::maxTimeout, graphicsAllocation->getTaskCount(0u));
+    memoryManager->getDefaultCommandStreamReceiver(0)->waitForCompletionWithTimeout(false, TimeoutControls::maxTimeout, graphicsAllocation->getTaskCount(0u));
 }
 
 void MemObj::destroyGraphicsAllocation(GraphicsAllocation *allocation, bool asyncDestroy) {

runtime/memory_manager/deferrable_allocation_deletion.cpp

@@ -8,6 +8,7 @@
 #include "runtime/command_stream/command_stream_receiver.h"
 #include "runtime/memory_manager/deferrable_allocation_deletion.h"
 #include "runtime/memory_manager/memory_manager.h"
+#include "runtime/os_interface/os_context.h"
 
 namespace OCLRT {
 
@@ -16,11 +17,16 @@ DeferrableAllocationDeletion::DeferrableAllocationDeletion(MemoryManager &memory
 void DeferrableAllocationDeletion::apply() {
     while (graphicsAllocation.isUsed()) {
 
-        for (auto contextId = 0u; contextId < memoryManager.getOsContextCount(); contextId++) {
-            if (graphicsAllocation.isUsedByContext(contextId)) {
-                auto currentContextTaskCount = *memoryManager.getCommandStreamReceiver(contextId)->getTagAddress();
-                if (graphicsAllocation.getTaskCount(contextId) <= currentContextTaskCount) {
-                    graphicsAllocation.resetTaskCount(contextId);
+        for (auto &deviceCsrs : memoryManager.getCommandStreamReceivers()) {
+            for (auto &csr : deviceCsrs) {
+                if (csr) {
+                    auto contextId = csr->getOsContext().getContextId();
+                    if (graphicsAllocation.isUsedByContext(contextId)) {
+                        auto currentContextTaskCount = *csr->getTagAddress();
+                        if (graphicsAllocation.getTaskCount(contextId) <= currentContextTaskCount) {
+                            graphicsAllocation.resetTaskCount(contextId);
+                        }
+                    }
                 }
             }
         }

runtime/memory_manager/host_ptr_manager.cpp

@@ -282,7 +282,7 @@ RequirementsStatus HostPtrManager::checkAllocationsForOverlapping(MemoryManager
         if (checkedFragments->status[i] == OverlapStatus::FRAGMENT_OVERLAPING_AND_BIGGER_THEN_STORED_FRAGMENT) {
             // clean temporary allocations
 
-            auto commandStreamReceiver = memoryManager.getCommandStreamReceiver(0);
+            auto commandStreamReceiver = memoryManager.getDefaultCommandStreamReceiver(0);
             auto allocationStorage = commandStreamReceiver->getInternalAllocationStorage();
             uint32_t taskCount = *commandStreamReceiver->getTagAddress();
             allocationStorage->cleanAllocationList(taskCount, TEMPORARY_ALLOCATION);

runtime/memory_manager/memory_manager.cpp

@@ -137,10 +137,10 @@ void MemoryManager::freeGraphicsMemory(GraphicsAllocation *gfxAllocation) {
 //if not in use destroy in place
 //if in use pass to temporary allocation list that is cleaned on blocking calls
 void MemoryManager::checkGpuUsageAndDestroyGraphicsAllocations(GraphicsAllocation *gfxAllocation) {
-    if (!gfxAllocation->isUsed() || gfxAllocation->getTaskCount(0u) <= *getCommandStreamReceiver(0)->getTagAddress()) {
+    if (!gfxAllocation->isUsed() || gfxAllocation->getTaskCount(0u) <= *getCommandStreamReceivers()[0][defaultEngineIndex]->getTagAddress()) {
         freeGraphicsMemory(gfxAllocation);
     } else {
-        getCommandStreamReceiver(0)->getInternalAllocationStorage()->storeAllocation(std::unique_ptr<GraphicsAllocation>(gfxAllocation), TEMPORARY_ALLOCATION);
+        getCommandStreamReceivers()[0][defaultEngineIndex]->getInternalAllocationStorage()->storeAllocation(std::unique_ptr<GraphicsAllocation>(gfxAllocation), TEMPORARY_ALLOCATION);
     }
 }
 
@@ -274,9 +274,13 @@ GraphicsAllocation *MemoryManager::allocateGraphicsMemory(const AllocationData &
     }
     return allocateGraphicsMemory(allocationData.size, MemoryConstants::pageSize, allocationData.flags.forcePin, allocationData.flags.uncacheable);
 }
-CommandStreamReceiver *MemoryManager::getCommandStreamReceiver(uint32_t contextId) {
-    UNRECOVERABLE_IF(executionEnvironment.commandStreamReceivers.size() < 1);
-    return executionEnvironment.commandStreamReceivers[contextId][0].get();
+
+const CsrContainer &MemoryManager::getCommandStreamReceivers() const {
+    return executionEnvironment.commandStreamReceivers;
+}
+
+CommandStreamReceiver *MemoryManager::getDefaultCommandStreamReceiver(uint32_t deviceId) const {
+    return executionEnvironment.commandStreamReceivers[deviceId][defaultEngineIndex].get();
 }
 
 } // namespace OCLRT