Extend copy engine selector data

DziubanMaciejIntel · Compute-Runtime-Automation · commit 6f6ee90aa595 · 2021-07-06T12:28:55.000+02:00
Signed-off-by: Maciej Dziuban &lt;maciej.dziuban@intel.com&gt;
diff --git a/opencl/source/cl_device/cl_device.cpp b/opencl/source/cl_device/cl_device.cpp
@@ -132,7 +132,7 @@ const HardwareInfo &ClDevice::getHardwareInfo() const { return device.getHardwar
 EngineControl &ClDevice::getEngine(aub_stream::EngineType engineType, EngineUsage engineUsage) { return device.getEngine(engineType, engineUsage); }
 EngineControl &ClDevice::getDefaultEngine() { return device.getDefaultEngine(); }
 EngineControl &ClDevice::getInternalEngine() { return device.getInternalEngine(); }
-std::atomic<uint32_t> &ClDevice::getSelectorCopyEngine() { return device.getSelectorCopyEngine(); }
+SelectorCopyEngine &ClDevice::getSelectorCopyEngine() { return device.getSelectorCopyEngine(); }
 MemoryManager *ClDevice::getMemoryManager() const { return device.getMemoryManager(); }
 GmmHelper *ClDevice::getGmmHelper() const { return device.getGmmHelper(); }
 GmmClientContext *ClDevice::getGmmClientContext() const { return device.getGmmClientContext(); }
diff --git a/opencl/source/cl_device/cl_device.h b/opencl/source/cl_device/cl_device.h
@@ -36,6 +36,7 @@ struct EngineControl;
 struct HardwareCapabilities;
 struct HardwareInfo;
 struct RootDeviceEnvironment;
+struct SelectorCopyEngine;
 
 template <>
 struct OpenCLObjectMapper<_cl_device_id> {
@@ -69,7 +70,7 @@ class ClDevice : public BaseObject<_cl_device_id> {
     EngineControl &getDefaultEngine();
     EngineControl &getInternalEngine();
     EngineControl *getInternalCopyEngine();
-    std::atomic<uint32_t> &getSelectorCopyEngine();
+    SelectorCopyEngine &getSelectorCopyEngine();
     MemoryManager *getMemoryManager() const;
     GmmHelper *getGmmHelper() const;
     GmmClientContext *getGmmClientContext() const;
diff --git a/opencl/source/command_queue/command_queue.cpp b/opencl/source/command_queue/command_queue.cpp
@@ -60,6 +60,9 @@ CommandQueue *CommandQueue::create(Context *context,
 }
 
 CommandQueue::CommandQueue(Context *context, ClDevice *device, const cl_queue_properties *properties)
+    : CommandQueue(context, device, properties, false) {}
+
+CommandQueue::CommandQueue(Context *context, ClDevice *device, const cl_queue_properties *properties, bool internalUsage)
     : context(context), device(device) {
     if (context) {
         context->incRefInternal();
@@ -83,8 +86,10 @@ CommandQueue::CommandQueue(Context *context, ClDevice *device, const cl_queue_pr
             deferredTimestampPackets = std::make_unique<TimestampPacketContainer>();
         }
         if (bcsAllowed) {
-            auto &selectorCopyEngine = device->getDeviceById(0)->getSelectorCopyEngine();
-            bcsEngine = device->getDeviceById(0)->getDevice().tryGetEngine(EngineHelpers::getBcsEngineType(hwInfo, selectorCopyEngine), EngineUsage::Regular);
+            auto &neoDevice = device->getDeviceById(0)->getDevice();
+            auto &selectorCopyEngine = neoDevice.getSelectorCopyEngine();
+            auto bcsEngineType = EngineHelpers::getBcsEngineType(hwInfo, selectorCopyEngine, internalUsage);
+            bcsEngine = neoDevice.tryGetEngine(bcsEngineType, EngineUsage::Regular);
         }
     }
 
@@ -109,6 +114,11 @@ CommandQueue::~CommandQueue() {
         if (this->perfCountersEnabled) {
             device->getPerformanceCounters()->shutdown();
         }
+
+        if (bcsEngine) {
+            auto &selectorCopyEngine = device->getDeviceById(0)->getSelectorCopyEngine();
+            EngineHelpers::releaseBcsEngineType(bcsEngine->getEngineType(), selectorCopyEngine);
+        }
     }
 
     timestampPacketContainer.reset();
diff --git a/opencl/source/command_queue/command_queue.h b/opencl/source/command_queue/command_queue.h
@@ -67,6 +67,8 @@ class CommandQueue : public BaseObject<_cl_command_queue> {
 
     CommandQueue() = delete;
 
+    CommandQueue(Context *context, ClDevice *device, const cl_queue_properties *properties, bool internalUsage);
+
     CommandQueue(Context *context, ClDevice *device,
                  const cl_queue_properties *properties);
 
diff --git a/opencl/source/command_queue/command_queue_hw.h b/opencl/source/command_queue/command_queue_hw.h
@@ -36,7 +36,7 @@ class CommandQueueHw : public CommandQueue {
     CommandQueueHw(Context *context,
                    ClDevice *device,
                    const cl_queue_properties *properties,
-                   bool internalUsage) : BaseClass(context, device, properties) {
+                   bool internalUsage) : BaseClass(context, device, properties, internalUsage) {
 
         auto clPriority = getCmdQueueProperties<cl_queue_priority_khr>(properties, CL_QUEUE_PRIORITY_KHR);
 
diff --git a/shared/source/device/device.h b/shared/source/device/device.h
@@ -26,6 +26,11 @@ class OSTime;
 class SourceLevelDebugger;
 class SubDevice;
 
+struct SelectorCopyEngine : NonCopyableOrMovableClass {
+    std::atomic<bool> isMainUsed = false;
+    std::atomic<uint32_t> selector = 0;
+};
+
 class Device : public ReferenceTrackedObject<Device> {
   public:
     Device &operator=(const Device &) = delete;
@@ -60,7 +65,7 @@ class Device : public ReferenceTrackedObject<Device> {
     EngineControl &getDefaultEngine();
     EngineControl &getInternalEngine();
     EngineControl *getInternalCopyEngine();
-    std::atomic<uint32_t> &getSelectorCopyEngine();
+    SelectorCopyEngine &getSelectorCopyEngine();
     MemoryManager *getMemoryManager() const;
     GmmHelper *getGmmHelper() const;
     GmmClientContext *getGmmClientContext() const;
@@ -166,7 +171,7 @@ class Device : public ReferenceTrackedObject<Device> {
     bool hasGenericSubDevices = false;
     bool engineInstanced = false;
 
-    std::atomic<uint32_t> selectorCopyEngine{0};
+    SelectorCopyEngine selectorCopyEngine = {};
 
     DeviceBitfield deviceBitfield = 1;
 
@@ -195,7 +200,7 @@ inline BuiltIns *Device::getBuiltIns() const {
     return executionEnvironment->rootDeviceEnvironments[getRootDeviceIndex()]->getBuiltIns();
 }
 
-inline std::atomic<uint32_t> &Device::getSelectorCopyEngine() {
+inline SelectorCopyEngine &Device::getSelectorCopyEngine() {
     return selectorCopyEngine;
 }
 
diff --git a/shared/source/helpers/engine_node_helper.h b/shared/source/helpers/engine_node_helper.h
@@ -15,6 +15,7 @@
 
 namespace NEO {
 struct HardwareInfo;
+struct SelectorCopyEngine;
 
 enum class EngineUsage : uint32_t {
     Regular,
@@ -29,7 +30,8 @@ using EngineTypeUsage = std::pair<aub_stream::EngineType, EngineUsage>;
 namespace EngineHelpers {
 bool isCcs(aub_stream::EngineType engineType);
 bool isBcs(aub_stream::EngineType engineType);
-aub_stream::EngineType getBcsEngineType(const HardwareInfo &hwInfo, std::atomic<uint32_t> &selectorCopyEngine);
+aub_stream::EngineType getBcsEngineType(const HardwareInfo &hwInfo, SelectorCopyEngine &selectorCopyEngine, bool internalUsage = false);
+void releaseBcsEngineType(aub_stream::EngineType engineType, SelectorCopyEngine &selectorCopyEngine);
 
 std::string engineTypeToString(aub_stream::EngineType engineType);
 std::string engineTypeToStringAdditional(aub_stream::EngineType engineType);
diff --git a/shared/source/helpers/engine_node_helper_extra.cpp b/shared/source/helpers/engine_node_helper_extra.cpp
@@ -13,10 +13,12 @@ bool isBcs(aub_stream::EngineType engineType) {
     return engineType == aub_stream::ENGINE_BCS;
 }
 
-aub_stream::EngineType getBcsEngineType(const HardwareInfo &hwInfo, std::atomic<uint32_t> &selectorCopyEngine) {
+aub_stream::EngineType getBcsEngineType(const HardwareInfo &hwInfo, SelectorCopyEngine &selectorCopyEngine, bool internalUsage) {
     return aub_stream::EngineType::ENGINE_BCS;
 }
 
+void releaseBcsEngineType(aub_stream::EngineType engineType, SelectorCopyEngine &selectorCopyEngine) {}
+
 std::string engineTypeToStringAdditional(aub_stream::EngineType engineType) {
     return "Unknown";
 }

Original file line number	Diff line number	Diff line change
`@@ -60,6 +60,9 @@ CommandQueue CommandQueue::create(Context context,`
`60`	`60`	`}`
`61`	`61`
`62`	`62`	`CommandQueue::CommandQueue(Context context, ClDevice device, const cl_queue_properties *properties)`
	`63`	`+ : CommandQueue(context, device, properties, false) {}`
	`64`	`+`
	`65`	`+CommandQueue::CommandQueue(Context context, ClDevice device, const cl_queue_properties *properties, bool internalUsage)`
`63`	`66`	`: context(context), device(device) {`
`64`	`67`	`if (context) {`
`65`	`68`	`context->incRefInternal();`
`@@ -83,8 +86,10 @@ CommandQueue::CommandQueue(Context context, ClDevice device, const cl_queue_pr`
`83`	`86`	`deferredTimestampPackets = std::make_unique<TimestampPacketContainer>();`
`84`	`87`	`}`
`85`	`88`	`if (bcsAllowed) {`
`86`		`- auto &selectorCopyEngine = device->getDeviceById(0)->getSelectorCopyEngine();`
`87`		`- bcsEngine = device->getDeviceById(0)->getDevice().tryGetEngine(EngineHelpers::getBcsEngineType(hwInfo, selectorCopyEngine), EngineUsage::Regular);`
	`89`	`+ auto &neoDevice = device->getDeviceById(0)->getDevice();`
	`90`	`+ auto &selectorCopyEngine = neoDevice.getSelectorCopyEngine();`
	`91`	`+ auto bcsEngineType = EngineHelpers::getBcsEngineType(hwInfo, selectorCopyEngine, internalUsage);`
	`92`	`+ bcsEngine = neoDevice.tryGetEngine(bcsEngineType, EngineUsage::Regular);`
`88`	`93`	`}`
`89`	`94`	`}`
`90`	`95`
`@@ -109,6 +114,11 @@ CommandQueue::~CommandQueue() {`
`109`	`114`	`if (this->perfCountersEnabled) {`
`110`	`115`	`device->getPerformanceCounters()->shutdown();`
`111`	`116`	`}`
	`117`	`+`
	`118`	`+ if (bcsEngine) {`
	`119`	`+ auto &selectorCopyEngine = device->getDeviceById(0)->getSelectorCopyEngine();`
	`120`	`+ EngineHelpers::releaseBcsEngineType(bcsEngine->getEngineType(), selectorCopyEngine);`
	`121`	`+ }`
`112`	`122`	`}`
`113`	`123`
`114`	`124`	`timestampPacketContainer.reset();`
Original file line number	Diff line number	Diff line change
`@@ -13,10 +13,12 @@ bool isBcs(aub_stream::EngineType engineType) {`
`13`	`13`	`return engineType == aub_stream::ENGINE_BCS;`
`14`	`14`	`}`
`15`	`15`
`16`		`-aub_stream::EngineType getBcsEngineType(const HardwareInfo &hwInfo, std::atomic<uint32_t> &selectorCopyEngine) {`
	`16`	`+aub_stream::EngineType getBcsEngineType(const HardwareInfo &hwInfo, SelectorCopyEngine &selectorCopyEngine, bool internalUsage) {`
`17`	`17`	`return aub_stream::EngineType::ENGINE_BCS;`
`18`	`18`	`}`
`19`	`19`
	`20`	`+void releaseBcsEngineType(aub_stream::EngineType engineType, SelectorCopyEngine &selectorCopyEngine) {}`
	`21`	`+`
`20`	`22`	`std::string engineTypeToStringAdditional(aub_stream::EngineType engineType) {`
`21`	`23`	`return "Unknown";`
`22`	`24`	`}`